# Spark > Spark designs, builds and runs one system tailored to how your business works, so your team stops fighting spreadsheets and software that doesn’t fit. Spark is a Jersey company that designs, builds and runs bespoke operations systems for growing businesses. Founded by Seb Lawson and Mark McNally, Spark works with engineering, trades, professional services and finance firms, replacing disconnected spreadsheets and off-the-shelf tools with one system built around how each business works. Spark also runs training programmes and builds Herd, an AI micro-learning app for people at work. Spark is based in Jersey, Channel Islands, and works with businesses across Jersey and the UK. The call to action is "Book a review": a free 30-minute call with Seb Lawson (Managing Director) or Mark McNally (co-founder and CTO). Prices on this site are marked £TBC until confirmed. Every page and article has a markdown version: add .md to its address. ## Pages - [Spark: one system built around your business](https://hellospark.co.uk/index.md): Who we are, what we build and how we work. - [Spark Systems](https://hellospark.co.uk/systems.md): What a Spark system covers, how we build it, and how the price works. - [Spark Learning](https://hellospark.co.uk/learning.md): Workshops, programmes and Herd: training on your team’s own work. - [Work](https://hellospark.co.uk/work.md): Case studies of systems we’ve built, written with each client and published once they approve. - [About Spark](https://hellospark.co.uk/about.md): Seb Lawson and Mark McNally, Jersey, and what we believe. - [Build your system](https://hellospark.co.uk/build.md): A short, optional way to spec your system before a review. - [Book a review](https://hellospark.co.uk/book-a-review.md): A free 30-minute call with Seb or Mark. - [Security and your data](https://hellospark.co.uk/security.md): How Spark keeps a client’s system and data safe, and what happens if you leave. - [Questions](https://hellospark.co.uk/faq.md): Straight answers on build or buy, how it works, cost and contracts, security and data, ownership and support, Jersey and regulated firms, AI, working with us and learning. - [The State of AI & Automation](https://hellospark.co.uk/research.md): Our research into how businesses are putting AI and automation to work. ## Services - [Spark Systems](https://hellospark.co.uk/systems): We map how your business works, design one system around it, build it, then run and improve it: CRM, quotes and invoices, job tracking, compliance records, onboarding, and the automations that remove the copy and paste. We keep the tools you already use where they work and join them up. - [Spark Learning](https://hellospark.co.uk/learning): Training that gets teams using AI and their systems well: half-day and full-day workshops, longer programmes over several weeks, and Herd. Practical, built on the team’s own work, and measured. - [Herd](https://hellospark.co.uk/learning#herd): Our AI micro-learning app for people at work: a short lesson a day, on the job they’re already doing. Its own brand, made by Spark. ## Writing - [Build or buy: buy for the jobs every business does, build for how yours works](https://hellospark.co.uk/writing/build-or-buy-bespoke-system.md): Buy software for the jobs every business does the same way. Build where your work is different, or where people spend their week bridging tools that don’t fit. Most growing businesses need both: good bought tools, and one system that joins them up. - [Bespoke software is now within reach of small businesses: what changed, and what didn’t](https://hellospark.co.uk/writing/why-bespoke-is-affordable-now.md): A bespoke system used to need a big budget and a long project. Ready-made parts, AI-assisted building and proven patterns have cut the cost of building. What hasn’t changed is that someone still has to understand how your business works. - [What one system changes: each step hands the next one what it needs](https://hellospark.co.uk/writing/what-one-system-unlocks.md): Joining up your tools means each step of a job hands the next one what it needs, so nobody retypes, chases or hunts for the latest version. We follow two jobs end to end to show what that looks like. Both are worked examples, not client results. - [A bespoke system is as secure as its builder’s answers to eight questions](https://hellospark.co.uk/writing/is-a-bespoke-system-secure.md): A bespoke system can be as secure as anything you buy, but only if the people building it can answer eight plain questions well. Here they are, from where your data lives to what happens with AI, with what a good answer sounds like. - [Owning your system: if you part ways with your builder, you keep everything](https://hellospark.co.uk/writing/owning-your-system.md): If you stop working with the people who built your system, you should keep the code, the data and the documentation, and be able to hand them to someone else. That belongs in the contract from day one. Here’s what to ask for. - [How we map a business: follow real jobs end to end before building anything](https://hellospark.co.uk/writing/how-we-map-a-business.md): Before we build anything, we sit with your team and follow real jobs from first contact to final invoice, workarounds included. You come away with a clear picture of how work flows and where it breaks, which is worth having even if you never build. - [Switch over a piece at a time: how a new system goes live without a big bang](https://hellospark.co.uk/writing/going-live-without-the-big-bang.md): The safest way to move a business onto a new system is a piece at a time: one part of the work switches over and settles before the next one starts. Here’s how that works, from moving the data to what the first fortnight looks like. - [Spreadsheets aren’t the problem, retyping is: start where the same detail gets keyed twice](https://hellospark.co.uk/writing/retyping-is-the-problem.md): Spreadsheets get the blame, but the real cost is typing the same detail into more than one place. Every retype takes time and invites mistakes. Find where it happens in your business and start there: it’s the cheapest fix with the clearest payback. - [A fixed price is fairer than hourly billing: how pricing a software build works](https://hellospark.co.uk/writing/fixed-price-software.md): A fixed price for a software build puts the risk of overruns on the builder, who is best placed to judge it. Here’s what the price should cover, what sits outside it, how changes are handled, and why it’s the fairer deal for the business paying. - [The Data Protection (Jersey) Law 2018 and your systems: what it asks, and how a good system helps](https://hellospark.co.uk/writing/data-protection-jersey-systems.md): Jersey’s data protection law expects every business to know what personal data it holds, why and where, and to act quickly on requests and breaches. A well-built system keeps the records that make that easier. Your business stays responsible. - [AI just got 75% cheaper. Your business case barely moved.](https://hellospark.co.uk/writing/ai-costs-2026-model-prices-copilot.md): Anthropic’s new small model costs around 75% less to run than the last one, and OpenAI cut its prices too. Run the numbers on a real workflow and the saving is $27 a month. The business case lives somewhere else. - [AI rules across the islands, the UK and the EU: the autumn 2026 map](https://hellospark.co.uk/writing/ai-regulation-channel-islands-uk-eu-2026.md): Five jurisdictions, five different instruments, and one set of questions underneath them all. Here is where AI regulation stands for Channel Islands firms this autumn, the dates to put in the diary, and the single control set that answers every regulator on the list. - [Give AI agents a job description, not the keys](https://hellospark.co.uk/writing/ai-agent-permissions-regulated-firms.md): This summer an AI agent escaped its test environment and reached another company’s systems while trying to get the answers to a test. It was not malicious. It was persistent. Here is how to write an agent’s permissions so persistence never becomes your problem. - [Voice and video are no longer proof of identity](https://hellospark.co.uk/writing/deepfake-payment-fraud-controls.md): Fraudsters have cloned the voice of a Jersey Chief Minister and a lawyer trusted by an Italian private bank. If your payment controls rely on recognising a voice or a face, they no longer work. Here is what does. - [The JFSC’s AI guidance, applied to ten real use cases](https://hellospark.co.uk/writing/jfsc-ai-guidance-ten-use-cases.md): The JFSC has said which AI needs light-touch controls and which needs more. We took ten use cases a Jersey trust or fund business actually runs, sorted them against the guidance, and wrote down the controls each one needs. - [The JFSC’s AI guidance: what it actually means for your firm](https://hellospark.co.uk/writing/jfsc-ai-guidance-what-it-means.md): The Jersey regulator has set out how it expects firms to adopt AI. It is principles-based, proportionate, and lands the responsibility firmly with the board. Here is what it means in practice, and the four moves worth making now. - [Where to start with AI: pick the smallest painful workflow first](https://hellospark.co.uk/writing/where-to-start-with-ai.md): Do not start with a strategy. Start with the smallest painful workflow you can name. Here is how to find the one first use case that pays for itself, teaches your team, and gives the regulator nothing to flinch at. - [RAG explained: how retrieval makes business AI reliable](https://hellospark.co.uk/writing/rag-explained.md): A general-purpose model knows a lot about the world and nothing about your firm. RAG fixes that. It feeds the model your own documents at the moment of the question, so answers are grounded in your sources and can be checked. Here is how it works and when you actually need it. - [AI governance for Channel Islands firms: getting ahead of the JFSC](https://hellospark.co.uk/writing/ai-governance-jersey-jfsc.md): Good AI governance is not a policy document. It is a small set of controls a regulator can inspect and a team can actually live with. Here is the framework we use with Channel Islands firms. - [AI for KYC and client take-on in trust and corporate services](https://hellospark.co.uk/writing/ai-for-kyc-client-take-on.md): Client take-on is the most manual, least billable week in a trust administrator’s month. AI can collect the documents, extract the data, run the screening and draft the risk assessment. The decision stays human, and the audit trail is there from the first click. - [How to measure ROI on AI automation, with a worked example](https://hellospark.co.uk/writing/how-to-measure-ai-roi.md): AI ROI is not mysterious. It is hours saved times loaded cost, plus errors removed, plus capacity freed. Here is the framework we use, and a worked example that shows the maths on a real-shaped workflow. - [Agents, chatbots and automations: what the difference actually is](https://hellospark.co.uk/writing/ai-agents-chatbots-automations.md): Three words get used interchangeably and shouldn’t be. A chatbot talks. An automation runs a fixed path without being asked. An agent decides what to do next. Confusing them is how firms buy the expensive one for a job the cheap one does better. Here is the honest breakdown. - [Automating NAV and investor reporting in fund administration](https://hellospark.co.uk/writing/automating-nav-investor-reporting-fund-administration.md): The NAV cycle is a deadline machine that runs on manual assembly. Automate the reconciliation, the validation and the drafting, surface only the exceptions, and the cycle runs faster with fewer late nights. The numbers still get checked by a human. - [Building an AI audit trail: how regulated firms keep AI decisions defensible](https://hellospark.co.uk/writing/building-an-ai-audit-trail.md): An AI audit trail is five fields and a habit. Capture the input, the model and prompt version, the output, the approver, and the timestamp, from day one. Retrofit it later and you are reconstructing history you no longer have. - [Build, buy or no-code: choosing the right automation approach](https://hellospark.co.uk/writing/build-buy-or-no-code.md): Build, buy, or no-code is the wrong question until you know the workflow. Here is how we choose between them for regulated firms, weighing total cost of ownership, lock-in, and whether a regulator can inspect the result. - [What MCP (Model Context Protocol) is, and why it matters for business AI](https://hellospark.co.uk/writing/what-is-mcp-model-context-protocol.md): MCP is a common plug for AI. Instead of building a bespoke connection between every model and every system, you build one standard socket and any compliant model can use it. It sounds like plumbing because it is, and it is the plumbing that decides whether you are locked in or free to switch. - [AI for law firms: document review and drafting, done safely](https://hellospark.co.uk/writing/ai-for-law-firms-document-review-drafting.md): AI drafts fast and reviews faster, which is exactly why it needs guardrails in a law firm. Contract review, first drafts and matter intake all speed up. Privilege, confidentiality and a lawyer’s verification are what keep it safe. - [Human-in-the-loop AI: what it means and when a regulator expects it](https://hellospark.co.uk/writing/human-in-the-loop-ai.md): Human-in-the-loop is not a person clicking approve. It is a person who can realistically catch an error in the time they are given. Here is how to tell real oversight from theatre, and how to match it to the risk. - [Why most AI pilots never reach production, and how to avoid it](https://hellospark.co.uk/writing/why-ai-pilots-fail.md): A working demo is not a working system. Most AI pilots stall in the same four places: no owner, no integration, no audit trail, and the wrong first use case. Here is how to close the gap between the demo and the day job. - [The EU AI Act and UK AI rules: what professional services firms actually have to do](https://hellospark.co.uk/writing/eu-ai-act-uk-rules-professional-services.md): The EU AI Act sorts systems into risk tiers and regulates the high-risk ones hard. The UK and Jersey lean on principles instead. If you serve regulated clients, you may sit under both. Here is what that actually requires. - [Prompt versioning: keeping AI outputs consistent and auditable](https://hellospark.co.uk/writing/prompt-versioning.md): The instruction you give an AI is not a throwaway sentence. It is the logic of the system, and quietly editing it changes every output that follows. Treat prompts like code: version them, test them, log the changes. Here is why, and what it looks like in a firm a regulator might inspect. - [AI for accountancy practices: from bookkeeping to advisory](https://hellospark.co.uk/writing/ai-for-accountancy-practices.md): The lowest-margin work in an accountancy practice is the most automatable: data entry, categorisation, reconciliations, first-draft accounts. Automate it and you free your best people for the advisory work clients will happily pay for. - [AI and data protection under UK GDPR and the Data Protection (Jersey) Law 2018](https://hellospark.co.uk/writing/ai-data-protection-uk-gdpr-jersey.md): Feeding client data to an AI tool is still processing personal data, and the same rules apply. Lawful basis, minimisation, and knowing where your prompts go. Here is what UK GDPR and the Jersey Law actually require. - [An AI readiness checklist for firms of 5 to 250 people](https://hellospark.co.uk/writing/ai-readiness-checklist.md): You do not need to be AI-ready in the abstract. You need to be ready for one workflow. This honest checklist covers the four things that actually matter: data, governance, skills, and a use case worth building. - [Claude vs GPT vs Gemini for professional services work](https://hellospark.co.uk/writing/claude-vs-gpt-vs-gemini-professional-services.md): The honest answer is that all three are good, and the differences that matter to a law or trust firm are not the ones the benchmarks argue about. Here is how we compare Claude, GPT and Gemini on the things a regulated firm actually cares about, and where we land. - [Document extraction: turning PDFs, contracts and statements into structured data](https://hellospark.co.uk/writing/ai-document-extraction.md): Document extraction is the unglamorous workhorse behind almost every useful AI system in a professional firm. It turns the PDFs, contracts and statements clogging your inbox into structured data. Here is how it works and where the value actually is. ## Optional - [Everything in one file](https://hellospark.co.uk/llms-full.txt): every page and every article, in plain text - [How an AI reads this site](https://hellospark.co.uk/ai.md): what we publish for AI assistants, and why - [Sitemap](https://hellospark.co.uk/sitemap.xml) --- # Pages # Spark: one system built around your business Spark designs, builds and runs one system tailored to how your business works, so your team stops fighting spreadsheets and software that doesn’t fit. Spark is a Jersey company that designs, builds and runs bespoke operations systems for growing businesses. Founded by Seb Lawson and Mark McNally, Spark works with engineering, trades, professional services and finance firms, replacing disconnected spreadsheets and off-the-shelf tools with one system built around how each business works. Spark also runs training programmes and builds Herd, an AI micro-learning app for people at work. ## What we do - **Spark Systems**: We map how your business works, design one system around it, build it, then run and improve it: CRM, quotes and invoices, job tracking, compliance records, onboarding, and the automations that remove the copy and paste. We keep the tools you already use where they work and join them up. https://hellospark.co.uk/systems - **Spark Learning**: Training that gets teams using AI and their systems well: half-day and full-day workshops, longer programmes over several weeks, and Herd. Practical, built on the team’s own work, and measured. https://hellospark.co.uk/learning - **Herd**: Our AI micro-learning app for people at work: a short lesson a day, on the job they’re already doing. Its own brand, made by Spark. https://hellospark.co.uk/learning#herd ## Who it’s for The person who owns how a growing business runs: the founder, managing director or operations lead. Most of our clients are engineering firms, trades and professional services businesses, plus a handful of finance firms, in Jersey and the UK. If your team spends its week copying information from one place to another, you’re probably a good fit. ## How we work 1. **Review** (A free 30-minute call): You tell us how the work flows and where it sticks. We ask the awkward questions and tell you honestly whether a system would help. You get: A straight answer, and where we'd start if it's worth doing. 2. **Map** (Your processes, systems and people): We sit with your team and follow real work end to end: every tool, every spreadsheet, every hand-off, who approves what, and the workarounds nobody wrote down. You get: A map of how your business really runs, a recommendation and a fixed price. The map's yours whether or not we build. 3. **Design** (A clickable prototype in your brand): Before anything is built, your team clicks through the system with your own words, your own steps and your logo, and tells us what's wrong. You get: A design you've signed off, so nobody is surprised later. 4. **Build** (On your real data, with you watching): We build it joined up with the tools you keep, and show you progress every week. Your team tries each piece as it's ready. You get: A weekly demo of real progress, never a big reveal at the end. 5. **Go live** (A piece at a time, nothing overnight): We move your data across, run the old way and the new side by side until your team is happy, and train everyone on their own work. You get: A system your team already knows how to use on day one. 6. **Run** (We stay, and keep improving it): We run it, back it up, keep it secure and watch for problems. Each month we agree what to improve next, as the business changes. You get: A system that keeps up with you, for one monthly fee. ## What we believe - **Built around you**: We start from how your business works, not from a product we want to sell. The system fits the team, not the other way round. - **Plain English**: No jargon, no hype, no black boxes. If we can’t explain it simply, we haven’t designed it well enough. - **We stay**: A system is never finished on handover day. We run it, measure it and keep improving it. ## Next step Book a review: a free 30-minute call with Seb or Mark about how your business works and whether one system would fix what’s slowing you down. https://hellospark.co.uk/book-a-review ## Contact - Book a review (a free 30-minute call with Seb or Mark): https://hellospark.co.uk/book-a-review - Email: seb@sparkcuriosityhq.com - Based in: Jersey, Channel Islands - Company: Spark Curiosity Ltd --- # Spark Systems: one system, built around how your team works We map how your business runs, design one system around it, build it, then run and improve it. CRM, quotes, jobs, invoices and compliance, joined up, with the automations that remove the copy and paste. ## It doesn’t have to be big You don’t have to rebuild the whole business to start. Most clients begin with one fix, such as getting enquiries into one place or quotes out the same day, and add pieces once it’s working. ## The pieces a system can include We build only the pieces you need: - Enquiries and clients: every enquiry and client in one place. Hands each new enquiry to quoting with the details already filled in. - Quotes and proposals: built from your own price list or templates. A won quote becomes a job in one click, with no retyping. - Jobs and projects: who’s doing what, and what’s next. Finished work goes straight to invoicing, with the hours and costs attached. - Invoicing: invoices out the day the work’s done. Syncs with Xero or QuickBooks, so your accountant sees the same numbers. - Documents: letters, reports and contracts from templates. Pulls in the client and job details, then files the result in Drive or SharePoint. - Compliance records: certificates, checks and audit evidence. Records are created as the work happens, so the audit is mostly done already. - Onboarding: new clients and new starters, step by step. Feeds the client list and the team rota, so nobody starts without the basics. - Timesheets: hours logged against the job, on a phone. Hours land on the job and the invoice, and payroll gets one clean export. - Reporting: the numbers you check every Monday, live. Reads from every other piece, so the figures are never a week old. - AI helpers: drafts, summaries and sorting, checked by a person. Reads from the system, drafts the reply or summary, and a named person approves it. - Team training: short lessons on using the system, and AI, well. From Spark Learning: short lessons built on your own system. ## We keep the tools you already use Usually you don’t have to replace your software. We start from the tools you already own and pay for, such as Xero, HubSpot, Gmail or Outlook, Google Drive and your project tracker, keep what works and join it up, so there’s one place to see everything. Where a tool genuinely doesn’t fit, we tell you why and what replacing it would cost before anything changes. ## How we build 1. **Review** (A free 30-minute call): You tell us how the work flows and where it sticks. We ask the awkward questions and tell you honestly whether a system would help. You get: A straight answer, and where we'd start if it's worth doing. 2. **Map** (Your processes, systems and people): We sit with your team and follow real work end to end: every tool, every spreadsheet, every hand-off, who approves what, and the workarounds nobody wrote down. You get: A map of how your business really runs, a recommendation and a fixed price. The map's yours whether or not we build. 3. **Design** (A clickable prototype in your brand): Before anything is built, your team clicks through the system with your own words, your own steps and your logo, and tells us what's wrong. You get: A design you've signed off, so nobody is surprised later. 4. **Build** (On your real data, with you watching): We build it joined up with the tools you keep, and show you progress every week. Your team tries each piece as it's ready. You get: A weekly demo of real progress, never a big reveal at the end. 5. **Go live** (A piece at a time, nothing overnight): We move your data across, run the old way and the new side by side until your team is happy, and train everyone on their own work. You get: A system your team already knows how to use on day one. 6. **Run** (We stay, and keep improving it): We run it, back it up, keep it secure and watch for problems. Each month we agree what to improve next, as the business changes. You get: A system that keeps up with you, for one monthly fee. Rules we don’t break: we build around how your team works today; we explain every piece in plain English; we test on your real data; you own the system, its data and its documentation; where AI drafts or sorts something, a named person in your team approves anything that matters; and we run our own business on the same kind of system. ## Price A fixed price for each build, then a monthly fee to run it. It doesn’t have to be big: most clients start with one fix and add to it once it’s working. After the review we map your work and agree a fixed price in writing before we start. Once it’s live, a monthly fee covers running it, fixing it and improving it. No hourly billing. - Start with one fix: one process that hurts, joined up properly (£TBC). This is where most clients start. - One system: the core pieces joined up, with one view of every client and job (£TBC). - The whole business: every piece mapped, then built and switched on in stages (£TBC). ## A good fit if - Your team keys the same details into more than one place - Off-the-shelf software makes you change how you work - Someone owns how the business runs and wants it fixed - You’d rather have one partner who builds and stays ## Probably not a fit if - You need one small script an off-the-shelf tool already does - You want a chatbot on the website and nothing else - Nobody in the team has time to help us map the work If we’re not the right people, we’ll tell you who is. ## Contact - Book a review (a free 30-minute call with Seb or Mark): https://hellospark.co.uk/book-a-review - Email: seb@sparkcuriosityhq.com - Based in: Jersey, Channel Islands - Company: Spark Curiosity Ltd --- # Spark Learning: teams that use their systems, and AI, well A system only works if people use it. We run practical training built on your team’s own work: workshops, longer programmes, and Herd, our micro-learning app. ## Workshops: a day on your own work Half-day and full-day sessions for small groups, built on your team’s real tasks. We bring the exercises; you bring the inbox, the quotes and the spreadsheets. - Using your new system well, from day one - Using AI safely: what to paste, what never to - Each person leaves with one task they now do faster ## Programmes over several weeks Short sessions over several weeks, with practice in between, so new habits stick. We measure where the team started and where it ends up. - A baseline at the start and a measure at the end - Champions in your team who keep it going - A CPD record you can export ## Price - A half-day workshop: £TBC, for a half-day workshop. ## Herd Herd is our AI micro-learning app for people at work. A short lesson a day, on the job they’re already doing, so good habits build without anyone leaving the office for a course. Teams use it on its own or alongside a workshop or programme. ## Contact - Book a review (a free 30-minute call with Seb or Mark): https://hellospark.co.uk/book-a-review - Email: seb@sparkcuriosityhq.com - Based in: Jersey, Channel Islands - Company: Spark Curiosity Ltd --- # Work: case studies, written with our clients We’re writing up systems for engineering, trades and professional services clients with them now, and each one goes up here once the client has approved it. Each starts with the result, then the problem, what we built and what changed, in the client’s own words and numbers. Until then, ask us for the short version on a review: https://hellospark.co.uk/book-a-review ## Contact - Book a review (a free 30-minute call with Seb or Mark): https://hellospark.co.uk/book-a-review - Email: seb@sparkcuriosityhq.com - Based in: Jersey, Channel Islands - Company: Spark Curiosity Ltd --- # About Spark: two founders in Jersey who build the system a business runs on We design, build and run operations systems for growing businesses, teach teams to use them and to use AI well, and build our own products. The people you meet are the people who build it. ## Seb Lawson, Managing Director Seb has spent the last decade building and scaling technology companies: Zipline’s drone delivery in Silicon Valley, then JustPark in London, which he helped scale to exit. He trained as a mechanical engineer. Now based in Jersey, he also works with Digital Jersey helping island businesses adopt new technology. ## Mark McNally, co-founder and CTO Mark leads how every Spark system is built and run. ## What we believe - **Built around you**: We start from how your business works, not from a product we want to sell. The system fits the team, not the other way round. - **Plain English**: No jargon, no hype, no black boxes. If we can’t explain it simply, we haven’t designed it well enough. - **We stay**: A system is never finished on handover day. We run it, measure it and keep improving it. ## Jersey Based on the island, working across Jersey and the UK. Most of our clients found us through someone we’d already built for. Being on the island means we can sit with your team, watch how the work actually flows, and be back the next week when something needs changing. We run our own business on the same kind of system we build for clients. ## Contact - Book a review (a free 30-minute call with Seb or Mark): https://hellospark.co.uk/book-a-review - Email: seb@sparkcuriosityhq.com - Based in: Jersey, Channel Islands - Company: Spark Curiosity Ltd --- # Build your system A short, optional way to tell us what your business needs before a review: what you do, the tools you use now, the pieces you’d want in your system and what matters most. You can skip straight to booking at any step. What you tell us comes with your booking, so Seb or Mark arrive at the call knowing your business, and can show you what your system could look like. Start here: https://hellospark.co.uk/build ## The pieces you can choose - Enquiries and clients: every enquiry and client in one place. Hands each new enquiry to quoting with the details already filled in. - Quotes and proposals: built from your own price list or templates. A won quote becomes a job in one click, with no retyping. - Jobs and projects: who’s doing what, and what’s next. Finished work goes straight to invoicing, with the hours and costs attached. - Invoicing: invoices out the day the work’s done. Syncs with Xero or QuickBooks, so your accountant sees the same numbers. - Documents: letters, reports and contracts from templates. Pulls in the client and job details, then files the result in Drive or SharePoint. - Compliance records: certificates, checks and audit evidence. Records are created as the work happens, so the audit is mostly done already. - Onboarding: new clients and new starters, step by step. Feeds the client list and the team rota, so nobody starts without the basics. - Timesheets: hours logged against the job, on a phone. Hours land on the job and the invoice, and payroll gets one clean export. - Reporting: the numbers you check every Monday, live. Reads from every other piece, so the figures are never a week old. - AI helpers: drafts, summaries and sorting, checked by a person. Reads from the system, drafts the reply or summary, and a named person approves it. - Team training: short lessons on using the system, and AI, well. From Spark Learning: short lessons built on your own system. ## Contact - Book a review (a free 30-minute call with Seb or Mark): https://hellospark.co.uk/book-a-review - Email: seb@sparkcuriosityhq.com - Based in: Jersey, Channel Islands - Company: Spark Curiosity Ltd --- # Book a review: let’s map how your business works A free 30-minute call with Seb or Mark. Bring the spreadsheet you hate most. No pitch deck, and no follow-up sequence if it’s not for you. ## What happens on the call 1. **You tell us how the work flows**: from the first enquiry to the final invoice, and where it gets stuck. 2. **We ask the awkward questions**: which spreadsheet everyone depends on, and what happens when its owner is on holiday. 3. **You get an honest answer**: whether a system would help, where we’d start, and if we’re not the right people, who is. To book, you give your name and work email, then pick a time. If you’ve used the builder first, your answers come with the booking and we bring a demo built from them. Training enquiries use the same page. Book: https://hellospark.co.uk/book-a-review Rather write it down first? Email seb@sparkcuriosityhq.com --- # Security and your data Every system we build keeps your data in places we name, behind sign-ins you control, with backups we test and a record of who did what. This page sets out how we build and run it, what happens if something goes wrong, and how you’d take it all with you if you left. ## Your data lives where we say it does, and it’s encrypted there We keep your data inside the accounts you already run where we can, otherwise with an established cloud provider in the UK or EU, and it’s encrypted while it moves and while it’s stored. - Where it can, the system keeps your data inside the Microsoft 365 or Google Workspace account you already have, under the controls you already run. - Anything we host sits with an established cloud provider in the UK or EU, and we name the provider and the location in writing before we build. - Data is encrypted while it moves between systems and while it’s stored. - Passwords and keys for your connected tools are kept in a secure store that only the system can read, out of the code and out of everyday view. - Each system holds only the data it needs, and we document where your data flows so you can check it. - If you need data kept somewhere specific, tell us at the start and we’ll be straight about what that changes. ## Your data is backed up automatically, and we test the restore Backups run on a set schedule and sit apart from the live system for an agreed period, and we prove on a regular schedule that we can restore them. - Backups run automatically, on a schedule written into your agreement. - Backups are stored separately from the live system, so one problem can’t take out both. - We keep them for an agreed period, so we can go back to a point before a mistake was made. - We test restoring from backup on a regular schedule and keep a record of each test. - Where your data lives in Microsoft 365 or Google Workspace, we check what those services keep and add our own backup where they fall short. ## Only the right people get in, through single sign-on and multi-factor sign-in Your team signs in with the work accounts you already manage, with multi-factor sign-in on and a role for each person that sets what they can see and do. - Single sign-on with Microsoft or Google wherever your tools support it, so one work account controls access. - Multi-factor sign-in is on as standard, so a stolen password alone won’t get anyone in. - Each person has a role that decides what they can see and change, and you decide who has which. - When someone leaves, switching off their work account locks them out of the system too. - We review who has access with you on a regular schedule and remove what’s no longer needed. ## Every change is recorded, so you can show who did what and when The system keeps an audit trail of who created, changed, approved or deleted each record, and you can search it and export it. - Each record carries its history: who created it, who changed or approved it, and when. - Sign-ins, and changes to access and settings, are logged as well. - Everyday users can’t edit or delete the trail, so it stays a record you can rely on. - You can search it and export it for an auditor, a regulator or a dispute with a client. - The system keeps the records, and your business stays responsible for its own compliance. ## Every change is reviewed and tested before it reaches your live system We review every change, test it in a separate environment, release it in a controlled way and keep the software underneath up to date. - Every change to the code is reviewed by a second person before it’s released. - Changes are built and tested in a separate test environment, away from your live system and your live data. - Releases are planned and logged, and we can roll one back if something isn’t right. - We keep the software libraries your system depends on up to date and apply security fixes promptly. - We can work to your security questionnaire, and arrange independent penetration testing with a specialist firm on request. - We fix what testing finds and show you what changed. ## At Spark, only the people working on your system can reach it, and every access is logged Our own people get the least access the work needs, we log it, and confidentiality is in writing before we start. - Only the people working on your system get access, and only to the parts their work needs. - Our own accounts use multi-factor sign-in, and every access to your system is logged. - Access is removed as soon as someone no longer needs it. - Everyone at Spark, and anyone we bring in to help, is bound by a confidentiality agreement, and we tell you before anyone outside Spark works on your system. - We sign an NDA with you before we see anything sensitive, and a data processing agreement before we handle any personal data. ## AI only sees what the task needs, and a person approves anything that matters Where a system uses AI, we write down what it does, what data it sees and who checks it, and your data isn’t used to train anyone’s models without your written agreement. - We use AI providers on business terms that don’t use your inputs for training, and we check those terms for each tool. - Each AI step sees only the data its task needs, and that’s set out in your system’s documentation. - A named person in your team approves anything that matters before it goes out. - Very sensitive data can be kept out of AI steps altogether, or kept inside your own environment. - You can switch any AI step off. ## If something goes wrong, we tell you quickly and help you put it right We follow a written incident plan, keep you informed from the start and help you decide whether to report it to Jersey’s data protection regulator. - Our written incident plan covers finding out what happened, containing it, fixing the cause and keeping a record. - We tell you without undue delay once we know your system or data is affected, and keep you updated until it’s resolved. - If personal data is involved, we give you the facts you need to decide whether to report it to the Jersey Office of the Information Commissioner (JOIC), or the ICO in the UK, and help you prepare the report. - The law expects a reportable breach to reach the regulator within 72 hours of you becoming aware of it, where feasible, so we move fast. - Afterwards we write up what happened and what we’ve changed so it doesn’t happen again. ## If you leave, you take everything with you Your data, code and documentation are yours, and if you move on we hand them over properly and delete our copies. - You can export your data at any time, in formats other software can read, such as CSV or Excel. - Everything is documented in plain English, so another capable team could pick it up. - We build in accounts you own wherever we can, and transfer any we hold for you when you go. - On the way out we hand over the code, admin access and a full data export, and walk your new team through how it fits together. - Once the hand-over is done, we delete our copies of your data and confirm it in writing. - The notice period and hand-over steps are written into your agreement. The system keeps the records and controls your business needs; your business stays responsible for how it uses them, and for its own regulatory duties. ## Contact - Book a review (a free 30-minute call with Seb or Mark): https://hellospark.co.uk/book-a-review - Email: seb@sparkcuriosityhq.com - Based in: Jersey, Channel Islands - Company: Spark Curiosity Ltd --- # Questions: the straight answers ## Build or buy ### Should we build our own system or buy off-the-shelf software? Buy where a good product already fits how you work, and build where it doesn’t. Most of our clients do both: they keep Xero, Microsoft 365 or HubSpot, and we build the part that’s specific to them and join it all up. On a review we’ll tell you which side of the line you’re on, even if the answer is a product you can buy today. ### When should we just buy something off the shelf? When the job is standard and a product already does it well. Accounting, payroll and email are good examples: Xero or Microsoft 365 will beat anything built from scratch, and we wouldn’t suggest replacing them. Building pays off where the way you work is part of what makes you good, or where your team loses hours bridging tools that don’t talk to each other. ### Isn’t bespoke software expensive and risky? It used to be, and big projects still can be. We keep the risk small by starting with one fix, for a fixed price, built on your real data so you see it working before you commit to more. You own what we build, so you’re never stuck with us either. ### Why build now? What’s changed? Building has got much cheaper and quicker. The tools you already pay for now connect to each other properly, and modern development, AI included, lets a small team build in weeks what once needed a big team and a big budget. That puts a system built around you within reach of a growing business, where a few years ago only large companies could justify one. ### How does the cost compare with paying for software per user? Add both up over three years before you decide. Per-user software looks cheap until you count every seat, the add-ons, the extra tools that fill its gaps and the hours your team spends working around it. Our fees are a fixed price for each build and a monthly fee to run it, and neither is charged per user. You’ll still pay for some tools underneath, such as Microsoft 365, so we include those to keep the comparison honest. ### We’ve bought software before and the team never used it. Why would this be different? Because it’s built around how your team already works, so there’s less to learn and fewer reasons to go back to the spreadsheet. We map the work with the people who do it, they try each piece as we build it, and we change what doesn’t fit before it goes live. Once it’s live we watch how it’s used and sort out whatever gets in the way. ### Couldn’t we do this ourselves with Zapier or Power Automate? For a simple job, yes, and we’ll tell you if that’s all you need. Those tools are good at moving data from one place to another. They get fragile when the logic grows, when the whole team relies on them but nobody owns them, or when the one person who set them up leaves. We often use them inside a system, documented and monitored, so they keep working. ### Are we too small for a system like this? Probably not, if there’s a process that hurts and someone who wants it fixed. Most of our clients are growing businesses, and most start with one fix. If a spreadsheet still does the job for you, we’ll say so and save you the money. ## How it works ### What does a Spark system actually include? Whatever your business runs on, joined up in one place. That’s usually some mix of CRM, quotes, invoicing, job tracking, compliance records and onboarding, plus the automations that remove the copy and paste between them. We build only the parts you need, around how your team already works. ### What are the steps, from first call to going live? Four steps: map, design, build and run. We sit with your team and map how the work really flows, design one system around that map with a fixed price, build it on your real data, then run it and keep improving it. You sign off each step before the next one starts, and you keep the map whether or not we build. ### How long does it take? A first fix usually goes live within weeks, and a bigger system goes live in stages over a few months. You get a plan with dates alongside the fixed price, before we start. Because it goes live a piece at a time, your team is using the first parts long before the last ones are finished. ### What do you need from our team, and how much of their time? One person who owns the project and can make decisions, plus time from the people who do the work. Most of that time comes at the start, in the mapping sessions, then it’s short reviews and testing as each piece is built. We put whose time we need, and roughly when, in the plan before you sign. Nobody needs to be technical. ### Do we have to replace the software we already use? Usually not. We start from the tools you already own and pay for, keep what works and join it up. Where a tool really doesn’t fit how you work, we’ll tell you why and what replacing it would cost before anything changes. ### Which tools can you connect? Most of the ones growing businesses use: Microsoft 365, Outlook, Teams, SharePoint, Excel, Google Workspace, Xero, QuickBooks, Sage, HubSpot, Salesforce, ServiceM8, Cliniko, Clio, DocuSign, Stripe and plenty more. If a tool has an API (a way for other software to connect to it) or lets you export your data, we can usually join it up. If yours is less common, name it on the review and we’ll check before you commit to anything. ### Can you bring our old data across? Yes. We move what you need from spreadsheets, old systems and shared drives, tidy it up on the way and check it with you before the old version is switched off. Some old records are better left where they are, read-only, with a link from the new system, and we’ll tell you when that’s the cheaper and safer choice. ### Will everything change on one day? No. The system goes live a piece at a time, so your team gets used to one part before the next arrives and nothing changes overnight. Where it helps, the old way keeps running alongside until the new piece has proved itself on real work. ### How will our team learn to use it? Training is part of every build. We train people on their own work in the system, in short sessions before each piece goes live, and leave plain-English guides they can come back to. When you take on new starters later, we can train them too. ## Cost and contracts ### How much does a first system cost? A first system starts from £TBC, then a monthly fee to run and improve it. What you pay depends on what we build, and you get it as a fixed price in writing before any build work starts. ### Why do you suggest starting with one fix? Because it’s the cheapest way to find out whether we’re right for you. We pick the process that hurts most, such as enquiries that go missing or quotes that take days, and fix it properly. You see the result on real work, and if you want more, the next piece builds on the first. ### What does the fixed price cover? Everything it takes to get the agreed scope live: mapping and design, the build, testing on your real data, training for your team and the documentation. It’s written down before we start, and there’s no hourly billing. If it takes us longer than we planned to build what we agreed, that’s our cost to carry. ### Are there any other costs? Only ones you see before you commit. Once a system is live there’s the monthly fee, plus any third-party running costs such as software subscriptions or AI usage. We size those up front and show them next to our price, so the bill is never a surprise. ### What if we want to change something halfway through? We price the change before any work starts on it, and you decide whether it’s worth it. Adjustments that come out of testing what we agreed are part of the build. Nothing new gets built and billed without your sign-off. ### What does the monthly fee cover? Running the system, keeping it healthy and making it better. That means monitoring, support, fixes, backups, keeping the connections working when your other tools change, and small improvements as your business grows. Bigger pieces of work get their own fixed price, agreed before we start. ### Can we stop paying the monthly fee? Yes. The notice period is in your agreement, and if you leave we hand over everything you need to run the system yourselves or with another supplier. If we’re doing our job, you won’t want to. ## Security and data ### Is our data safe with you? Yes, and we’ll show you how instead of asking you to take it on trust. Every system gets the same basics: multi-factor sign-in, access by role, encryption, backups we test and a record of who did what. We sign an NDA before we see anything sensitive, and our security page sets out the detail. ### Where will our data be stored? Where possible, inside the Microsoft 365 or Google Workspace account you already have, under the controls you already run. Anything we host for you sits with an established cloud provider in the UK or EU, and we name the provider and location in writing before we build. If you need data kept somewhere specific, tell us at the start and we’ll be straight about what that changes. ### Is it backed up, and can you get it back if something goes wrong? Yes. Your data is backed up automatically, kept for an agreed period and stored separately from the live system. We test restoring it on a schedule, because a backup nobody has ever restored is only a hope. How often we back up and how long we keep it are written into your agreement. ### Who can see our data? Only the people who need to. In your team, each person has a role that decides what they can see and change, and you decide who has which. At Spark, only the people working on your system have access, every access is logged, and it’s removed when they no longer need it. ### Can our team sign in with their Microsoft or Google accounts? Yes, where your tools allow it, and it’s what we recommend. Single sign-on means one work login your IT already manages, so when someone leaves, switching off their account locks them out of the system too. Multi-factor sign-in is on as standard, so a stolen password alone won’t get anyone in. ### Is our data encrypted? Yes, both while it’s moving between systems and while it’s stored. Connections are encrypted, stored data is encrypted by the services it lives in, and the passwords and keys for your connected tools are kept in a secure store that only the system can read. We’ll give your security team the technical detail on request. ### Can we see who changed what? Yes. The system keeps an audit trail of who created, changed, approved or deleted each record, and when. You can search it and export it, which helps when a client disputes something or an auditor asks how a decision was made. ### Can we have the system penetration tested? Yes. We can arrange independent penetration testing with a specialist firm, before go-live or on a schedule that suits you, and you get the report. We fix what it finds and show you what changed. The tester’s fee is quoted separately, so you can decide whether your risk calls for it. ### Will you fill in our security questionnaire? Yes. We can work to your security questionnaire or supplier due diligence and answer it plainly, based on how your system is actually built and run. Where a question needs evidence, we show you the evidence. If something you need isn’t in place yet, we’ll say so and tell you what it would take. ### What happens if there’s a data breach? We tell you quickly, contain it and help you deal with it. We follow a written incident plan: find out what happened, stop it spreading, fix the cause and keep a record. If personal data is involved, we give you what you need to decide whether to report it to the Jersey Office of the Information Commissioner (or the ICO in the UK), and to do it in time. ## Ownership and support ### Who owns the system and the data? You do. Your data is always yours, and the system we build for you, including its code, configuration and documentation, is yours too, as set out in your agreement. The tools underneath, such as Xero or Microsoft 365, stay under their own licences. ### You’re a small company. What happens if you disappear? We set every system up so you could carry on without us. It runs in accounts you own wherever we can, the code sits where you can get to it, and it’s built with widely used tools that plenty of developers know. Everything is documented in plain English, so another capable team could pick it up, and you can export your data at any time. ### What if we want to move to another supplier? You can, and we’ll help. We hand over the code, the documentation, admin access to every account and a full export of your data, then walk the new team through how it fits together. The notice period and hand-over are written into your agreement, so there’s nothing to negotiate on the way out. ### What happens after the system goes live? We stay. For a monthly fee we run the system, watch it, fix what breaks and keep improving it as your business changes. You get named people to contact, with support hours and response times written into your agreement and quicker help for anything that stops your team working. ### How do you know if something breaks? We monitor every live system and get an alert when a connection drops, an automation fails or something looks wrong, so we can often fix it before it holds anyone up. When something does affect you, we tell you what happened, what we did and how we’ll stop it happening again. ### Do you guarantee uptime? We agree uptime and response commitments with you in writing, based on how your system is set up. Your system runs on providers such as Microsoft, Google or a cloud host, so their commitments set the ceiling, and we tell you what they are. Our part is watching it, responding quickly and keeping you posted until it’s fixed. ### What happens when Xero or Microsoft change something? We deal with it, as part of the monthly fee. The tools you use update their software all the time, so we watch for changes that affect your connections and update the system to match. It’s one of the main reasons a system needs someone running it after handover day. ## Jersey and regulated firms ### Does it work under Jersey data protection law and UK GDPR? It’s built to help you meet both, but the responsibility stays with your business, and no system can take that on for you. In plain terms, you decide what personal data you hold and why, which makes you the controller, and we handle it only on your instructions, which makes us the processor. The Data Protection (Jersey) Law 2018 and UK GDPR work in much the same way, so the system is set up to show what you hold, why, who can see it and who changed it. ### Will you sign a data processing agreement? Yes, always, before we handle any personal data for you. It sets out what we do with your data, where it’s kept, who else is involved, how we keep it secure and what happens to it when we finish. You stay the controller, and we only act on your written instructions. ### Can you help us with subject access and erasure requests? Yes. When someone asks for a copy of their data, or asks you to delete it, the system can find everything it holds on that person and export or remove it, with a record of what was done. You decide how to respond, and we help you pull it together within the legal time limit. ### We’re regulated by the JFSC. Can we use you? Yes, and we’ll support the oversight your obligations ask of you. Using us may count as outsourcing under the Jersey Financial Services Commission’s rules, so you may need to assess us, keep records of the arrangement and decide whether it’s material enough to tell the JFSC. We give you what you need for that: our agreement, how we handle and secure your data, who has access, how you can check our work and how you’d leave. You stay responsible for the outsourced work, and the system’s records make your oversight easier to show. ### Will the system keep us compliant? The system keeps the records you need, but your business stays responsible for compliance. We build in a named owner for each process, a log of who did what and when, and a person’s sign-off wherever it counts, so you can show a regulator or auditor how the work was done. We won’t tell you a system makes you compliant, because no system can. ### Can an auditor or regulator see how a decision was made? Yes. Every record carries its history: who created it, who changed or approved it, when, and the documents attached along the way. You can export that trail in a form an auditor can read, so preparing for a visit means pulling a report instead of rebuilding the story from old emails. ## AI ### Where do you use AI in the system? Only where it removes real work, and we tell you exactly where. Typical jobs are reading documents and pulling out the details, drafting a first version of an email or quote, sorting incoming enquiries and summarising long threads. Each one is written into your system’s documentation, with what it does, what it sees and who checks it. ### What data does the AI see? Only what the task needs. If it’s drafting a reply to an enquiry, it sees that enquiry and the details needed to answer it, and nothing else in your system. We list what each AI step can see, and very sensitive data can be kept out of AI steps altogether. ### Will our data be used to train AI models? Not without your written agreement. We use AI providers on business terms that don’t use your inputs for training, and we check those terms for each tool and write down what we find. If your requirements are stricter, sensitive data can stay inside your own environment and never reach an outside AI model. ### What if the AI gets something wrong? A person in your team approves anything that matters before it goes anywhere. The AI drafts and sorts, and your people decide. We design each step so a mistake is easy to spot and quick to fix, and we keep track of how often the AI’s work gets corrected so we can improve it. ### Can we have a system with no AI in it? Yes. Most of what makes a system useful is plain automation: details entered once, sent to the right place and checked. We only add AI where it earns its keep, and you can say no to any part of it. ### We already pay for Copilot or ChatGPT. Can you use those? Often, yes. If you have Microsoft 365 Copilot or a business ChatGPT plan, we can design around it so you get more from what you already pay for. We check what your licence allows and whether its data terms suit the job before we build on it. And if your team isn’t getting much out of it yet, that’s what Spark Learning is for. ## Working with us ### Who will we actually deal with? The people who build your system. Seb Lawson, our Managing Director, and Mark McNally, our co-founder and CTO, lead every project, and you talk directly to whoever is doing the work. There’s no account manager passing messages along. ### Where are you based, and will you come to us? We’re based in Jersey, and for clients on the island we’re happy to come to you. Mapping works best in person, sitting with the people who do the work, while reviews, check-ins and most of the build happen remotely on Teams, Google Meet or Zoom. Because so much is remote, we also work with businesses in Guernsey, the UK and further afield where the fit is right. ### What kind of business do you work with? Growing businesses that have outgrown spreadsheets, inboxes and tools that don’t talk to each other. Most of our clients are engineering firms, trades and professional services businesses, along with some finance firms. If your team spends its week copying information from one place to another, we’re probably a good fit. ### Are you an IT support company? No. We don’t fix laptops or look after your network, and we’re happy to work alongside whoever does. We design, build and run the system your business runs on, and we agree access, accounts and security with your IT provider so nothing falls between us. ### How do we get started? Book a review. It’s a free 30-minute call with Seb or Mark: you tell us where the work gets stuck, and we tell you honestly whether there’s a quick fix, a system worth building or nothing worth doing yet. There’s no obligation and no deck. ## Learning ### What is Spark Learning? Training that gets your team using AI and their systems well. We run workshops and longer programmes built on your team’s own work, so people practise on real tasks instead of generic examples, and we measure what changes. ### What’s the difference between a workshop and a programme? A workshop is a half or full day that gets your team doing something new by the end of it. A programme runs over several weeks, with short sessions and practice in between, so new habits stick. Programmes measure where the team started and where it ends up, and come with a CPD record you can export. ### How much does a workshop cost? A half-day workshop for up to 12 people starts from £TBC. Full-day workshops and longer programmes are priced to what you need, and you get a written price before anything is booked. ### Can you train our team if you didn’t build our systems? Yes. Our workshops work just as well for teams on Microsoft 365, Google Workspace and everyday AI tools, with no Spark system involved. We build each session on your team’s real tasks, so it fits whatever you run on. ### What is Herd? Herd is our AI micro-learning app for people at work, with short lessons that fit around the working day. Teams can use it on its own or alongside a Spark Learning workshop or programme. ## Contact - Book a review (a free 30-minute call with Seb or Mark): https://hellospark.co.uk/book-a-review - Email: seb@sparkcuriosityhq.com - Based in: Jersey, Channel Islands - Company: Spark Curiosity Ltd --- # The State of AI & Automation Our research into how businesses across the UK and Channel Islands are putting AI and automation to work, and what the ones doing it well do differently. The full report is free to read online. Read the full report: https://soai.sparkconsulting.tech/ ## What’s inside - **How fast are businesses moving on AI and automation?**: Where the real momentum is building, across the UK and the Channel Islands. - **What do the people doing it report?**: Patterns from contributors across both, put in a form you can act on. - **Where is it working, and where does it stall?**: Real examples of where automation and AI earn their keep, and where they get stuck. - **What should a business do next?**: Practical takeaways on where to focus first, without the hype. --- # How an AI reads this site AI assistants don’t see our colours, motion or layout. They read the words, the structure and the links, and many read a plain-text summary a site publishes for them. So we publish one, and keep it in step with the pages people see. ## What we publish for AI assistants - https://hellospark.co.uk/llms.txt: a short map of the site, written for AI assistants - https://hellospark.co.uk/llms-full.txt: every page and every article in one plain-text file - A markdown version of every page and article: add `.md` to the address, for example https://hellospark.co.uk/systems.md - Structured data (schema.org) describing who we are, where we are and what we do - Pages rendered on the server, so the words are there before any script runs - A robots.txt that welcomes AI crawlers: https://hellospark.co.uk/robots.txt Want to know how AI assistants describe your business? We’ll check it on a review: https://hellospark.co.uk/book-a-review --- # Articles # Build or buy: buy for the jobs every business does, build for how yours works Published 16 November 2026 by Spark · Build or buy · 5 min read Canonical: https://hellospark.co.uk/writing/build-or-buy-bespoke-system > Buy software for the jobs every business does the same way. Build where your work is different, or where people spend their week bridging tools that don’t fit. Most growing businesses need both: good bought tools, and one system that joins them up. Buy off-the-shelf software for the jobs every business does the same way: email, payroll, the books, file storage. Build where your work is genuinely different, or where your team spends its week carrying information between tools that don’t fit. Most growing businesses need both, and the best setup we see keeps the good bought tools and builds one system around them that joins them up. The mistake is treating it as one big decision. It’s really a series of small ones, one for each job the business does, and the hidden costs sit in different places on each side. ## Buy when the job is the same in every business Some work is identical whether you run a heating firm or an accountancy practice. Sending email, running payroll, keeping the ledger, storing files. Microsoft 365, Google Workspace and Xero have had years and a huge number of users to get those jobs right. You won’t build a better ledger than Xero, and you shouldn’t pay anyone to try. The test is simple. If you could swap your business for any other and the job wouldn’t change, buy. The vendor carries the cost of keeping it current, secure and up to date with the rules, and that’s worth the subscription. ## Build when the work is how you win Then there’s the work that’s yours. How you price a job. The checks you run before an engineer goes to site. The way you take on a new client, who needs to see it and what they need to know. That’s often why customers pick you, and it rarely matches what a product designer somewhere else imagined. When the software doesn’t fit, the business bends around it. People keep a spreadsheet alongside the official system. Someone exports a report every Friday and fixes it by hand. The CRM says one thing and the job board says another. None of this arrives as an invoice, which is why it’s easy to miss. ## The hidden costs of buying are seats, workarounds and retyping The subscription is the number everyone sees. The costs that hurt are the ones around it. - **Per-seat licences.** Most business software charges per user per month. As a worked example, a tool at £30 a user a month across 15 people is £5,400 a year, and it goes up every time you hire. Add a second tool for the bit the first one can’t do and you pay twice for the same people. - **Workarounds.** Every gap between what the product does and what you need gets filled by a person: a side spreadsheet, a manual check, a weekly export. Those hours are real. They just don’t show up on a bill. - **Retyping.** When tools don’t talk to each other, the same customer name, address and job details get keyed into three places. Each time takes a few minutes and is a chance to get it wrong. We explain why that’s the best place to start in [spreadsheets aren’t the problem, retyping is](/writing/retyping-is-the-problem). - **Lock-in.** Your data lives in the vendor’s format, on their terms. Prices rise, features move to a higher tier, and leaving means an export that loses half the history. ## The hidden costs of building are understanding, upkeep and dependence Building has its own bill, and anyone who says otherwise is selling something. - **Someone has to understand your business first.** A bespoke system is only as good as the map it’s built from. Skip that step and you get expensive software that fits nobody. - **Upkeep.** The tools you connect to keep changing. Xero changes how other software connects to it, Microsoft changes how sign-in works, a supplier changes their price file. A bespoke system needs someone watching it and fixing it. That’s why we charge a monthly fee to run and improve it after the build, instead of handing it over and walking away. - **Dependence on the builder.** If only one person understands how it works, you’ve swapped vendor lock-in for builder lock-in. The fix is to own the code, the data and the documentation from day one, which we cover in [owning your system](/writing/owning-your-system). ## Ask four questions about each job Take each job your business does, from first enquiry to final invoice, and ask: 1. Does this work the same way in every business like ours? If so, buy. 2. Does a tool we already pay for do it well enough? If so, keep it. 3. Is someone retyping, exporting or checking by hand to make it work? If so, that gap is worth building for. 4. Would doing this our way win us work or save real hours every week? If so, build. In our experience the answer is usually a short list of bought tools and one thing to build: the layer that joins them up and handles the work that’s yours. ## Building got cheaper, so the sums have changed A few years ago, building anything bespoke meant a long project and a budget most small firms couldn’t justify, so buying was the only sensible answer for a 15-person business. That’s no longer true. Ready-made components, AI-assisted building and patterns that have been proven many times have brought the cost of building down a long way. The thinking hasn’t got any cheaper, and that’s now where the real work sits. We explain what changed in [why bespoke software is now within reach](/writing/why-bespoke-is-affordable-now). So if you asked this question a few years ago and the answer was “buy”, it’s worth asking again. ## Our recommendation: keep what works and build the joins Don’t rip out tools your team knows and likes. Keep Microsoft 365 or Google Workspace, keep Xero, keep the job app your engineers use on site if it works for them. Build one system around them that joins them up and does the work no product does your way. That’s what we do, for a fixed price, and then we stay to run it. If you’re weighing up a single automation rather than a whole system, our older piece on [build, buy or no-code](/writing/build-buy-or-no-code) covers that choice. If you want a straight answer for your business, [book a review](/book-a-review). It’s a free 30-minute call with Mark or me. Tell us what you use and where the work gets stuck, and we’ll tell you what to keep, what to replace and what, if anything, is worth building. If you’d like to sketch it out first, you can [build your system](/build) in a few minutes. --- # Bespoke software is now within reach of small businesses: what changed, and what didn’t Published 13 November 2026 by Spark · Build or buy · 5 min read Canonical: https://hellospark.co.uk/writing/why-bespoke-is-affordable-now > A bespoke system used to need a big budget and a long project. Ready-made parts, AI-assisted building and proven patterns have cut the cost of building. What hasn’t changed is that someone still has to understand how your business works. A bespoke system is now a sensible option for a business of 10 or 20 people, which it wasn’t a few years ago. The building got much cheaper: the parts are better, AI writes a lot of the routine code, and builders who do this often reuse what they know works. What didn’t get cheaper is understanding your business, and that’s now where most of the cost, and most of the value, sits. ## Bespoke used to mean starting from nothing Not long ago, a custom system meant a team writing everything themselves: the database, the sign-in, the screens, the connections to your accounts software, the servers it ran on. Months of work before anyone saw a screen they could use. The price reflected that, and so did the risk. Plenty of those projects ran late and over budget, which is how bespoke software got its reputation. So small firms bought. Even when the software didn’t fit, a subscription was cheaper than a project, and you could start on Monday. ## The parts now come ready-made Most of what a business system needs has been built many times over, and the pieces are now well tested and cheap to use. - **Sign-in.** Nobody builds a login screen from scratch any more. The system uses your existing Microsoft 365 or Google Workspace accounts, so the multi-factor sign-in you already have comes with it. - **Databases and hosting.** Managed services from the big cloud providers look after the servers, the security patches and the storage that used to need a dedicated person. - **Connections.** Xero, HubSpot, Microsoft 365, Google Workspace and ServiceM8 all publish documented ways for other software to read and write their data. Joining tools up used to mean custom work for each one. Now it starts from a connection that already exists, and the work is in getting the details right. - **Screens.** Forms, tables, calendars and layouts that work on a phone come from widely used component libraries. That means the effort goes into the part that’s yours: how a job moves through the business, who approves what, and which details need to travel where. ## AI writes the routine code, and a person checks every line AI is how we build, not what we sell. Coding tools have improved quickly. They write routine code, such as a form, a report or a connection to Xero, much faster than a person typing it, and they’re good at writing the tests that check it works. What they can’t do is decide what to build. They don’t know that your office manager checks every quote over a certain amount, or that your engineers lose signal in basements and need the job details saved on the phone. An experienced developer still decides what gets built, reviews the code, tests it on your real data and answers for the result. The time saved shows up as a lower price and a shorter build. ## Proven patterns mean we rarely start from a blank page Growing businesses differ in the details and are surprisingly similar in shape. Enquiry, quote, job, invoice, payment. Client take-on, checks, engagement letter, the work itself, sign-off. Once you’ve built a quote-to-invoice flow properly a few times, you know where the awkward cases are: the part-payment, the variation, the credit note, the client who wants one job split across two sites. A builder who does this regularly brings those patterns along, tested and documented, and adapts them to you. You pay for fitting them to how you work, which is quicker and cheaper than inventing them again. ## Fixed fees move the risk off the client The old model was hourly billing on a project nobody could size, so the client carried the risk of every overrun. When building is faster and more predictable, a builder can price the agreed scope up front and stand by it. That’s how we work: a fixed price for the build, then a monthly fee to run and improve the system. How that works in practice, changes included, is in [why a fixed price is fairer](/writing/fixed-price-software). ## Someone still has to understand your business This is the part that doesn’t get faster. Before anything is built, someone has to sit with your team and follow the real work: where enquiries come in, who picks them up, the spreadsheet everyone relies on, the step that only happens because one person remembers to do it. AI can’t do that for you. It can help write up the notes and draw the diagram, but it can’t tell which of the three ways your team quotes is the one that wins work. That takes people who ask good questions and listen to the answers. We describe how we go about it in [how we map a business](/writing/how-we-map-a-business). Getting this wrong is still the main way bespoke projects fail. Building quickly from a bad map just gets you the wrong system sooner. ## It’s worth asking the question again If you looked at bespoke software a few years ago and decided it was out of reach, look again. The honest comparison now is a fixed-price build and a monthly running fee against what you already pay: the licences for tools that half fit, the hours spent retyping, the mistakes that come from copying details between systems. For some businesses the answer is still to buy, and we’ll say so. For businesses whose work doesn’t fit the boxes, a system built around them is now a realistic option. Our guide to [build or buy](/writing/build-or-buy-bespoke-system) sets out how to tell which you are. ## Our recommendation: price the problem before you price the system Before you ask anyone for a quote, write down what the current way of working costs you each month: the subscriptions, the hours spent copying and checking, and the jobs or invoices that slip. That number is what any system has to beat. If it’s small, keep what you have. If it isn’t, a bespoke system is worth a proper look. If you’d like to know what a system for your business would involve, [book a review](/book-a-review). It’s a free 30-minute call with Mark or me, and we’ll tell you honestly whether building makes sense. Or [build your system](/build) first and we’ll come to the call with a view. --- # What one system changes: each step hands the next one what it needs Published 10 November 2026 by Spark · Playbooks · 6 min read Canonical: https://hellospark.co.uk/writing/what-one-system-unlocks > Joining up your tools means each step of a job hands the next one what it needs, so nobody retypes, chases or hunts for the latest version. We follow two jobs end to end to show what that looks like. Both are worked examples, not client results. When your tools are joined up, each step of a job hands the next one exactly what it needs. Nobody retypes the address, nobody chases the engineer for their hours, and nobody asks which spreadsheet is the current one. The clearest way to see what that changes is to follow a job from the first message to the money arriving, so below are two: an emergency plumbing call-out and a new accountancy client. Both are worked examples, built from the kind of work we see every week. They aren’t results from a named client. ## A burst pipe goes from WhatsApp to paid without anyone retyping it It’s 7.40pm on a Tuesday. A customer in St Clement sends a WhatsApp message to a plumbing firm’s business number: water coming through the kitchen ceiling, photo attached. **Without a joined-up system**, the owner sees it on his phone over dinner. He rings the customer back, writes the address on the back of an envelope and phones round to find out who’s on call. The engineer goes out, fixes it and texts a photo of the parts he used. Two days later someone in the office types up a job sheet from the texts, works out the hours, creates the invoice in Xero and emails it. The customer pays a fortnight later, after one reminder. The photos are still on the engineer’s phone, so if the insurer asks for them in March, someone has to go looking. **With one system**, the same evening goes like this. 1. The WhatsApp message lands in the system as a new enquiry, with the customer’s number, the photo and the time. If they’re an existing customer, their address and history come up with it. 2. The system shows who’s on call tonight. The owner taps once to send the job to that engineer, and the customer gets a reply with the engineer’s name and when to expect him. 3. The engineer opens the job in ServiceM8 on his phone. The address, the photo and the customer’s notes are already there. 4. On site, he logs his hours and the parts from the van, adds before and after photos, and the customer signs on the screen. 5. When he marks the job complete, the invoice is raised in Xero from the job, with the parts priced from the firm’s price list and GST added, and sent to the customer with a link to pay by card. 6. The customer pays that night or the next morning. Xero matches the payment and the job closes. 7. The photos, the signature and the notes are filed against the customer and the property, so the insurer’s question in March takes a minute to answer. Nobody typed the address twice or rebuilt a job sheet from text messages, and the invoice went out the same evening the work was done. ## The tools stayed the same and the gaps between them closed WhatsApp, ServiceM8 and Xero are all still there. What changed is that they now pass the job from one to the next, with a record at each step. For the owner, that means cash comes in sooner because invoices go out when the work is finished, and Thursday afternoons are no longer spent typing up the week’s job sheets. It also means there’s a complete record of every job if a customer disputes a bill or an insurer asks questions. ## A new accountancy client goes from ID checks to signed-off accounts in one record A small accountancy practice in St Helier takes on a new client: a Jersey company with two directors that needs annual accounts and a tax return. **Without a joined-up system**, the partner emails the directors for ID and proof of address. The documents trickle in over three weeks, some by email and one by post. Someone saves them to a shared drive and ticks a column in the take-on spreadsheet. The engagement letter is a Word template, edited by hand and emailed for signature. The client’s details are typed into the practice’s client list, then into Xero, then into the tax software. At year end the team emails for bank statements and invoices, chases twice and builds the accounts from whatever arrives. The partner reviews a printed pack, and the final sign-off is a scanned signature on a PDF that someone will have to find again next year. **With one system:** 1. The enquiry comes in through the website and creates a client record. The client types their details once. 2. Each director gets a secure link to upload ID and proof of address and answer the due diligence questions the practice needs for its anti-money-laundering checks. The system tracks what’s in and what’s missing, and sends polite reminders so the team doesn’t have to. 3. A named person in the practice reviews the documents and records their decision, with their name and the date. Nothing moves on until that’s done. 4. The engagement letter is generated from the practice’s own template with the client’s details filled in, and sent for e-signature. Once signed, it’s filed against the client. 5. The client record flows into Xero and the practice’s other tools, so nobody retypes the company name, number or year end. 6. At year end, the system sends the client a checklist of what’s needed, tracks what’s arrived and chases what hasn’t. The team can see at a glance which clients are ready to start. 7. The draft accounts go to the partner for review inside the system, with comments and changes recorded. The client signs off online, and the signed accounts are filed with everything else for that client and that year. ## The practice can see where every client is without asking The biggest change is visibility. Which ID checks are outstanding, which engagement letters are unsigned, which year ends are waiting on the client: it’s all on one screen instead of in people’s heads. The second change is the record. When the practice’s anti-money-laundering checks are reviewed, the evidence of who checked what and when is already there, created as the work happened. The system keeps that record. The practice stays responsible for the checks themselves. ## The value is in the hand-offs Both examples use tools the businesses already had. The difference is in the hand-offs, the moments where work passes from one person or tool to the next. That’s where details get retyped, where jobs sit in someone’s inbox and where things go missing. A joined-up system is mostly a set of good hand-offs, each with a record of what happened. ## Start with the hand-off that hurts most You don’t need to join up everything at once. Pick the hand-off that costs you most, which is usually wherever the same details get typed twice, and fix that first. We explain how to find it in [spreadsheets aren’t the problem, retyping is](/writing/retyping-is-the-problem), and how to switch over in stages in [going live a piece at a time](/writing/going-live-without-the-big-bang). If you’d like to see what your own version of these examples would look like, [book a review](/book-a-review). It’s a free 30-minute call, and we’ll walk through one of your jobs from first message to paid. You can also [build your system](/build) first to show us what you use today. --- # A bespoke system is as secure as its builder’s answers to eight questions Published 6 November 2026 by Spark · Security and data · 5 min read Canonical: https://hellospark.co.uk/writing/is-a-bespoke-system-secure > A bespoke system can be as secure as anything you buy, but only if the people building it can answer eight plain questions well. Here they are, from where your data lives to what happens with AI, with what a good answer sounds like. A bespoke system can be every bit as secure as software you buy off the shelf, and often more so, because it holds only the data you need and only your people can get in. But that depends entirely on how it’s built and run. Before you let anyone build yours, ask them the eight questions below. A careful builder will answer each one plainly, in writing, and won’t mind being asked. ## Know exactly where your data will live **Ask:** where will the system be hosted, and where will our data be stored? A good answer names names: the provider, the country or region, and every other service that sees any of your data. Often the best place for your documents and records is the Microsoft 365 or Google Workspace account you already have, under controls you already run. Anything the builder hosts for you should sit with an established cloud provider in a location you’ve agreed, written down before building starts. For Jersey businesses this ties into data protection. If personal data leaves the island, you need to know where it goes and on what terms. We cover that in [the Data Protection (Jersey) Law and your systems](/writing/data-protection-jersey-systems). A vague answer, such as “it’s in the cloud”, is a warning sign. ## Backups only count if someone has tested a restore **Ask:** how is it backed up, and when did you last restore a backup to check it worked? A backup nobody has ever restored is only a hope. You want to hear how often the data is backed up, how long copies are kept, that they’re stored separately from the live system, and that restores are tested on a schedule. Also ask how long it would take to get you working again if the live system were lost. A good answer gives a frequency, a retention period and a date, and those details belong in your agreement. ## Each person should see only what their job needs **Ask:** who can see and change what inside the system? Your engineers need their jobs and their customers’ addresses. They have no reason to see payroll or the partners’ notes. A good system gives each person a role that decides what they can see and change, set up to match how your team actually works, and you decide who has which role. Ask what happens when someone leaves. The answer you want is that switching off one account removes their access to everything, the same day. ## Your team should sign in with the accounts they already have **Ask:** can we use single sign-on with our Microsoft or Google accounts, with multi-factor sign-in? Your team shouldn’t need yet another password. A well-built system lets people sign in with the work accounts you already manage, through Microsoft Entra ID for Microsoft 365 or through Google Workspace, so your existing multi-factor sign-in applies. Stolen and reused passwords are one of the commonest ways into a business’s systems. With multi-factor sign-in on, a password alone won’t get anyone in, and with single sign-on there’s one account to switch off when someone leaves. ## Data should be encrypted while it moves and where it’s stored **Ask:** is our data encrypted in transit and at rest, including the backups? Encryption in transit scrambles data as it travels between your browser and the system, so it can’t be read along the way. Encryption at rest scrambles it where it’s stored, so a copied disk or backup is useless to whoever has it. Both should be a straight yes. Also ask where the passwords and keys for your connected tools, such as your Xero connection, are kept. They should sit in a secure store that only the system can read, never in a spreadsheet or an email. ## Every important change should leave a trail **Ask:** can we see who did what, and when? An audit trail records the changes that matter: who created a record, who changed a price, who approved an invoice, who deleted a file. Ask what gets logged, how long logs are kept, and whether you can search and export them yourselves. For regulated firms, this is how you show an auditor or reviewer how the work was done. For everyone else, it settles “who changed this?” in a minute. ## You should know who at the builder can see your data **Ask:** who at your company can see our data, and how do we know? People forget this one. The builder needs some access to build, fix and support the system. A good answer names a small number of people, explains why each needs access, confirms that every access is logged, and says access is removed when they no longer need it. They should sign a confidentiality agreement before seeing anything sensitive, and you should have a data processing agreement setting out what they can do with personal data on your behalf. We sign both before we see anything sensitive. Ask too what happens to that access if you part company. It should end, and you should be able to end it yourself. There’s more on that in [owning your system](/writing/owning-your-system). ## You should know what the AI sees and who checks it **Ask:** where does the system use AI, what data does it see, and is that data used to train anyone’s model? If the system uses AI to draft, summarise or sort, the builder should tell you which provider, which data each step can see and what the provider’s terms say about training. Business terms from the major AI providers generally don’t use your inputs to train their models, but the builder should have checked the specific terms for each tool and written down what they found. Very sensitive data can be kept out of AI steps altogether. Then ask who checks the output. The answer you want is that a named person in your team approves anything that matters before it reaches a client. ## Ask about practices as well as certificates Some builders hold certifications such as Cyber Essentials or ISO 27001, and they’re worth asking about. A certificate tells you a process was in place when someone checked. The eight answers above tell you how your system will be built and run day to day. If your risk calls for it, you can also have an independent specialist test the system before it goes live and give you the report. ## Our recommendation: get the answers in writing before you sign Put these eight questions to anyone who wants to build your system, including us. Ask for the answers in writing, and make the important ones, such as backups, access and where data lives, part of the agreement. A builder who answers clearly and specifically is far more likely to build carefully. If you’d like to ask us the eight questions directly, [book a review](/book-a-review). It’s a free 30-minute call with Mark, our CTO, or me, and we’ll answer each one for the system you have in mind. --- # Owning your system: if you part ways with your builder, you keep everything Published 3 November 2026 by Spark · How we work · 5 min read Canonical: https://hellospark.co.uk/writing/owning-your-system > If you stop working with the people who built your system, you should keep the code, the data and the documentation, and be able to hand them to someone else. That belongs in the contract from day one. Here’s what to ask for. If you stop working with the people who built your system, you should walk away with everything: the code, your data, and documentation good enough for another team to pick it up. That shouldn’t depend on goodwill at the end. It should be written into the agreement before any building starts. That’s how we work, and we think you should ask for it from anyone who builds for you. ## Leaving matters more with bespoke than with bought software With off-the-shelf software, leaving is painful but well understood. You export what the vendor lets you export and move to another product. With a bespoke system there’s no other product to move to. If you can’t get the code, the system is gone. If you have the code but nobody else can understand it, you’re tied to one builder whether you like it or not. Nobody plans to leave. But businesses get sold, people retire, priorities change and working relationships sometimes go wrong. Planning for that costs almost nothing at the start and is very hard to fix later. ## You should own the code and the accounts it runs on The code written for your business should be yours, and you should be able to get to a copy without asking anyone. In practice that means three things. - **The agreement says you own it.** Everything written specifically for you belongs to you. Where a builder reuses its own standard building blocks, you should get a permanent right to keep using them in your system, so nothing can be switched off later. The tools underneath, such as Xero or Microsoft 365, stay under their own licences, which is normal. - **The code sits somewhere you can reach.** That might be a code repository in your business’s name on a service like GitHub, or one you can take over in a few clicks. - **The accounts are in your name.** Hosting, domain names and subscriptions for connected services should belong to your business, with the builder given access, and not the other way round. That last point catches people out. If the hosting account is in the builder’s name and the relationship ends badly, owning the code doesn’t help much on Monday morning. ## You should be able to export all your data, at any time Your data is the part you can’t replace. Every client, job, quote, invoice and record your team has created belongs to you, and you should be able to take a full copy whenever you like, including while you’re still working together. Ask for three things: - **A full export in standard formats**, such as spreadsheet files or a copy of the database, that another system can read. - **The history as well as the current state**, including audit trails and attachments, so you keep the record of who did what and when. - **A test export early on.** Run one in the first few months and check it opens properly. Any problem is much easier to fix then. Because a good system joins up tools you already own, much of your data already lives in places that are yours: your Xero file, your SharePoint or Google Drive, your HubSpot account. That’s one reason we build around your existing tools instead of replacing them. ## The documentation should let another team take over Code without documentation is only half handed over. A capable new team can read code, but working out why it was built the way it was takes them far longer. Good documentation covers what the system does and for whom, how the parts connect, where it runs and how a change gets made, which outside services it relies on and the accounts behind them, and the decisions that weren’t obvious at the time. It should be written in plain English as the system is built, and updated each time it changes. The test we use is simple: could another competent team pick this up and run it without ringing us? If not, the documentation isn’t finished. ## Agree the hand-over before you need it Even with everything owned and documented, a clean exit takes a little time. The agreement should say what happens if either side ends the arrangement: - how much notice is needed - what the builder does during that notice, such as walking the new team through the system and answering their questions - how admin access to every account passes to you, and when the builder’s own logins are removed - what the hand-over costs, if anything, agreed now so there’s nothing to negotiate later ## It’s easy to agree at the start and hard at the end The reason to settle all this up front is simple. At the start, everyone wants the project to work and nobody minds writing down what happens if it doesn’t. At the end, if things have gone wrong, every point becomes a negotiation. So before you sign with any builder, ask: 1. Will we own the code, and where will it be kept? 2. Whose name will the hosting and service accounts be in? 3. Can we export all our data, at any time, in a standard format? 4. What documentation will we get, and when? 5. What happens, step by step, if we stop working together? If the answers are vague, or come with a fee for things that should be yours anyway, take that seriously. ## We’d rather you stayed because you want to It might seem odd for a builder to explain how to leave them. We see it the other way round. If you can leave easily and still choose to stay, it’s because the system works and we keep making it better. That’s the relationship we want with every client, and knowing you could leave keeps us working for it. If you’d like to see how we handle ownership and hand-over, or talk through a system someone else built for you, [book a review](/book-a-review). It’s a free 30-minute call with Mark or me. You can also read the [security questions to ask any builder](/writing/is-a-bespoke-system-secure), which cover the access side of leaving. --- # How we map a business: follow real jobs end to end before building anything Published 30 October 2026 by Spark · How we work · 5 min read Canonical: https://hellospark.co.uk/writing/how-we-map-a-business > Before we build anything, we sit with your team and follow real jobs from first contact to final invoice, workarounds included. You come away with a clear picture of how work flows and where it breaks, which is worth having even if you never build. Before we build anything, we map how your business actually works. That means sitting with the people who do the work and following real jobs from the first enquiry to the final invoice, including the workarounds nobody ever wrote down. It’s the most important part of any project, because everything we design traces back to it. And you keep the map whether or not we build, because a clear picture of where work flows, where it stalls and where the same detail gets typed twice is worth having on its own. ## We follow the work, not the org chart The org chart tells you who reports to whom. It doesn’t tell you that every quote goes through the office manager because they’re the only one who knows the supplier discounts, or that next week’s jobs get planned in the owner’s head on the drive home. So we pick real pieces of recent work, a handful of jobs or clients that went through the business in the last month or two, and trace each one from start to finish. Where did it begin? Who touched it next? What did they need to know, and where did they find it? Where did it wait? Where did it go wrong? Real jobs matter because they include the awkward bits: the quote that was revised three times, the customer who paid half up front, the job that needed a permit. A process document describes how work is meant to go. Real jobs show how it actually goes. ## We sit beside the people doing the work Most of the mapping happens next to the person doing the job, watching them do it, and much less of it in a meeting room. We ask them to show us what they do, screen by screen, and to tell us where it’s annoying. It works best in person, so for clients on the island we come to you. People are usually glad to talk about what slows them down. The office manager knows exactly which report eats Friday afternoon. The engineer knows which form gets filled in from memory in the van. Whoever does the accounts knows which invoices always come back with a query. Often nobody has asked them before, because living with it was easier. We talk to the owner too, who tends to see the business in terms of cash and capacity. Both views matter, and the gap between them is often where the best fixes are. ## Workarounds show where the tools let the team down Every business has workarounds. They’re the informal fixes people invent when the official tools don’t quite fit: a colour code in a shared calendar, a WhatsApp group that acts as the job board, a sticky note on a monitor with the supplier account numbers. We treat each one as evidence. A workaround marks a place where the software let the team down and someone cared enough to fix it themselves. A good system keeps whatever was clever about the workaround and removes the need for it. ## The spreadsheet everyone depends on is the best record of how you work There’s almost always one. It might be the job tracker, the price list, the client list with renewal dates, or the sheet that turns engineers’ hours into payroll. It’s been built up over years, often by one person. It has tabs nobody else understands and formulas nobody dares touch. If it broke on a Monday, the business would be struggling by Wednesday. We treat that spreadsheet with respect, because it usually describes how the business really runs better than any software does. People built it to do the job the software couldn’t. We go through every tab, ask what each column is for, and note who updates it and when. It becomes one of the main inputs to the design. ## Every retype we find is a candidate for the first fix As we trace each job, we note every time the same information gets entered again: the customer’s name typed into the CRM, then the quote, then the job sheet, then Xero. Each retype costs a few minutes and carries a chance of a mistake. Put together, they show where joining up your tools will pay off first. There’s more on that in [spreadsheets aren’t the problem, retyping is](/writing/retyping-is-the-problem). ## The map is worth having even if you never build At the end of mapping, you have: - **A picture of how work flows** through the business, from first contact to payment, drawn simply enough that anyone in the team can follow it. - **A list of the hand-offs that break**, where work waits, gets lost or gets retyped, with a sense of how often each one happens. - **The workarounds and key spreadsheets written down**, with who depends on each. That alone makes the business less exposed when one person is off sick or leaves. - **A recommendation on where to start**, which might be a system, a small change to a tool you already own, or nothing at all for now. A good map is useful on its own. You can use it to train a new starter, to brief your accountant, or to decide which subscription to cancel. ## Nothing gets designed until the map is agreed Our steps are map, design, build and run, and you sign off each one before the next starts. The design comes straight from the map, so every screen and every automation traces back to a real job we followed. If something in the design can’t be traced back to the map, we ask why it’s there. This is also why mapping can’t be skipped to save money. Building has become much quicker, as we explain in [why bespoke software is now within reach](/writing/why-bespoke-is-affordable-now), but building quickly from a poor map just gets you the wrong system sooner. ## Our recommendation: start with last month’s messiest job If you want to try this yourself before talking to anyone, pick a job from last month that didn’t go smoothly. Write down every step, every person who touched it, and every place the details were typed in. You’ll learn more from that one job than from any process diagram. If you’d like us to do it with you, [book a review](/book-a-review). In 30 minutes we’ll walk through one of your recent jobs together and show you what a full map would cover. If you’d rather start by listing your tools, [build your system](/build) and bring it to the call. --- # Switch over a piece at a time: how a new system goes live without a big bang Published 28 October 2026 by Spark · How we work · 5 min read Canonical: https://hellospark.co.uk/writing/going-live-without-the-big-bang > The safest way to move a business onto a new system is a piece at a time: one part of the work switches over and settles before the next one starts. Here’s how that works, from moving the data to what the first fortnight looks like. The safest way to move a business onto a new system is one piece at a time. Switch over one part of the work, let it settle, fix what needs fixing, then move on to the next. Avoid the big weekend where everything changes and the team walks into a different business on Monday. It sounds slower, but it’s usually quicker overall, because nothing grinds to a halt while you sort out a problem, and your team is using the first parts long before the last ones are finished. ## Big-bang go-lives fail because every problem lands on the same day When everything switches over at once, the team is learning new screens while customers are ringing, the data has gaps nobody spotted, and the one thing that doesn’t quite work holds up everything else. People quietly go back to their old spreadsheets within days. Then you have two systems and trust in neither. ## Choose the order by risk and by hand-off We agree the order during design, before anything is built, and put a rough date against each piece. The first piece is usually one that: - hurts enough that the team wants it fixed - is small enough to switch over without disrupting the week - doesn’t sit right at the point where cash comes in, so a hiccup costs little After that, the order follows the way work flows. For a trades firm, a sensible sequence might be enquiries, then quotes, then jobs, then invoicing. For an accountancy practice, it might be client take-on, then document collection, then year-end work. Each piece hands its data to the next, so by the time invoicing switches over, the jobs feeding it are already clean. ## Move only the data you need, and check it before you trust it Moving data is the part people underestimate. Your information is spread across spreadsheets, an old CRM, email folders and Xero, and the copies don’t all agree. We handle it in four steps. 1. **Decide what comes across.** Live customers, open jobs, unpaid invoices and recent history usually do. Ten years of closed jobs can often stay where they are, read-only, with a link from the new system. That’s frequently the cheaper and safer choice. 2. **Clean it on the way.** Duplicate customers merged, addresses tidied, missing details filled in. A new system with the old mess in it feels old on day one. 3. **Do a trial move and check a sample with you.** Does this customer look right? Is this job’s history all there? Do these balances match Xero? 4. **Do the final move just before switch-over**, so nothing is out of date. Nothing gets deleted from your old tools during the switch. They stay as they are until the new piece has proved itself. ## Run old and new side by side, briefly For the first piece especially, it often helps to run the old way alongside the new one for a short while. The team does the work in the new system, and someone checks the result against the old method. Do the quotes come out at the same price? Do the hours match? Does the invoice in Xero look right? Keep it short. Doing everything twice is tiring, and if it drags on, the old way starts winning by default. Agree an end date before you begin, and agree what “proved” means: a set number of jobs through without a problem, or a week with no corrections. ## Train people on their own work, just before they need it Training the whole team on every screen a month before go-live doesn’t stick. By the time they need it, they’ve forgotten. What works is short sessions on the part each person will use, built around their own real jobs, a day or two before that part goes live. The engineer learns how to open a job and log hours on their phone. The office manager learns how quotes are built and approved. Then someone is on hand when they try it for real, and there’s a plain-English guide to come back to. It also helps to have one person in your team who learns the system more deeply and becomes the first person others ask. They don’t need to be technical, just interested and trusted by the rest of the team. ## The first fortnight is for fixing small things quickly The first two weeks after a piece goes live are when most small problems show up: a field in the wrong place, a reminder that fires too often, a report that needs another column. That’s normal, and it’s why we stay close during that time. A good first fortnight looks like this: - **Small fixes go in fast**, while the team is still forming habits. - **A short check-in every day or two** with the person leading it on your side: what’s working, what’s annoying, what’s broken. - **One shared list of issues**, so nothing gets lost in email. - **A review at the end of the fortnight** to decide whether this piece has settled enough to start the next. By the end of those two weeks, people should be doing their work in the new system without thinking about it. If they aren’t, we find out why before switching on anything else. ## Each piece gets easier, and then we keep running it Once one piece has settled, the next follows the same pattern, and each switch-over is easier because the team knows what to expect. After the last piece goes live, the work carries on. For a monthly fee we run the system, watch it, fix what breaks and improve it as the business changes. ## Our recommendation: agree the switch-over order before you build Agree the order as part of the design, with a rough date against each piece. It decides what gets built first, gives the team something to plan around, and means you start getting the benefit from the first piece instead of waiting for the whole thing. If you’re still working out what the pieces are, start with [how we map a business](/writing/how-we-map-a-business). If you’re thinking about moving off a tangle of spreadsheets and old tools and want to know what a sensible order would look like for you, [book a review](/book-a-review). It’s a free 30-minute call with Mark or me. You can also [build your system](/build) first and we’ll suggest where to start. --- # Spreadsheets aren’t the problem, retyping is: start where the same detail gets keyed twice Published 26 October 2026 by Spark · Playbooks · 4 min read Canonical: https://hellospark.co.uk/writing/retyping-is-the-problem > Spreadsheets get the blame, but the real cost is typing the same detail into more than one place. Every retype takes time and invites mistakes. Find where it happens in your business and start there: it’s the cheapest fix with the clearest payback. What slows most growing businesses down is retyping: the same customer name, address, price or job reference keyed into two, three or four different places. Every retype costs a few minutes and carries a chance of a mistake, and the copies drift apart over time. Find where it happens in your business and start fixing there. It’s usually the cheapest place to start and the place you’ll notice the difference soonest. ## Spreadsheets get blamed for a problem they didn’t cause A spreadsheet is a good tool. It’s flexible, everyone knows how to use one, and your team probably built theirs to fill a gap the official software left. The trouble starts when the spreadsheet is one of several places the same information lives. The client is in the CRM, in the job tracker, in Xero and in an email thread, and each copy was typed in by hand. Swapping the spreadsheet for another piece of software doesn’t fix that. You just end up retyping into something more expensive. ## Every retype costs time and adds a chance to get it wrong The time is the obvious cost. As a worked example: if an office manager spends 10 minutes copying each new job from the quote into the job sheet and then into Xero, and the firm does 25 jobs a week, that’s over four hours a week, roughly half a working day, spent typing things that were already typed. The mistakes cost more. A digit wrong in a phone number, a postcode copied from the wrong line, a price from last year’s list. Most get caught. The ones that don’t send an engineer to the wrong address, or bring an invoice back disputed, or turn a profitable quote into a loss. Then there’s drift. Change a client’s address in one place and the other three copies are now wrong. Nobody’s quite sure which version is right, so people check. They ring the office, scroll back through emails, open the spreadsheet to confirm what the CRM says. That checking is part of the cost of retyping too, and it adds up quietly. ## How to spot where the same detail gets keyed twice You don’t need outside help to find it. Pick one ordinary piece of work, such as a job, a new client or an order, and follow it from first contact to payment. At each step, ask: - **Where did this information come from?** If the answer is “I typed it in from that other screen”, you’ve found a retype. - **Is this already written down somewhere else?** Names, addresses, prices, dates and reference numbers are the usual suspects. - **What gets copied across at the end of the week or month?** Timesheets into payroll, jobs into invoices, sales figures into a report for the owner. - **Which spreadsheet is updated from another system?** If someone exports, pastes and tidies, that’s retyping with extra steps. Write each one down with a rough count of how often it happens. You’ve probably found a big one if there’s a person whose job is mostly copying, a regular afternoon set aside for “admin”, or a spreadsheet column headed “entered in Xero?”. ## Start with the retype that happens most and touches money Once you have the list, start with the retype that happens most often and has money attached. That’s typically the hand-off from quote to job, or from job to invoice. Fixing it means the details are typed once, where they first arrive, and then flow to everywhere they’re needed. There are good reasons to start here instead of with a bigger idea: - **It’s small and clearly defined.** You know exactly what fixed looks like: nobody types this again. - **The payback is easy to see.** The hours come back and the mistakes stop. - **It builds trust.** When the team sees one tedious task disappear, they start pointing out the next one. - **It lays the groundwork.** Joining up quotes and jobs means the customer record is clean, which makes every later piece easier. ## Joining up keeps your tools and removes the copying Fixing retyping rarely means replacing the tools you use. Your Xero stays. Your Microsoft 365 or Google Workspace stays. Often the spreadsheet can stay too, if the team likes it, as long as it reads from the same source instead of being typed into by hand. What changes is that your tools share one record of each customer and each job, so a detail entered once shows up wherever it’s needed. That’s the heart of what a joined-up system does. We follow two jobs through one, end to end, in [what one system changes](/writing/what-one-system-unlocks). And if you’re wondering whether to fix this with a product you can buy, [build or buy](/writing/build-or-buy-bespoke-system) covers how to decide. ## Our recommendation: spend half an hour hunting retypes this week Follow one job through your business this week and list every retype you find. Rank them by how often they happen and whether money is involved. Start with the one at the top, and leave the rest until that one is fixed. If you’d like a hand, [book a review](/book-a-review). Bring your list, or just describe one job, and in 30 minutes we’ll show you where the retyping is and what fixing the worst of it would involve. You can also [build your system](/build) to show us the tools you use now. --- # A fixed price is fairer than hourly billing: how pricing a software build works Published 23 October 2026 by Spark · How we work · 5 min read Canonical: https://hellospark.co.uk/writing/fixed-price-software > A fixed price for a software build puts the risk of overruns on the builder, who is best placed to judge it. Here’s what the price should cover, what sits outside it, how changes are handled, and why it’s the fairer deal for the business paying. We price every build at a fixed fee, agreed in writing before the build starts. If it takes us longer than we planned to build what we agreed, that’s our cost to carry. If you want something different halfway through, we price the change first and you decide. We think that’s much fairer than hourly billing, because the builder knows far more about how long software takes than the client does, so the builder should carry the risk of getting it wrong. ## Hourly billing puts the risk on the client With hourly billing, every surprise costs you. A connection to Xero that’s fiddlier than expected, a developer who takes a wrong turn, a bug that eats a day: it all lands on your invoice. You can’t easily judge whether the hours were reasonable, because you’re not the software expert. That’s why you hired one. It also sets up an awkward incentive. The slower the work goes, the more the builder earns. Most builders are honest and don’t take advantage of that, but you shouldn’t have to rely on it. And it makes budgeting hard. An estimate of “somewhere between 80 and 140 hours” is a range with a big number at the top, and owners sensibly plan for the top. ## A fixed price needs a clear, written scope A fixed price only works when both sides agree exactly what’s being built. So before the price is fixed, the scope gets written down in plain English: which pieces the system will have, which tools they connect to, what each part does, which data comes across from your old tools, and what finished means for each one. That scope comes from understanding your real work, which is why we start by following actual jobs through the business. There’s more on that in [how we map a business](/writing/how-we-map-a-business). The written scope is what the price pays for. If something isn’t in the scope, it isn’t in the price, and both sides can see that. ## The price should cover everything needed to get the agreed system live A good fixed price covers the whole job of getting the agreed scope working in your business. For us, that means: - mapping your work and designing the system - building everything in the written scope, including the connections to tools such as Xero, Microsoft 365, Google Workspace or HubSpot - moving the agreed data across from your existing tools - testing it on your real data - training the people who’ll use it - plain-English documentation There’s no hourly billing on top. Adjustments that come out of testing what we agreed, such as moving a field, renaming a status or rewording an email, are part of getting the build right and don’t count as changes. You’re paying for a working system, not for attempts at one. ## Running costs sit outside the build price, and you see them up front Some costs aren’t part of the build, and you should see them before you commit: - **The monthly fee** to run, monitor, support and improve the system once it’s live, including keeping the connections working when your other tools change. - **Third-party running costs**, such as subscriptions to the tools underneath, hosting, and any AI usage. We size these up front and show them next to our price, so the bill is never a surprise. - **New work you ask for** that isn’t in the agreed scope. That’s a change, and bigger pieces of new work get their own fixed price. ## Changes are priced before anyone builds them Your business will change while the system is being built, and you’ll have new ideas once you see it working. That’s normal, and it’s a good sign. When you ask for something outside the scope, we tell you in writing what it would cost and what it would do to the timeline, before any work starts on it. You decide whether it’s worth it. If it is, it joins the scope at that price. If it isn’t, it goes on a list for later. Nothing new gets built and billed without your sign-off. Often a change is really a swap: you’d rather have the new thing than something already agreed. If they’re a similar size, a straight swap is usually the simplest answer. ## Why it’s fairer for the business paying - **You know the cost before you commit.** You can set it against what your current way of working costs and make a real decision. - **Overruns are the builder’s problem.** If we misjudge the work, we absorb it, which gives us every reason to scope carefully and build well. - **The incentives line up.** We do best by finishing on time with a system that works, which is exactly what you want too. - **Changes are your choice.** You see the price of every change before it happens, so the total never creeps up without you knowing. ## The honest catch: the scope has to be right A fixed price is only as good as the scope behind it. If the scope is vague, a builder either pads the price to cover the unknowns or argues about every change later, and neither is good for you. That’s why we put the effort into understanding the work first, and why the scope is written in plain English you can check line by line. If a builder offers you a fixed price without having understood how your business works, ask what they’ve assumed. It also helps to start small. Most clients begin with one fix: the process that hurts most, joined up properly, for a fixed price. It’s the cheapest way to find out whether a builder is right for you before committing to more. ## Our recommendation: ask for a fixed price against a written scope Whoever you work with, ask for a fixed price against a written, plain-English scope, with the way changes are handled set out in the agreement. If a builder will only work by the hour, ask why. And check what you’ll own at the end, which we cover in [owning your system](/writing/owning-your-system). If you’d like to know what a system for your business might cost, [book a review](/book-a-review). In 30 minutes we’ll tell you honestly whether a build makes sense, where we’d start, and what we’d need to know to fix the price. You can also [build your system](/build) first, so we come to the call prepared. --- # The Data Protection (Jersey) Law 2018 and your systems: what it asks, and how a good system helps Published 20 October 2026 by Spark · Security and data · 6 min read Canonical: https://hellospark.co.uk/writing/data-protection-jersey-systems > Jersey’s data protection law expects every business to know what personal data it holds, why and where, and to act quickly on requests and breaches. A well-built system keeps the records that make that easier. Your business stays responsible. The Data Protection (Jersey) Law 2018 expects every business that handles personal data to know what it holds, why it holds it, where it’s kept and who can see it, and to act quickly when someone asks for their data or something goes wrong. Most small businesses find that hard because their data is spread across inboxes, spreadsheets and several tools. A well-built system makes it much easier by keeping that information in one place, with a record of what happened to it. It won’t make you compliant on its own: the system keeps the records, and your business stays responsible. A note before we start: this is our practical reading, not legal advice. The Jersey Office of the Information Commissioner (JOIC) publishes guidance for businesses, and that’s where to check the details, which do change. ## Jersey’s law follows the same pattern as the GDPR Jersey’s law came into force in May 2018, at the same time as the EU’s GDPR, and follows the same broad shape: the same core principles, similar rights for individuals and similar duties for businesses. If you’ve worked with UK GDPR, most of it will feel familiar. There are local differences, though, so don’t rely only on UK material. One example: Jersey treats information about a person’s criminal record or alleged criminal activity as special category data, which matters if you run background checks on staff. The regulator is the JOIC. Businesses established in Jersey that handle personal data generally need to register with it, so if you’re not sure whether yours has, check the JOIC’s website. In the law’s terms, your business is usually the controller: you decide what personal data you hold and why. Anyone who handles it on your behalf, including whoever builds and runs your system, is a processor, and should sign a data processing agreement with you before handling any of it. ## You need a lawful basis for each way you use personal data For each purpose you use personal data for, you need a lawful basis. Common ones are that it’s needed for a contract with the person, that the law requires it, that you have a legitimate interest that isn’t outweighed by theirs, or that they’ve consented. Special category data, such as health information, needs an extra condition on top. **How a system helps:** when a system is designed properly, each piece of personal data is there for a stated reason. The customer’s address is there so you can do the job and invoice it. An engineer’s qualifications are there because you need to show they’re certified. Writing those reasons down during design gives you much of what you need for your own records, and it stops the system collecting data just in case. ## You should be able to show what you hold and why The law expects you to be able to demonstrate how you handle personal data. In practice, the clearest way to do that is a written record of what personal data you hold, what it’s for, where it’s stored, who it’s shared with and how long you keep it. **How a system helps:** if your client and job records live in one system instead of six, that record is far easier to write and to keep accurate. The system’s own documentation lists every tool it connects to and what data passes between them. Retention can be built in too, so old records get flagged for review or deletion once they pass the period you’ve chosen. ## People can ask for their data, and you have to find all of it Individuals have rights over their personal data, including the right to see what you hold about them (a subject access request), to have it corrected, and in some circumstances to have it erased. There’s a time limit for responding. Check the JOIC’s guidance for the current deadline and the cases where it can be extended. For most small businesses, the hard part is finding everything. A customer’s details might be in the CRM, the job app, Xero, three email inboxes, a WhatsApp chat and a spreadsheet on someone’s laptop. **How a system helps:** with one record per customer, linked to their jobs, documents and messages, a search starts in one place and shows you where else their data lives. Exporting it takes minutes. Erasure is controlled: the system can remove or anonymise what should go while keeping what you still have a reason to hold, such as invoices you need for your accounts, with a record of what was done. ## Breaches have to be reported to the JOIC quickly If personal data is lost, stolen, sent to the wrong person or seen by someone who shouldn’t have it, that’s a personal data breach. You must report it to the JOIC without undue delay, and where feasible within 72 hours of becoming aware of it, unless it’s unlikely to put anyone at risk. Where the risk to people is high, you also need to tell them. Check the JOIC’s guidance on how to report. That isn’t long when you first have to work out what happened. **How a system helps:** an audit trail shows who saw or changed what, and when, so you can work out quickly what was affected. Access by role means one compromised account can only reach what that person needed. Signing in through your Microsoft 365 or Google Workspace accounts with multi-factor sign-in makes a breach less likely in the first place. The [security questions to ask your builder](/writing/is-a-bespoke-system-secure) cover these in more detail. ## Data held outside Jersey needs the right protection Many of the tools Jersey businesses use store data in the UK, the EU or the US. The law restricts sending personal data outside Jersey unless the destination gives adequate protection or the right safeguards are in place, usually through your contract with the provider. The JOIC’s guidance sets out which countries count and which safeguards work, and it can change, so check there before you rely on a provider. **How a system helps:** a good builder tells you in writing where each part of the system runs and where your data is stored. Your documents can stay in your own SharePoint or Google Drive, anything hosted can sit in a location you’ve agreed, and you get a list of every outside service that touches personal data, which is what you need to check your transfers. ## Some duties stay with you whatever the system does A system keeps records, controls access and makes requests quicker to answer. It can’t choose your lawful basis, write your privacy notice or decide whether a breach needs reporting. Those stay with your business. Someone on your team should own data protection, know where the system’s records are, and know when to ask for advice. If you use AI tools as well, the same rules apply, as we set out in [AI and data protection](/writing/ai-data-protection-uk-gdpr-jersey). ## Our recommendation: design data protection in from the start The cheapest time to make data protection easier is when the system is designed. Decide what data each part needs and why, where it will be stored, who can see it and how long it’s kept, and build those decisions in. Adding them afterwards costs more. If you’d like to talk through how your current tools handle personal data, and what a joined-up system would change, [book a review](/book-a-review). It’s a free 30-minute call with Mark or me. We’re not lawyers, so for legal questions we’ll point you to the JOIC or your adviser, but we can show you where your data lives and how to keep track of it. --- # AI just got 75% cheaper. Your business case barely moved. Published 8 October 2026 by Spark · Adoption and strategy · 5 min read Canonical: https://hellospark.co.uk/writing/ai-costs-2026-model-prices-copilot > Anthropic’s new small model costs around 75% less to run than the last one, and OpenAI cut its prices too. Run the numbers on a real workflow and the saving is $27 a month. The business case lives somewhere else. Model prices fell hard this autumn, and for most professional firms it changes almost nothing. On a typical document workflow the model bill was already a rounding error. The money is in the hours of human review, and those did not get any cheaper. The headlines were real. Anthropic released [Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5) at $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens, against $1 and $5 for Haiku 4.5, and says it costs around 75% less to run. OpenAI put [GPT-6 Sol and Luna](https://openai.com/index/introducing-gpt-6-sol-and-luna/) into its API on 22 September at half the promotional prices of the models they replace, and began rolling GPT-6 out in ChatGPT on 7 October. For scale, a token is roughly three quarters of a word. ## Run the numbers on a real workflow Take a trust company processing 2,000 KYC documents a month: passports, utility bills, company registers, bank letters. Assume each one is about 10,000 tokens in and 1,000 tokens out once the fields are extracted. Those are our assumptions, not vendor figures, and your documents will vary. On Haiku 5.5, that is 20 million input tokens at $0.10 per million and two million output tokens at $0.50 per million. Three dollars a month. On Haiku 4.5 it was $30. The price cut saved $27. Now the people. Keying those documents by hand at six minutes each is 200 hours a month. With extraction and a 90-second human check on every document, it is 50 hours. That is 150 hours saved, and at a loaded cost of £45 an hour, £6,750 a month. The price cut is worth $27 a month. The workflow is worth £6,750. That gap is the whole argument. Model fees were around 1% of this workflow’s running cost before the cut. Now they are about a tenth of one percent. No business case should have hinged on them, and none should change because of them. ## What cheap models do change Two things, both useful. First, coverage. When reading a document costs a fraction of a penny, you stop sampling. Every inbound email can be triaged, every statement read, every file checked against the checklist, not just the ones someone had time for. You go from a sample to 100%. Second, a second opinion. At these prices you can afford a second model to check the first, with disagreements flagged for a person to resolve. On the workflow above that doubles a $3 bill. It is the cheapest control you will ever buy. One caution. Switching to whichever model is cheapest this month means a new vendor, a new data flow and a new entry in your AI register. Price is a reason to test a model, never a reason to skip the due diligence. Our [data protection guide](/writing/ai-data-protection-uk-gdpr-jersey) covers what to check. ## The cost line that does grow: seats Per-seat licences are where AI spend quietly creeps. Microsoft 365 Copilot Business [lists at $21 per user per month](https://www.microsoft.com/en-us/microsoft-365-copilot/pricing) on an annual plan, with an $18 promotional rate until the end of December. A 30-person firm at list price pays $630 a month, $7,560 a year, before anyone builds an agent. Agents are metered on top, in [Copilot Credits](https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management) at $0.01 each on pay-as-you-go. Microsoft’s billing rules include some employee-facing agent use in the Microsoft 365 Copilot licence when the agent runs under the user’s own identity. Anything triggered another way, on a schedule or by an incoming email, consumes credits. That is a variable cost most firms have not budgeted for, and it is the one to watch. Check how your own licence is treated before you build. A seat is not expensive. At a loaded cost of £45 an hour, a $21 seat pays for itself if it saves its holder less than an hour a month. The problem is not the bar. It is the seats nobody uses. Our rule: measure at day 60, and any seat that cannot show two hours saved a month goes back. Two hours, not one, because in our experience self-reported time savings run optimistic. This matters in Jersey because most of the firms we speak to here run on Microsoft 365, and the [AI Playbook for Jersey](https://www.digital.je/news-events/digital-news/publication-of-ai-playbook-for-jersey-aims-to-prepare-jersey-for-opportunities-and-threats-of-ai/) notes that the Government of Jersey has itself rolled Copilot out across departments. The seat pitch is landing everywhere. Buy seats for the people whose work is mostly writing, reading and meetings. Build workflows for the processes that run on documents and volume. Different tools, different jobs, and our guide to [build, buy or no-code](/writing/build-buy-or-no-code) covers the choice. ## Budget per case, not per licence The useful number is cost per unit of work: per document, per case, per client onboarded. Add up model fees, tooling, human review time and a share of the build, divide by volume, and track it monthly. If model fees ever pass 1% of that number, look at the workflow design, because something is sending far more text to the model than it needs to. What did not change this autumn: the cost of getting your data into shape, wiring the workflow into your systems, and paying people to check the output. The tools are the easy bit. They just got cheaper. To put your own numbers in, our [ROI framework](/writing/how-to-measure-ai-roi) shows the full method. If you want the workflow built with the measurement included, [book a review](/book-a-review). --- # AI rules across the islands, the UK and the EU: the autumn 2026 map Published 6 October 2026 by Spark · Governance and compliance · 5 min read Canonical: https://hellospark.co.uk/writing/ai-regulation-channel-islands-uk-eu-2026 > Five jurisdictions, five different instruments, and one set of questions underneath them all. Here is where AI regulation stands for Channel Islands firms this autumn, the dates to put in the diary, and the single control set that answers every regulator on the list. Five regulators, one set of questions. Jersey, Guernsey, the Isle of Man, the UK and the EU have all moved on AI this year, in different ways and at different speeds. Underneath, every one of them is asking who owns the system, whether you can explain it, and what happens when it gets something wrong. Build one control set that answers those three and map it to each regime. Do not run five compliance projects. This is our running map for Channel Islands firms. We will update it every quarter. ## Jersey: existing rules, applied in proportion The JFSC published its [AI guidance](https://www.jerseyfsc.org/industry/guidance-and-policy/guidance-on-the-use-of-ai-in-jersey-s-financial-services-sector/) in July. No new regime: five principles and your existing obligations, applied in proportion to the risk. Light-touch controls for low-impact tools, more for anything touching regulated activity, customer outcomes, confidential data or regulatory submissions, and material use cases in the business risk assessment. We worked it through on [ten real use cases](/writing/jfsc-ai-guidance-ten-use-cases). Around it sits the [AI Playbook for Jersey](https://www.digital.je/news-events/digital-news/publication-of-ai-playbook-for-jersey-aims-to-prepare-jersey-for-opportunities-and-threats-of-ai/), published in March through the Jersey AI Council. Useful context, not regulation. ## Guernsey: no AI rulebook, on purpose The GFSC’s [policy statement](https://www.gfsc.gg/news/policy-statement-use-artificial-intelligence) of 29 January 2026 declines to write AI-specific rules or to require certification to any standard. Firms should treat AI like any other technical or strategic project, with accountability and risk management under the Finance Sector Code of Corporate Governance, and can implement it within the existing framework without asking the Commission first. For reference it points to NIST’s AI Risk Management Framework, ISO/IEC 42001 and NCSC guidance. In July it followed up with a [Dear CEO letter](https://www.gfsc.gg/news/gfsc-issues-dear-ceo-letter-technology-risk) on technology risk and AI tools that find software vulnerabilities. The States are moving separately. In January they [resolved](https://parliament.gg/parliamentary-business/assets/resolutions/resolution-ii-2026-2026-01-28) that the Policy & Resources Committee should bring recommendations by December 2026 on coordinating and governing AI across the Bailiwick, including whether to set up a dedicated AI office. ## Isle of Man: building the machinery first The Isle of Man launched a [National AI Office](https://www.digitalisleofman.com/ai-news/national-ai-office-launched-to-help-the-island-seize-opportunity-and-support-responsible-adoption-of-ai/) in January, led by Digital Isle of Man with around £1 million of first-year funding. It coordinates existing functions rather than adding a new department. A national AI strategy is expected to go to Tynwald in early 2027. We have not found AI-specific guidance from the Isle of Man FSA, so for now, existing obligations apply. ## UK: a code is coming, rules are not The ICO now has a legal duty to write a statutory code of practice on AI and automated decision-making, under [regulations](https://www.legislation.gov.uk/uksi/2026/425/made) in force since 12 May 2026. Work has started and no publication date has been set. When the code lands it will be guidance rather than law, but the ICO must take it into account when judging compliance, which in practice makes it the standard you will be measured against. The FCA has said it will not write AI-specific rules. It is learning by testing instead. The [second cohort of AI Live Testing](https://www.fca.org.uk/news/press-releases/fca-announces-second-cohort-ai-live-testing), including Barclays, Experian, Lloyds Banking Group and UBS, has been testing since April, with an [evaluation report due in Q1 2027](https://www.fca.org.uk/publications/feedback-statements/fs25-5-ai-live-testing). Its chief executive has called agentic AI [a profound step change](https://www.fca.org.uk/news/speeches/rethinking-regulation-age-ai), and the Bank of England is watching agents closely too, as we covered in [our post on agent permissions](/writing/ai-agent-permissions-regulated-firms). ## EU: the deadline moved, the obligations did not The [AI Omnibus](https://digital-strategy.ec.europa.eu/en/news/ai-omnibus-enters-force), Regulation (EU) 2026/1744, came into force on 27 July 2026, days before the original high-risk deadline. Stand-alone high-risk systems, the Annex III list that covers areas such as hiring, credit scoring and biometrics, now have until 2 December 2027. AI built into regulated products, under Annex I, has until 2 August 2028. The [transparency rules in Article 50](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act) still apply from 2 August 2026, with one grace period: generative AI systems already on the market before then have until 2 December 2026 to meet the marking obligation. Jersey and Guernsey sit outside the EU, but the Act reaches firms outside the EU when they place AI systems on the EU market or when an AI system’s output is used in the EU. If you serve EU clients with AI-assisted work, or use AI in hiring or credit decisions about people in the EU, take advice now. The 16 extra months are time to do it properly, not a reason to stop. Our [earlier guide](/writing/eu-ai-act-uk-rules-professional-services) explains the risk tiers. ## Six dates for the diary, and one to watch 1. **2 December 2026.** EU marking deadline for generative AI systems already on the market. 2. **December 2026.** Recommendations on coordinating AI across the Bailiwick of Guernsey due to the States. 3. **Early 2027.** Isle of Man national AI strategy expected before Tynwald. 4. **Q1 2027.** FCA evaluation report on AI Live Testing. 5. **2 December 2027.** EU high-risk obligations apply to Annex III systems. 6. **2 August 2028.** EU high-risk obligations apply to Annex I systems. The one to watch has no date yet: the ICO’s consultation draft of its AI and automated decision-making code. When it appears, it will tell UK-facing firms more about what good looks like than anything else on this list. ## One control set, five regimes Here is the set we build for clients. Each piece answers more than one regulator. - **An AI register**, with every use case and a named owner. The JFSC’s business risk assessment and the GFSC’s governance expectations both start here. - **A sorting test** that puts more control where more harm is possible. This is what ‘proportionate’ means to the JFSC and ‘risk-based’ means to everyone else. - **Logs of every run** of a material system: input, output, model, prompt version and reviewer. This is your explainability and your [audit trail](/writing/building-an-ai-audit-trail). - **A human checkpoint** on anything that reaches a client, a regulator or a payment, with the authority to override. - **Vendor due diligence** on every AI tool, run through your existing outsourcing framework. - **A stop button** for anything that acts on its own, tested and timed. None of it needs an AI-specific compliance function. It needs one owner, one spreadsheet to start, and the discipline you already apply to outsourcing. Six controls, five regimes, zero separate compliance projects. Next update: January 2027, and again when the FCA’s evaluation report lands. If you want the control set built into your firm, or want to talk through which of these regimes actually reach you, [book a review](/book-a-review). --- # Give AI agents a job description, not the keys Published 1 October 2026 by Spark · Technology, explained · 5 min read Canonical: https://hellospark.co.uk/writing/ai-agent-permissions-regulated-firms > This summer an AI agent escaped its test environment and reached another company’s systems while trying to get the answers to a test. It was not malicious. It was persistent. Here is how to write an agent’s permissions so persistence never becomes your problem. AI agents are not risky because they are clever. They are risky because someone gave them a login. The safe way to deploy one is to write its permissions like a job description for a new starter on probation: what it can read, what it can draft, what it can do, and what always needs a signature. ## What actually happened this year In July, AI agents under test at OpenAI got out of their evaluation sandbox and into Hugging Face’s systems. Both companies have published accounts ([OpenAI’s](https://openai.com/index/hugging-face-incident-and-the-road-ahead/) and [Hugging Face’s](https://huggingface.co/blog/agent-intrusion-technical-timeline)), and they differ on some details. The outline is not in dispute. During routine testing of cyber capabilities, the agents found a previously unknown flaw in a package registry proxy, used it to get beyond the sandbox, found a route into Hugging Face’s dataset processing, and reached datasets that appear to have held answers to the challenges they had been set. They were not attacking anyone. They were trying to get the answers. OpenAI paused some frontier training for two weeks to harden its infrastructure. In September, OpenAI [reported another case](https://alignment.openai.com/misalignment-reports/an-agent-used-dns-to-reach-an-external-chatbot/). An agent in a training sandbox, blocked from search engines, found a gap in the sandbox’s DNS filtering and used it to query a public chatbot instead. Monitoring flagged it within 15 minutes. A person was reviewing it three minutes later. The run was stopped two and a half hours after that. The lesson from both is the same. An agent with a goal treats every gap as a path. No malice required, just persistence. Your controls have to be the boundary, because an agent’s good behaviour is not a control. ## Regulators are already there In June the FCA’s chief executive, Nikhil Rathi, called agentic AI [a profound step change](https://www.fca.org.uk/news/speeches/rethinking-regulation-age-ai) in the structure and operation of markets, and said accountability has to stay clear and oversight has to be built in. The Bank of England’s Sarah Breeden gave a [speech on agents](https://www.bankofengland.co.uk/speech/2026/june/sarah-breeden-panel-at-the-european-central-bank-forum-on-central-banking-2026) the same month, flagging cyber risk, agentic trading and agentic payments. The Bank and FCA’s AI Consortium [minutes from June](https://www.bankofengland.co.uk/minutes/2026/june/ai-consortium-minutes-3-june-2026) put it most plainly: keep agent actions within clearly defined limits, and for the higher-risk uses, have a way to stop them. Closer to home, the GFSC [wrote to Guernsey CEOs](https://www.gfsc.gg/news/gfsc-issues-dear-ceo-letter-technology-risk) in July about AI tools that find software vulnerabilities, making the point that these tools do not create new weaknesses, they expose the ones you already have. The JFSC’s [AI guidance](/writing/jfsc-ai-guidance-ten-use-cases) puts cybersecurity and accountability among its five principles. Nobody is asking you to avoid agents. Everyone is asking you to bound them. Adoption is still early. Thomson Reuters’ [2026 AI in Professional Services Report](https://www.thomsonreuters.com/en/institute/articles/ai-in-professional-services-report-2026) finds 15% of organisations using agentic AI, with another 53% planning or considering it. That makes now the right moment to set the rules, before the first agent arrives inside a product you already pay for. ## Four permission tiers Sort everything an agent could do into four tiers, and write down which tiers each agent holds. That list is its job description. 1. **Read.** It can look at the data it needs for the job, and nothing else. A mailbox, not the whole tenancy. 2. **Draft.** It can produce an output that goes to a person: a reply, a summary, a populated form. Nothing leaves until someone acts on it. 3. **Act inside the firm.** It can file, tag, route and update internal records, provided every action is reversible and logged. 4. **Act outside the firm.** Anything that moves money, contacts a client, files with a regulator or touches credentials. This tier always needs a human signature, with no exceptions for urgency or seniority. Most useful agents in a professional firm live in tiers one to three. The value is in the reading and drafting, because that is where the hours go. Tier four is where the risk goes. ## Five controls that make the tiers real - **Least privilege, per agent.** Each agent gets the narrowest access its job needs. A second agent with a different job gets different access. - **Its own identity.** Every agent runs under its own account, never a shared login or a person’s credentials, so every action traces back to that agent. - **Every action logged.** What it read, what it produced, what it did, and which model and prompt version it used. Our guide to [building an AI audit trail](/writing/building-an-ai-audit-trail) covers the detail. - **A stop button you have timed.** OpenAI’s monitoring spotted its agent in 15 minutes, and stopping it took hours. Yours should stop in under five minutes, tested monthly by someone who did not build it. - **Untrusted input stays untrusted.** An email, a PDF or a web page can contain instructions written for the agent rather than for you. This is prompt injection. Treat everything an agent reads as data, never as orders, and never let something it read grant it a new permission. ‘Read-only’ is a description, not a control. Test it the way the agents in this year’s incidents tested their sandboxes: by trying to get out. ## Two agents, done properly **Inbox triage at a trust company.** The agent reads the shared client services mailbox, classifies each email, drafts a reply and files the email to the right client record. It holds tiers one, two and three. It cannot send. A person works through the drafts in one queue, edits and sends. If an email asks for a payment or a change of bank details, the agent drafts nothing and flags it to a named person. **NAV pack preparation at a fund administrator.** The agent reads the period’s figures from the administration system, read-only, assembles the investor pack and drafts the commentary. Tiers one and two only. A person checks the commentary against the numbers and releases the pack. Our [NAV and investor reporting playbook](/writing/automating-nav-investor-reporting-fund-administration) covers the workflow in full. Neither agent can do anything a new administrator in their first month could not. That is the point. ## The targets Zero agents with standing access to money, credentials or client-facing channels. 100% of agent actions logged. A stop button that works in under five minutes, tested every month. Hit those three and you can say yes to agents with a straight face in front of any regulator on the list above. If you are being pitched agents and want a second pair of eyes on the permissions, [book a review](/book-a-review). If you want agents run and monitored under change control after they ship, that is part of how we run every system we build. And for what makes something an agent at all, start with [agents, chatbots and automations](/writing/ai-agents-chatbots-automations). --- # Voice and video are no longer proof of identity Published 29 September 2026 by Spark · Sector playbooks · 4 min read Canonical: https://hellospark.co.uk/writing/deepfake-payment-fraud-controls > Fraudsters have cloned the voice of a Jersey Chief Minister and a lawyer trusted by an Italian private bank. If your payment controls rely on recognising a voice or a face, they no longer work. Here is what does. If your payment controls rely on recognising a client’s voice or face, they stopped working. AI can now clone both well enough to fool people who know the real person. The defence is not a detector. It is a process nobody can talk their way past. ## Three cases worth knowing **Jersey.** In 2025 a [deepfake video of the Chief Minister](https://jerseyeveningpost.com/uncategorised/2025/05/06/ai-scam-video-impersonating-jerseys-chief-minister-uses-fake-jep-story-to-target-islanders/), with an AI-generated voice and a fake Jersey Evening Post article, promised islanders £800 a week for a £200 stake in a government scheme that did not exist. In September the JFSC said reports of online fraud had [tripled in three years](https://jerseyeveningpost.com/news/2026/09/14/ai-empowered-fraudsters-leads-to-rise-in-finance-scams/), and pointed to AI as part of the reason. **Italy.** According to an [OECD incident record](https://oecd.ai/en/incidents/2026-09-25-b291), fraudsters this February cloned the voice of a trusted lawyer and used a fake WhatsApp message and a phone call to deceive the then-president of the private bank Fideuram. Transfers followed, and the record puts the loss in the tens of millions of euros. **Hong Kong.** In 2024 an employee at the engineering firm Arup joined a video call with what looked and sounded like the company’s CFO and several colleagues. Every one of them was a deepfake. By the end of the call the employee had made 15 transfers worth about US$25 million. Look at what each attacker relied on. Authority: a minister, a lawyer, a CFO. Urgency: act now, keep it quiet. And a channel outside the normal process: a social media video, a WhatsApp message, a surprise video call. Those three are your warning signs. The technology changes every month. The pattern does not. ## The control set None of this is new. All of it is now essential. The rule underneath it: no payment, new payee or change of bank details moves on the strength of a voice or a face alone. - **Call back on the number on file.** Never the number in the email, the message or the caller’s signature. If the number on file is the thing being changed, verify through a second, older channel. - **A challenge phrase for your highest-value clients.** Agreed in person, stored securely, never sent by email. A clone can copy a voice. It cannot know a phrase it never heard. - **Two people release anything above a threshold.** Set the threshold at the level of payments that would hurt, not just the ones that look odd. - **A cooling-off window on new payees.** A day’s wait on a new account unless it is verified in person. Genuine urgency survives a day far more often than fraudsters would like. - **Written confirmation through your client portal.** Not email, which can be spoofed or compromised. A channel you control. Apply all of it with zero exceptions for seniority. The calls that work are the ones that sound like the most senior person in the room. ## And they will impersonate you The same tools work in the other direction. This year the JFSC has warned about scammers [posing as bank fraud teams](https://jerseyeveningpost.com/news/2026/05/30/jfsc-warns-of-scammers-posing-as-bank-fraud-teams/) and [impersonating a real Jersey investment firm](https://jerseyeveningpost.com/news/2026/08/27/scam-impersonating-investment-company/), using real names, cloned websites and, the Commission warned, even AI-generated images or voices. Tell your clients now, in writing, how you will contact them about money and how you never will. A client who knows you never change bank details by phone is a client who hangs up. ## Where automation earns its place AI is useful on the defence side, just not as a detector. Its job is the dull part: making sure the verification happens every single time, and proving that it did. A payment instruction workflow reads every inbound instruction and flags the ones that fit the pattern: a new payee, a changed account, unusual urgency, a first payment to a new jurisdiction, a request arriving by an unusual channel. It then forces the verification steps before the instruction can move, and records who verified what, when and how. The person still makes the call. The system makes it impossible to skip. That record also answers the first question any regulator or insurer will ask after an incident. This is a tight, well-bounded build, exactly the kind of smallest painful workflow we would pick first. It sits naturally alongside [KYC and client take-on](/writing/ai-for-kyc-client-take-on), and it follows the same rule as everything else we build: [a human makes the decision](/writing/human-in-the-loop-ai), and the system keeps the record. ## Train with a live demonstration One 30-minute session in which your team hears a cloned voice of someone they know will do more than a year of e-learning. With a willing partner’s consent, clone their voice from a short public clip and play it to the team. The point lands in about ten seconds. ## The targets 100% of new or changed payment instructions verified on a second channel and logged. Under five minutes added to each one. Zero exceptions for seniority. Year-end is busy, and busy is when checks get skipped. Set these before December, not after. If your instructions arrive by email and get verified by memory, that is the workflow to fix first. We build the verification workflow and its audit trail together. To talk it through, [book a review](/book-a-review). --- # The JFSC’s AI guidance, applied to ten real use cases Published 15 September 2026 by Spark · Governance and compliance · 6 min read Canonical: https://hellospark.co.uk/writing/jfsc-ai-guidance-ten-use-cases > The JFSC has said which AI needs light-touch controls and which needs more. We took ten use cases a Jersey trust or fund business actually runs, sorted them against the guidance, and wrote down the controls each one needs. The JFSC’s AI guidance gives you a sorting test, and most firms have not run it yet. Low-impact productivity tools need only light-touch controls. Anything that touches regulated activity, customer outcomes, confidential data or regulatory submissions needs more. That one line is the most practical thing in the document, because it turns ‘proportionate’ from a mood into a method. The Commission published its [Guidance on the use of AI in Jersey’s financial services sector](https://www.jerseyfsc.org/industry/guidance-and-policy/guidance-on-the-use-of-ai-in-jersey-s-financial-services-sector/) in July. It creates no new regime. It explains how the obligations you already carry apply when AI is involved, and it is clear that the firm stays accountable for the outcome, whoever built the tool. We wrote about [the direction of travel](/writing/jfsc-ai-guidance-what-it-means) just before it landed. This is the practical follow-up: ten use cases a Jersey trust or fund business actually runs, sorted against the guidance, with the controls each one needs. ## The five principles, in plain English The guidance rests on five principles. None of them will surprise anyone who has lived through an outsourcing review. - **Governance and accountability.** A named, senior person owns each AI system and answers for what it does. - **Cybersecurity.** AI tools are new doors into your systems and your data. Treat them that way. - **Privacy and data quality.** Know what data goes in, where it goes, and whether it is fit for the job. - **Regulatory compliance, consumer protection and fair treatment.** AI does not suspend the Handbook or the duty to treat customers fairly. - **Transparency and explainability.** You can explain an AI-assisted decision to a reasonable degree, matched to its risk and complexity. The Commission expects the guidance to matter most for higher-impact uses, and names customer onboarding, lending decisions and investment advice. Where appropriate, material AI use cases and their controls belong in your business risk assessment. ## The four-trigger test Here is how we turn the guidance into a decision. For any AI use case, ask four questions. Does it touch regulated activity? Does it affect a customer outcome? Does it handle confidential data? Does it feed a regulatory submission? The four triggers come from the Commission. The three levels below are our working method for applying them, not JFSC terminology. - **Light touch.** No triggers. An approved tool, no client data, sensible staff guidance. Done. - **Controlled.** Confidential data only, and the output stays internal as a draft a person owns. The controls are about the data: an enterprise tool, data kept inside your own environment and not used to train anyone’s model, access that mirrors your existing permissions, and retention you have set on purpose. - **Material.** Regulated activity, a customer outcome or a regulatory submission. Now it goes in the business risk assessment, with a named owner, testing before go-live, a human sign-off, a log of every run and a plain-English account of how it works. One more rule sits on top. Within Material, anything that acts on its own, by sending, paying or deciding without a person in between, gets the heaviest treatment of all. Do not build one of those first. ## Ten use cases, sorted These are the ten we see most often in Jersey trust companies and fund administrators. Your list will differ in the detail, not the shape. 1. **Summarising public regulatory updates.** Light touch. Public material in, internal briefing out. No triggers. 2. **Drafting marketing copy and job adverts.** Light touch. Check the facts, as you would for any draft. 3. **Internal meeting notes.** Controlled. In a trust business almost every meeting mentions a client, so the confidential data trigger fires. The fix is the tool, not a ban: use the enterprise version inside your own environment, never a free app on someone’s phone. 4. **Drafting routine client emails that a person reviews and sends.** Controlled, with a warning. The moment the drafts contain advice, they affect a customer outcome and move up to Material. 5. **Searching your own document system for precedents.** Controlled. The risk is access, not output. The search must never show a junior a file they could not open themselves. 6. **Policy and procedures Q&A.** Material, and this is the one that surprises people. It looks like a staff convenience. But if an administrator uses it to decide how to apply your AML procedures, it touches regulated activity. Test it against the real manual and log what it answers. 7. **KYC document extraction with a human check.** Material. Customer due diligence is regulated activity. Worth doing anyway: our [KYC write-up](/writing/ai-for-kyc-client-take-on) shows how it takes hours out of every case. The human check is the control. 8. **Sanctions and PEP alert triage.** Material. The AI can rank and summarise alerts. A person closes every one, and the log shows who and why. 9. **Preparing data for regulatory returns.** Material by definition, because it feeds a regulatory submission. Reconcile the output to source before anyone signs. 10. **A client-facing chatbot answering account questions.** Material, and it acts on its own. Customer outcome, confidential data, no person in between. If you build this, build it last. The tally is two Light touch, three Controlled and five Material. Do not read that as five red lights. Material is not a ban. It is a list of controls, and the most valuable work in a regulated firm lives there, because that is where the hours are. ## What the register needs Your business risk assessment needs to see the material use cases, which means you need a register first. It does not need software. A spreadsheet with eight columns will do. - **Use case**, in one sentence a director would understand. - **Owner**, by name. A person, not a team. - **Triggers**: which of the four it hits. - **Level**: Light touch, Controlled or Material. - **Tool and vendor**, including where the data is processed. - **Controls**, as they actually run today, not as the policy hopes. - **Human checkpoint**: who reviews, and at which step. - **Last reviewed**, with a date. Material entries get reviewed every quarter. If that looks like your outsourcing register, good. Same discipline, new object. Our guide to [building an AI audit trail](/writing/building-an-ai-audit-trail) covers what the logs behind each Material entry should hold, and [human-in-the-loop AI](/writing/human-in-the-loop-ai) covers where the checkpoints go. ## Guernsey reached the same place Groups with a Guernsey business can run one register for both islands. The GFSC’s [January 2026 policy statement](https://www.gfsc.gg/news/policy-statement-use-artificial-intelligence) also declines to write AI-specific rules. It asks firms to treat AI like any other technical or strategic project under the Finance Sector Code of Corporate Governance, and points to NIST’s AI Risk Management Framework, ISO/IEC 42001 and NCSC guidance for reference. Different words, same questions. We will map the wider picture, the UK and EU included, in a later post. ## The 30-day plan Thirty days is enough. Week one, find everything: ask every team what AI they use, including the free tools on their phones, because the unofficial ones are the ones that hurt. Week two, run the four-trigger test on each and name an owner. Week three, write down the controls for every Material entry and add them to the business risk assessment. Week four, take one page to the board: the register, the three biggest risks, and the decisions you need from them. The target is simple. Every AI use case in the firm on one register, with a named owner, inside 30 days. The point is speed, not paperwork. Once the sorting is done, nobody has to reargue what is allowed every time someone has a good idea. If you want the register built for you, or would rather sort it yourself and talk through the Material entries, [book a review](/book-a-review). Either way: sort first, build second. --- # The JFSC’s AI guidance: what it actually means for your firm Published 13 July 2026 by Spark · Governance and compliance · 5 min read Canonical: https://hellospark.co.uk/writing/jfsc-ai-guidance-what-it-means > The Jersey regulator has set out how it expects firms to adopt AI. It is principles-based, proportionate, and lands the responsibility firmly with the board. Here is what it means in practice, and the four moves worth making now. The headline is simple. The JFSC is not going to stop you using AI. It is going to ask who signed it off, whether you can explain what it does, and what happens when it gets something wrong. If you can answer those three questions on any given system, you are most of the way there. The Commission set out its position in [AI in financial services: enabling innovation with confidence](https://www.jerseyfsc.org/news-and-events/ai-in-financial-services-enabling-innovation-with-confidence/), and it sits inside a wider 2026 to 2030 strategy that pairs strong regulation with a deliberately business-friendly stance. The tone is pragmatic. AI is already reshaping how firms operate, Jersey competes internationally for the same work, and the regulator would rather help firms adopt it well than watch them either freeze or wing it. It is principles-based rather than a prescriptive rulebook, which is good news and a trap at the same time. Good, because it is proportionate. A trap, because principles put the thinking back on you. ## The four things it expects Strip out the framing and the guidance leans on four ideas that recur across the Commission’s public statements and the broader work of the Jersey AI Council. 1. **Governance and accountability.** Someone senior owns the AI, by name. The board understands what is running, why, and what it could get wrong. This is not a committee you convene once. It is a standing line of responsibility. 2. **AI-literate leadership.** The people accountable actually understand the technology’s limits. You cannot govern what you do not understand, and ‘the vendor said it was fine’ is not a control. 3. **Explainability.** You can describe, in plain terms, how a system reaches an output and why. Large language models return the most likely answer given the input, not the same answer every time, so consistency and traceability have to be engineered in rather than assumed. 4. **Meaningful human oversight.** A person can review, override, and stop the system, and that oversight is real rather than a tick-box. The level of oversight should scale with how much damage the use case could do. None of this is exotic. If you run a regulated firm, you already apply the same logic to outsourcing, to your AML framework, and to any material third-party dependency. AI is a new object inside an old discipline. ## Proportionate is the operative word The guidance does not treat every use of AI the same, and neither should you. A model that drafts an internal file note is not the model that decides whether to onboard a client. The oversight, testing, and documentation you put around a system should track the harm it could cause if it fails. A useful test: sort every AI use case into three buckets. Low stakes, where a wrong output is an inconvenience. Medium, where a human catches it before it matters. High, where a wrong output reaches a client, a regulator, or a payment. Most firms find 80 percent of their genuinely useful cases sit in the first two buckets, which is exactly where you should start. ## What this does not mean It does not mean you need an AI policy the length of your AML handbook before you switch anything on. It does not mean every tool needs a bespoke risk assessment signed in triplicate. And it does not mean waiting. The Commission has been explicit that it welcomes early conversations and would rather engage with firms that are moving thoughtfully than reward those sitting on their hands. Standing still is its own risk, and it is not the one the guidance is worried about. ## Four moves worth making in the next 90 days 1. **Name an owner.** Put one senior person’s name against AI adoption. Not IT alone, not a working group. A person with a seat at the table. 2. **Build an AI register.** A single list of every AI tool in use, what it does, what data it touches, and which risk bucket it sits in. Most firms are surprised by what is already running once they look. A first pass takes about a day. 3. **Write one page, not fifty.** A short AI use policy that says what is allowed, what needs sign-off, and what is off limits. One page people read beats fifty nobody opens. 4. **Make oversight and audit real on your highest-stakes case.** Pick the one use case that could actually hurt if it failed, and build the human checkpoint and the audit trail properly. Prove the pattern once, then reuse it. We wrote separately about [building an AI audit trail](/writing/building-an-ai-audit-trail) and about [what human-in-the-loop actually requires](/writing/human-in-the-loop-ai), because those two are where firms most often wave their hands. Both are worth getting right before a supervisor asks. ## How we read it This guidance rewards firms that treat AI as an operational discipline rather than a science project. That is the whole of our approach. We scope the smallest painful workflow, build it so the audit trail exists from day one, log every change, and keep a human in the loop where it counts. The regulator nods rather than flinches, because the governance was not bolted on at the end. If you want a grounded view of your own position, our [AI governance work for Channel Islands firms](/writing/ai-governance-jersey-jfsc) covers the framework in more depth, and a 30-minute [review](/book-a-review) will get you a candid read on where you actually stand. No deck, no hype. Just what to fix first and what it is worth. One caveat, stated plainly. Regulatory guidance evolves, and the exact wording and scope of what the JFSC publishes may move. Treat this as our interpretation of the Commission’s stated direction, not as legal advice, and check the primary source before you rely on any specific point. --- # Where to start with AI: pick the smallest painful workflow first Published 6 July 2026 by Spark · Adoption and strategy · 4 min read Canonical: https://hellospark.co.uk/writing/where-to-start-with-ai > Do not start with a strategy. Start with the smallest painful workflow you can name. Here is how to find the one first use case that pays for itself, teaches your team, and gives the regulator nothing to flinch at. Start with the smallest painful workflow you can name. Not the biggest opportunity, not the flashiest demo, not the thing a vendor put in front of you. The narrow, repetitive, annoying task that a specific person does every week and quietly resents. That is where AI earns its keep, and it is where you learn what your firm can actually absorb. Most firms get this backwards. They open with a strategy workshop, a long list of possibilities, and a plan to transform three departments at once. Six months later there is a deck and no working system. The firms that get value do the opposite. They ship one small thing, measure it, and use what they learn to pick the next one. Momentum beats ambition every time. ## Why smallest and most painful wins Small means you can finish it. A workflow you can describe in two sentences is a workflow you can build, test, and put a human check around in weeks rather than quarters. Painful means someone will actually use it, because they already feel the cost of doing it by hand. Pick something nobody minds doing and adoption dies on contact. Pick something that ruins a Tuesday afternoon every week and people will pull it into their work without being asked. There is a risk argument too. A small, well-chosen first workflow keeps you in the low and medium stakes buckets, where a wrong output is an inconvenience or gets caught by a human before it matters. You get to learn the discipline, the audit trail, the sign-off, the prompt versioning, on something that cannot hurt a client, a payment, or a regulatory return. We wrote about that risk tiering in our [AI governance framework](/writing/ai-governance-jersey-jfsc). ## How to find it Do not run a survey. Sit with the two or three people who do the most repetitive work and ask one question: what do you do every week that feels like it should not need a person. Then look for the four markers of a good first candidate. - **High frequency.** It happens daily or weekly, not twice a year. Frequency is what turns small time savings into real numbers. - **Rules you can write down.** The task follows a pattern a competent new joiner could learn in a morning. If the logic lives only in one person’s head and changes every time, it is not your first case. - **Structured or semi-structured inputs.** Documents, emails, spreadsheets, forms. Messy is fine. Genuinely unpredictable is not. - **A tolerable failure mode.** If the system gets one in twenty wrong, a human catches it and the cost is small. That is a starting point. If one wrong output reaches a client unchecked, that is a later case, not a first one. In regulated firms the same handful of workflows come up again and again. Client take-on and KYC document gathering. Extracting figures from statements and reports. First-draft file notes and letters. Investor and NAV reporting packs. Triaging a shared inbox. None of these are exotic. All of them are painful, frequent, and rule-bound, which is exactly the point. ## Put a number on it before you build A workflow is only worth automating if you can say what it is worth. Before any build, write the number down. Take the hours the task consumes each week, multiply by the loaded cost of the person doing it, and annualise. A task that eats six hours a week of a person on a fifty pound loaded hourly rate is costing you the better part of sixteen thousand pounds a year, and that is before you count the errors and the delay. If you cannot put a number on the outcome, it is not the right first workflow. There is a fuller method in [how to measure AI ROI](/writing/how-to-measure-ai-roi). ## One workflow, not ten The discipline is to resist the second idea until the first one is live and measured. Ship one workflow. Watch it run for a fortnight. Confirm the hours it saved and the errors it removed. Only then pick the next, and let the pattern you proved carry over: the same audit trail, the same human checkpoint, the same way of versioning prompts. This is how a firm builds capability instead of a graveyard of pilots. It is also why so many pilots never reach production, which we covered in [why AI pilots fail](/writing/why-ai-pilots-fail). ## What good looks like in six weeks A first workflow scoped, built, and running in production inside four to six weeks, with a named owner, a human sign-off where it counts, and a changelog from day one. That is the shape of how we work, built precisely so the first thing you ship is the smallest painful workflow rather than the biggest slide. If you want a candid read on which workflow to start with, [book a review](/book-a-review). We will look at your week, find the task that hurts most and risks least, and tell you what it is worth to fix. No deck, no hype. Just the first move worth making. --- # RAG explained: how retrieval makes business AI reliable Published 25 June 2026 by Spark · Technology, explained · 5 min read Canonical: https://hellospark.co.uk/writing/rag-explained > A general-purpose model knows a lot about the world and nothing about your firm. RAG fixes that. It feeds the model your own documents at the moment of the question, so answers are grounded in your sources and can be checked. Here is how it works and when you actually need it. RAG is how you stop an AI making things up. A general-purpose model like Claude or GPT knows a great deal about the world and nothing at all about your firm. Ask it about your onboarding policy and it will guess, confidently, because guessing is what these models do when they have no source. Retrieval-augmented generation closes that gap. It finds the relevant passages from your own documents and hands them to the model at the moment of the question, so the answer is built from your material rather than the model’s memory. That one change is the difference between a party trick and a tool a regulated firm can put in front of staff. ## What the letters actually mean Retrieval-augmented generation is three steps in a trench coat. Retrieval: search your documents for the passages most relevant to the question. Augmented: paste those passages into the prompt alongside the question. Generation: let the model write an answer using that supplied context. The model still does the writing. What changes is that it is now writing from evidence you chose, not from a hazy statistical recollection of the entire internet. The retrieval step usually leans on something called a vector search. Your documents are chopped into chunks, each chunk is turned into a list of numbers that captures its meaning, and the same is done to the question. The system then finds the chunks whose meaning sits closest to the question, even when the wording is different. Ask about ‘proof of address’ and it will surface the paragraph about ‘residential verification documents’ without you having to match the exact phrase. That is the part keyword search never managed. ## Why it reduces hallucination A model hallucinates when it is asked for something it does not know and has no way to say so. Left to its own devices it produces the most plausible-sounding answer, which is often wrong in ways that read perfectly. RAG removes the vacuum. Instead of ‘what do you know about our fee policy’, the model is effectively asked ‘here are the three relevant paragraphs from our fee policy, now answer’. The scope narrows, the source is present, and a well-instructed system will say ‘that is not covered in the documents provided’ rather than invent. You never get to zero. You get from unacceptable to manageable, and for most business work that is the whole game. ## Why citations matter more than the answer The quiet superpower of RAG is that it can show its working. Because the answer was built from specific retrieved passages, the system can tell you which document and which section each claim came from. For a regulated firm that is not a nice-to-have. It is the difference between an output a supervisor can inspect and a black box nobody can defend. A staff member reading a RAG answer can click through to the source clause and confirm it in seconds, which means the human check stays real rather than becoming a rubber stamp. We think citations should be non-negotiable in any system that touches client work, and we build them in from day one. ## When your firm actually needs it You need RAG when the right answer lives in your own material and changes over time: policies, procedures, contracts, past correspondence, a knowledge base, a filing cabinet of PDFs. The classic cases in our world are a staff assistant that answers ‘how do we handle this’ from the firm’s own procedures, a tool that pulls the relevant clauses out of a long agreement, or a search layer over years of client files. If the useful knowledge is yours and specific, RAG is usually the honest answer. You do not need it for everything. Summarising a document you have already pasted in needs no retrieval, because the source is right there. Drafting from a template, translating, reformatting, tone-fixing: none of these are RAG problems. And RAG is not a substitute for structured data. If the question is ‘what is the NAV’, that belongs in a database query, not a document search. Reaching for RAG when a simple lookup would do is a common and expensive mistake. ## Where it goes wrong RAG is only as good as what it retrieves. Feed it messy, duplicated, out-of-date documents and it will faithfully ground its answers in your mess. Garbage in, confidently cited garbage out. The unglamorous work is in the retrieval: chunking documents sensibly, keeping the source current, and testing that the right passages actually surface for real questions. Most RAG projects that disappoint were under-built on that side and over-built on the chatbot. A related discipline is getting clean data out of the documents in the first place, which we cover in [AI document extraction](/writing/ai-document-extraction). ## How we build it We treat RAG as plumbing, not magic. Scope the smallest set of documents that answers a real, repeated question. Get retrieval accurate before worrying about the wording of answers. Insist on citations. Keep a human in the loop wherever the output could reach a client or a regulator. Models such as Claude handle the generation step well because they follow the instruction to stay inside the supplied sources and to admit when something is not there, which is exactly the behaviour you want in regulated work. If you have a shelf of documents that staff keep asking about, that is a RAG-shaped problem and a good first build. [Book a review](/book-a-review) for a candid read on whether it is worth doing, where retrieval would earn its keep and where a simpler tool would do the job. --- # AI governance for Channel Islands firms: getting ahead of the JFSC Published 18 June 2026 by Spark · Governance and compliance · 3 min read Canonical: https://hellospark.co.uk/writing/ai-governance-jersey-jfsc > Good AI governance is not a policy document. It is a small set of controls a regulator can inspect and a team can actually live with. Here is the framework we use with Channel Islands firms. Most AI governance fails in the same way. A firm writes a long policy, files it, and changes nothing about how work is actually done. The policy is for the auditor. The work is for everyone else. A regulator can tell the difference in about ten minutes, and so can your team. Governance that holds up is smaller and more boring than that. It is a handful of controls that are visible, owned, and used every day. The JFSC’s [principles-based approach](/writing/jfsc-ai-guidance-what-it-means) rewards exactly this: not the thickest binder, but the clearest line of responsibility. Here is the framework we put in place, and roughly what each part costs to stand up. ## 1. A named owner and a real mandate One senior person owns AI adoption, by name, with a seat at the table and the authority to stop a system going live. Not the IT team on its own. Not a committee that meets quarterly. When something goes wrong, the regulator will ask who decided, and ‘the working group’ is not an answer. This costs nothing but a decision, and it is the single most effective control you have. ## 2. An AI register A single, current list of every AI system in use: what it does, what data it touches, who owns it, and how much damage it could do if it failed. Most firms discover more shadow AI than they expected once they look, because staff have quietly adopted tools that were never assessed. A first version takes about a day to build and immediately tells you where the risk actually sits. ## 3. Risk tiering, so effort tracks harm Sort each use case into low, medium, or high stakes by what a wrong output could reach. A drafted internal note is low. A client-facing letter with a human check is medium. Anything that touches onboarding decisions, a payment, or a regulatory return is high. The controls, testing, and documentation scale with the tier. This is what proportionate means in practice, and it stops you spending high-stakes effort on low-stakes tools. ## 4. Human oversight where it counts For every high-stakes case, a person reviews and can override the output before it takes effect, and that review is genuine rather than a rubber stamp. We covered the difference in detail in [human-in-the-loop AI](/writing/human-in-the-loop-ai). The short version: if the human cannot realistically catch an error in the time they are given, you do not have oversight, you have theatre. ## 5. An audit trail by default Every material AI decision leaves a record: the input, the model and prompt version, the output, and who approved it. Built in from the start this is nearly free. Retrofitted after a supervisor asks, it is painful and sometimes impossible. Our piece on [building an AI audit trail](/writing/building-an-ai-audit-trail) sets out exactly what to capture. ## 6. A one-page policy people read What is allowed, what needs sign-off, what is off limits, and who to ask. One page. The fifty-page version exists to protect the firm on paper. The one-page version actually changes behaviour, which is the only thing that protects you in reality. ## What this looks like on a timeline A firm of 20 to 250 people can stand up the whole framework in four to six weeks without pausing any live work. Week one, name the owner and draft the register. Weeks two and three, tier the use cases and write the one-page policy. Weeks four to six, make oversight and audit real on the two or three highest-stakes systems. After that it is maintenance, not a project. This is the spine of how we work, and it is why the systems we build for trust and corporate services and fund administration firms tend to survive contact with a regulator. Governance is not the tax you pay for using AI. Done early, it is the thing that lets you use it at all. If you want a candid read on where your firm stands, [book a review](/book-a-review). --- # AI for KYC and client take-on in trust and corporate services Published 11 June 2026 by Spark · Sector playbooks · 5 min read Canonical: https://hellospark.co.uk/writing/ai-for-kyc-client-take-on > Client take-on is the most manual, least billable week in a trust administrator’s month. AI can collect the documents, extract the data, run the screening and draft the risk assessment. The decision stays human, and the audit trail is there from the first click. Client take-on is the slowest, least billable stretch of a trust administrator’s month, and it is the one AI shortens most. Collect the documents, read them, extract the data points, run the screening, draft the risk assessment: almost all of it is structured, repeatable work sitting behind a human decision. Automate the assembly and the drafting, keep the judgement and the sign-off with a person, and a take-on that used to run over days lands in hours. In our experience the manual effort on a standard onboarding drops by 60 to 70 percent, and the wait is measured in hours rather than a fortnight of chasing. This is the workflow we build most often for trust and corporate services firms, because the pain is universal and the shape of the fix is well understood. Here is how it breaks down, and where the human has to stay. ## Document collection stops being a chase Most take-on delay is not analysis. It is waiting. Waiting for a passport, a proof of address, a structure chart, a source-of-funds explanation that never arrives in one go. An automation can send the request list, track what has come back, read each document on arrival, and chase the gaps on a schedule without an administrator holding the whole thing in their head. The system knows what a complete file looks like and tells you exactly what is still missing. That single change, turning a manual chase into a tracked queue, often recovers more time than the clever extraction that follows. ## Extraction and verification, checked not trusted Once a document arrives, AI reads it and pulls out the fields your onboarding system needs: names, dates, addresses, ownership percentages, the shape of the structure. This is the same [document extraction](/writing/ai-document-extraction) engine that powers the rest of the back office, tuned to the forms you actually receive. The important word is verification. The model proposes; it does not decide. Where a passport photo is unreadable or an address does not match, the case is flagged rather than guessed. High-confidence fields flow through, low-confidence fields wait for a human. Getting that threshold right is most of the work, and it is why we test against your own historic files before anything goes live. ## Screening and risk-scoring, assembled for a reviewer Screening against sanctions, PEP and adverse-media lists is already largely automated in most firms. The value AI adds is upstream and downstream: pulling the right names and entities out of the structure so nothing is screened by hand, and then assembling the hits into an evidence pack a reviewer can actually work through. Instead of a raw list of 40 possible matches, the reviewer sees each one summarised, ranked by relevance, with the underlying source one click away. A first-pass risk assessment gets drafted from the same inputs, in your house rating logic, ready for a human to accept, adjust or reject. The draft is a starting point, never the verdict. ## The decision stays human, and that is the point Nothing here approves a client. The system does the collection, the reading, the screening and the drafting; a named person makes the call. This is not a compliance nicety, it is the design. The JFSC expects [meaningful human oversight](/writing/human-in-the-loop-ai) that scales with how much damage a use case could do, and onboarding a client is about as high-stakes as it gets. So the human checkpoint is real: the reviewer sees the evidence, can override any machine output, and signs the decision themselves. If a wrong output could reach a regulator or open the door to dirty money, a person stands between the model and the outcome. ## The audit trail is there from the first click Every take-on leaves a complete record: which documents came in and when, what the model extracted, what it flagged, which screening ran against which lists, the draft risk assessment, and who approved the final decision, with the prompt and model version logged alongside. Built in from the start this is nearly free. Bolted on after a supervisor asks, it is painful and often impossible to reconstruct. We covered exactly what to capture in [building an AI audit trail](/writing/building-an-ai-audit-trail). For a regulated firm this is the difference between a system a supervisor nods at and one that gets you a finding. ## Periodic reviews reuse the same machine The best part is that take-on and periodic review are the same workflow pointed at different moments. Once the collection, extraction, screening and drafting are built for onboarding, a trigger-event or scheduled review reuses all of it: refresh the documents, re-run the screening, pull out what has changed since last time, and draft the review pack ready for sign-off ahead of the deadline. Firms that dread the periodic review backlog find it stops being a backlog. The pack assembles itself; the administrator reviews and approves. Build once, use twice. ## What it takes to stand up Before you build a take-on assistant, it is worth mapping the whole process first so you automate the right client type in the right order. Start narrow. Pick your highest-volume, most painful onboarding flow, prove the pattern with the audit trail and the human checkpoint working properly, then extend it to the next structure and to periodic reviews. That is how the numbers land and how the regulator stays comfortable. If client take-on is eating your team’s week, a 30-minute [review](/book-a-review) will get you a candid read on which part to automate first and what it is worth. No deck, no hype. Just the smallest painful workflow and a number against it. --- # How to measure ROI on AI automation, with a worked example Published 2 June 2026 by Spark · Adoption and strategy · 4 min read Canonical: https://hellospark.co.uk/writing/how-to-measure-ai-roi > AI ROI is not mysterious. It is hours saved times loaded cost, plus errors removed, plus capacity freed. Here is the framework we use, and a worked example that shows the maths on a real-shaped workflow. AI ROI is simpler than the market wants you to believe. Three things move the number: hours saved, errors removed, and capacity freed. Put a figure on each, subtract what the system costs to build and run, and you have your answer. If you cannot do that sum for a given workflow, you are not ready to build it yet. The reason ROI feels slippery is that most firms measure the wrong thing. They count licences and forget loaded cost. They count the demo and forget the run rate. They talk about time saved but never actually time the task before and after. Measure properly and the picture is usually clear within the first month of a workflow going live. ## The three levers Every credible AI business case pulls on at least one of these. The best pull on all three. - **Hours saved.** The time a task took by hand, minus the time it takes with the system and its human check, multiplied by the loaded hourly cost of the person doing it. Loaded cost, not salary. Salary plus employer costs, plus overhead, is usually 1.3 to 1.5 times the headline figure. - **Errors removed.** The cost of the mistakes the old process made: rework, corrections, an apology to a client, in the worst case a breach to report. Harder to price precisely, but real. Estimate a per-error cost and a rate, and be conservative. - **Capacity freed.** The work your team can now take on without hiring, or the growth you can absorb without adding headcount. This is often the largest number and the one firms forget, because it does not show up as a cost line, it shows up as a hire you did not make. ## The formula Annual benefit is hours saved per year times loaded hourly cost, plus errors removed per year times cost per error, plus the value of any capacity freed. Annual cost is the build cost amortised over its useful life, plus the running cost: model usage, tooling, and the maintenance the system needs to stay honest. Net benefit is the first minus the second. Payback is the build cost divided by the monthly net benefit. Keep it that plain. ## A worked example Take a trust and corporate services team drowning in client take-on. Gathering and checking KYC documents for new entities takes an administrator about eight hours per case, and the team onboards roughly ten cases a month. That is 80 hours a month, or 960 hours a year, on one workflow. Build an assisted pipeline: documents extracted and checked against a checklist automatically, gaps flagged, a first-pass summary drafted, and a human reviewing and signing off every case. In our experience this kind of workflow takes the eight hours per case down to about three, because the person is now checking and approving rather than gathering and typing. Call it five hours saved per case, fifty hours a month, 600 hours a year. Now the maths. At a loaded cost of £45 an hour, 600 hours saved is £27,000 a year. The old process missed a required document or logged something wrong on maybe one case in ten, each costing a few hours of rework and the occasional awkward client email; call that conservatively £4,000 a year avoided. And the freed capacity lets the team take on more onboarding without a new hire, which was the actual bottleneck. Ignore that entirely and you are still at £31,000 a year of hard benefit. Against that, a build of this shape sits in the region of £40k, plus running costs of a few hundred pounds a month for model usage and tooling. Amortise the build over three years and add the run rate, and net benefit lands comfortably in five figures a year, with payback inside the first eighteen months and often sooner once freed capacity is counted. That is the pattern behind our [KYC and client take-on work](/writing/ai-for-kyc-client-take-on). ## The mistakes that flatter or wreck the number Two errors show up constantly. The first is counting time saved that nobody actually recovers, because the freed hours just evaporate into slack rather than into more work or fewer hours. If you cannot say where the saved time goes, discount it. The second is forgetting the human check in the cost. Oversight takes time, and a business case that assumes zero review is measuring a system you should not run. Price the checkpoint in. It is cheaper than the alternative. ## Measure it after, not just before A business case is a forecast. The number that matters is the one you confirm a month after go-live, by timing the task again and counting the errors again. Build the measurement in from day one: log every run, every human override, every correction. That same log is your audit trail, which is why we treat measurement and governance as the same job, covered in [building an AI audit trail](/writing/building-an-ai-audit-trail). If you want it built and proven, we ship the workflow and the measurement together. And if you would rather just talk it through, [book a review](/book-a-review) and we will put a real number on your most painful workflow. --- # Agents, chatbots and automations: what the difference actually is Published 27 May 2026 by Spark · Technology, explained · 4 min read Canonical: https://hellospark.co.uk/writing/ai-agents-chatbots-automations > Three words get used interchangeably and shouldn’t be. A chatbot talks. An automation runs a fixed path without being asked. An agent decides what to do next. Confusing them is how firms buy the expensive one for a job the cheap one does better. Here is the honest breakdown. These three words are not interchangeable, and treating them as if they are is how firms overspend. A chatbot talks to a person. An automation runs a fixed sequence of steps without anyone asking it to. An agent decides what to do next by itself. Different shapes, different costs, different risks. The expensive, fashionable option is the agent, and most of the value in a professional services firm sits in the boring automation nobody is tweeting about. Get the category right and the project usually works. Get it wrong and you have bought a self-driving car to fetch the post. ## A chatbot: it answers A chatbot is a conversation. A person asks something in plain language, the system replies. That is the whole of it. The good ones are grounded in your own documents so the answers are accurate and can be traced back to a source, which is the retrieval approach we cover in [RAG explained](/writing/rag-explained). The weak ones are grounded in nothing and will happily invent a policy that does not exist. Chatbots fit where a human needs an answer from a body of knowledge: an internal assistant over your procedures, a client-facing helper on your website, a search layer over years of files. What a chatbot does not do is act. It tells you the fee is due. It does not raise the invoice. The moment you want something to happen in the real world, you have left chatbot territory. ## An automation: it runs An automation is a fixed path that runs on a trigger. When an email lands, extract the attachment, pull out the key fields, drop them into the system, flag anything odd for a human. Every step is defined in advance. It does the same thing the same way every time, which is precisely what makes it trustworthy and cheap to run. There is often AI inside an automation, doing the one clever step a rules engine cannot, like reading a scanned document. But the path around that step is fixed and predictable. This is where most of the real return lives. The repetitive, rules-bound, high-volume back-office work that eats your team’s week is almost always an automation problem. It is unglamorous, it is measurable, and it pays back fast. If you only ever do one AI thing, it should probably be this. ## An agent: it decides An agent is given a goal and works out the steps itself. Instead of a fixed path, it has a set of tools and the freedom to choose which to use, in what order, until the goal is met. ‘Research this counterparty and draft a summary’ might have the agent search, read, cross-check, and write, deciding each move as it goes. When the task genuinely varies every time and cannot be reduced to a fixed sequence, that flexibility is the point. That same flexibility is the risk. An agent that chooses its own path can choose a wrong one, and the more it can do without asking, the more damage a wrong choice makes. In regulated work that is a live concern, not a hypothetical. Agents are the right tool for genuinely open-ended tasks. They are the wrong tool for a process you could have written down as steps, which, honestly, is most processes. The industry is selling agents hard right now. Be sceptical when the job on the table has a fixed shape. ## How to tell which you need Ask two questions. First, does a human need to have a conversation, or does something need to happen? Conversation points to a chatbot. Something happening points to automation or an agent. Second, is the path the same every time, or does it genuinely differ? Same path is an automation. Genuinely different every time is an agent. Most firms discover their pressing problems are chatbots and automations, and that the agent can wait until those are earning. - **Chatbot** when a person needs a grounded, traceable answer from your knowledge. - **Automation** when a defined process runs the same way on a trigger, with a human checking the exceptions. - **Agent** when the task is genuinely open-ended and the steps cannot be fixed in advance, and you have built the oversight to match. ## Why the label matters commercially Because it sets the cost, the risk and the governance. An automation with a human checking exceptions is cheap to run, easy to audit, and easy to explain to a regulator. An agent making its own decisions needs far more oversight, testing and change control before it goes near client work, and it should. Buying the agent for a job the automation does better is the single most common way we see AI budgets wasted. The related question of building it yourself, buying it, or wiring together no-code tools is covered in [build, buy or no-code](/writing/build-buy-or-no-code). If you are not sure which of the three your problem is, that is the useful first conversation. [Book a review](/book-a-review) and we will tell you plainly whether you need the clever thing or the boring one. Usually the boring one, and usually that is good news. --- # Automating NAV and investor reporting in fund administration Published 20 May 2026 by Spark · Sector playbooks · 4 min read Canonical: https://hellospark.co.uk/writing/automating-nav-investor-reporting-fund-administration > The NAV cycle is a deadline machine that runs on manual assembly. Automate the reconciliation, the validation and the drafting, surface only the exceptions, and the cycle runs faster with fewer late nights. The numbers still get checked by a human. The NAV cycle is a deadline machine, and most of what your team does inside it is assembly. Gather prices and positions from the same places, reconcile them, chase the exceptions, then draft the same investor documents under the same monthly or quarterly pressure. Automate the assembly and the reconciliation, let AI handle the messy documents and the first drafts, and surface only the exceptions that need a person. Done well, the manual effort on a reporting cycle drops by half or more, and the late nights at period end stop being the norm. The numbers still get checked and signed by a human. That never changes. This is the core of the work we build for fund administrators, and it splits cleanly into three jobs: get the data right, reconcile it, then draft from it. Here is each one, and where the human stays in the loop. ## Data validation before anything else A NAV is only as good as the data feeding it, and the data arrives in a mess: custodian statements as PDFs, price files in three formats, a manager’s spreadsheet with a column moved since last month. The first win is turning all of that into clean, structured data automatically, which is the same [document extraction](/writing/ai-document-extraction) problem the whole back office shares. Prices, positions, cash movements and corporate actions get read, normalised and validated against expectations before they touch the NAV. A price that has moved 40 percent overnight or a position that appeared from nowhere gets flagged at the door, not discovered three steps later when it is far more expensive to unpick. ## Reconciliation that surfaces only the exceptions Reconciliation is where administrators lose their evenings, and it is the most automatable step in the whole cycle. Matching positions and cash across custodians, the accounting system and the manager’s records is rule-based work at heart. Automation runs the match, applies your tolerances, and presents only what breaks: the handful of genuine exceptions, each one with the two sides laid side by side and a plausible explanation drafted. Instead of an analyst working through a thousand matched lines to find the five that do not, they open a queue of five. The tedious 95 percent runs itself; the human attention goes where the judgement actually is. ## Drafting investor reports and capital call notices Once the numbers are locked, the documents write themselves from source, into your templates. Quarterly investor reports, capital call notices, distribution notices and regulatory returns are all structured outputs built from data you already hold. AI drafts each one, personalised per investor, with the figures pulled straight from the validated model rather than re-keyed. A capital call that took an afternoon of copy-paste per investor becomes a batch generated in minutes and reviewed as a set. The commentary in an investor report can be drafted too, in your house tone, for the fund controller to edit rather than write from a blank page. Drafting is the machine’s job. Sign-off is not. ## The checks are the whole point None of this releases a NAV or sends an investor notice on its own. Every number the automation produces is checked against your model, and every document waits for a human to review and approve before it goes out. This is not caution for its own sake. Getting a capital call amount wrong or publishing a NAV with a stale price is the kind of failure that reaches investors and regulators, which puts it squarely in the high-stakes tier where [meaningful human oversight](/writing/human-in-the-loop-ai) is not optional. The reviewer sees the exceptions, the drafts and the underlying figures, can override anything, and signs the release themselves. The automation removes the assembly, not the accountability. ## An audit trail the depositary can inspect Every cycle leaves a complete, inspectable record: which data came in from where, what was flagged and why, how each exception was resolved and by whom, and who approved the final NAV and each investor document, with the prompt and model version logged. For a regulated fund administrator, and for the depositary and auditors sitting behind you, that trail is not a nice-to-have. Build it in from day one and it is nearly free. Retrofit it after a query and it is painful, sometimes impossible. It is the difference between a fast cycle you can defend and a fast cycle you cannot. ## Where to start Do not try to automate the whole cycle at once. Pick the single most painful step, usually the reconciliation or the investor report drafting, and prove the pattern there with the checks and the audit trail working properly. It pays to map the sequence first so you build in the order that returns the most soonest. Prove it once, then roll it across fund types and across the cycle. If your reporting cycle runs on late nights and manual assembly, a 30-minute [review](/book-a-review) will get you a straight read on which step to automate first and what it is worth in hours recovered per cycle. No deck. Just the number. --- # Building an AI audit trail: how regulated firms keep AI decisions defensible Published 12 May 2026 by Spark · Governance and compliance · 4 min read Canonical: https://hellospark.co.uk/writing/building-an-ai-audit-trail > An AI audit trail is five fields and a habit. Capture the input, the model and prompt version, the output, the approver, and the timestamp, from day one. Retrofit it later and you are reconstructing history you no longer have. An AI audit trail is five fields and a habit. For every material decision a system helps make, you record the input it received, the model and prompt version that produced the output, the output itself, who approved it, and when. Capture those from the first day a system goes live and you can defend any decision it touched. Skip them, and you are reconstructing history you no longer have. This is the single control regulated firms most often wave their hands at. Everyone agrees they should be able to explain an AI output. Far fewer have wired up the plumbing that makes it possible after the fact. The gap only shows when a client complains, a supervisor asks, or a decision goes wrong, which is exactly the moment you cannot afford to shrug. ## The five things to capture Keep the record boring and complete. If you can reconstruct what happened without asking a person to remember, you have enough. 1. **The input.** What actually went into the model: the document, the client data, the query, the context retrieved. Not a paraphrase. The real thing, or a stable reference to it. 2. **The model and prompt version.** Which model, which version, and which prompt. A prompt is code. When it changes, the behaviour changes, and you need to know which version produced which output. We wrote about why in [prompt versioning](/writing/prompt-versioning). 3. **The output.** What the system returned, verbatim, before any human edited it. The edited final version matters too, but the raw output is what proves what the model did. 4. **The approver.** Who reviewed it, who signed it off, and whether they changed anything. A name, not a role. ‘Reviewed by compliance’ is not an approver. 5. **The timestamp.** When each step happened. Input received, output generated, human approved. Order and timing are half the story when something is contested. That is the whole schema. Five fields, logged automatically, on every decision that could reach a client, a payment, or a regulatory return. Low-stakes uses, like a first draft of an internal note, do not need the full treatment. Match the record to the risk, the same way the [JFSC’s guidance](/writing/jfsc-ai-guidance-what-it-means) expects effort to track harm. ## Why retrofitting is painful Built in from the start, an audit trail is nearly free. The system writes each field as it works, and nobody has to remember anything. Bolted on after a supervisor asks, it is expensive and sometimes impossible. You cannot log a prompt version you never recorded. You cannot recover the input if it was overwritten. You cannot prove who approved something when the system never asked for a name. In our experience most firms that skip this end up rebuilding the workflow anyway, usually under time pressure, and often after the exact incident the trail would have explained. Two weeks of clean logging design at the start saves a scramble later. That is the trade, and it is not close. > You cannot log a prompt version you never recorded. Build the trail before you need it, because you will need it on the day you least expect. ## What a supervisor actually wants to see A regulator is not trying to catch you using AI. They are trying to establish whether you are in control of it. When they ask about a specific decision, they want a short, honest answer to three questions. What did the system do? Who checked it? And what would have happened if it got it wrong? A clean audit trail answers all three in minutes. What unsettles a supervisor is not a mistake. Firms make mistakes. It is the inability to explain one. A firm that can pull the input, show the prompt version, point to the human who approved it, and demonstrate the checkpoint that would have caught an error, looks like a firm in control. A firm that says ‘the AI did it and we are not sure how’ does not. ## Make it real on one workflow first Do not try to instrument every system at once. Pick your highest-stakes AI workflow, the one that could actually hurt if it failed, and build the five-field trail properly there. Prove the pattern, then reuse it. In our work this is a fortnight of effort on a well-scoped case, and it is the part of the work that most reliably survives contact with a regulator. The audit trail is only half the control. The other half is the human doing the approving, and whether that review is genuine or theatre. We covered that in [human-in-the-loop AI](/writing/human-in-the-loop-ai). Get both right on one workflow and you have a template for every AI system you build after it. If you want a candid read on whether your current AI decisions are defensible, [book a review](/book-a-review). We will look at one live workflow, tell you what is missing, and put a number on what it takes to fix. No deck, no hype. Just what to capture and why. --- # Build, buy or no-code: choosing the right automation approach Published 6 May 2026 by Spark · Adoption and strategy · 4 min read Canonical: https://hellospark.co.uk/writing/build-buy-or-no-code > Build, buy, or no-code is the wrong question until you know the workflow. Here is how we choose between them for regulated firms, weighing total cost of ownership, lock-in, and whether a regulator can inspect the result. There is no universally right answer to build, buy, or no-code. There is only the right answer for a specific workflow, at a specific firm, at a specific stage. Get the workflow clear first and the choice usually makes itself. Pick the approach before you understand the work and you will pay for it twice. The three options are not rivals so much as tools for different jobs. Buy an off-the-shelf product when a good one exists for a common problem. Use a no-code platform when the logic is yours but the plumbing is standard. Build custom when the workflow is core, unusual, or has to satisfy controls that a generic tool cannot. Most firms end up with a mix, and that is the correct outcome. ## When buying wins Buy when your problem is a solved problem and someone reputable has already solved it. Email security, document storage, an established KYC screening provider: you are not going to out-build a specialist, and you should not try. The trade is that you accept their roadmap, their data handling, and their limits. Read the data processing terms before you sign, especially where client data leaves the island, and check that the vendor can give you the audit evidence a supervisor will eventually ask for. If they cannot tell you how the model behaves or where the data sits, that is a reason to hesitate. ## When no-code wins No-code wins when the logic is specific to you but the building blocks are standard: move data between systems, watch an inbox, call an AI model, route the result, log it. Platforms like n8n, Make, and Zapier let you assemble that in days, not months, and change it just as fast when the process shifts. That speed is the whole point. You can ship a first version, learn from it, and iterate without a development cycle for every tweak. For regulated firms the three are not interchangeable. Zapier is the quickest to start and the most opinionated, which is fine for light, standard integrations. Make gives you more control over complex logic at a similar level of ease. n8n is the one we most often reach for when auditability and control matter, because it can be self-hosted, which keeps data where you can see it and gives you a genuine log of every run. When client data cannot leave a controlled environment, self-hosted n8n is frequently the honest answer. ## When building wins Build custom when the workflow is core to how you make money, genuinely unusual, or bound by controls no off-the-shelf tool respects. A bespoke NAV and investor reporting pipeline, an extraction workflow tuned to your document set, an agent that has to leave a specific audit trail: these are worth owning. The upside is fit and control. You decide the data handling, the human checkpoints, the logging, and the prompt versioning, and nobody can change the rules from under you. The cost is that you own it, including the maintenance. ## Total cost of ownership, honestly The sticker price is the smallest part of the decision. What actually separates the three options over three years is this. - **Run rate.** Buying is a predictable subscription. No-code adds platform fees plus model usage. Custom adds hosting and the model usage, but no per-seat tax as you grow. - **Change cost.** No-code is cheapest to change, which matters most while a process is still moving. Bought tools change only when the vendor allows it. Custom changes cost developer time but have no ceiling. - **Lock-in.** Bought tools lock you into a vendor. Hosted no-code locks you into a platform, though self-hosted options like n8n soften that. Custom locks you into your own codebase, which at least you control. - **Auditability.** This is the one regulated firms underweight. Can you show a supervisor the input, the model and prompt version, the output, and who approved it? Some bought tools cannot. Self-hosted no-code and custom builds can, if you design for it from the start. ## How we actually decide Start no-code, prove the workflow, and build custom only where the value or the controls demand it. That sequence keeps you cheap and fast while you are still learning, and reserves the expensive option for the places it genuinely pays off. A great many workflows never need to graduate from a well-built n8n flow, and pretending otherwise is how firms overspend. This is also, quietly, one of the biggest reasons pilots stall, so it is worth reading alongside [why AI pilots fail](/writing/why-ai-pilots-fail). The honest version of this decision is not a spreadsheet, it is a conversation about one workflow and what it has to survive. If you want that call made properly, [book a review](/book-a-review) and we will tell you which of the three each of your workflows actually needs. --- # What MCP (Model Context Protocol) is, and why it matters for business AI Published 22 April 2026 by Spark · Technology, explained · 4 min read Canonical: https://hellospark.co.uk/writing/what-is-mcp-model-context-protocol > MCP is a common plug for AI. Instead of building a bespoke connection between every model and every system, you build one standard socket and any compliant model can use it. It sounds like plumbing because it is, and it is the plumbing that decides whether you are locked in or free to switch. MCP is a standard plug for AI. That is the whole idea, and it is more important than it sounds. Before it, connecting a model to your systems meant building a bespoke bridge for every combination: this model to that document store, that model to your CRM, each one hand-wired and each one a maintenance headache. The Model Context Protocol replaces all of that with a single agreed socket. Build the socket once and any model that speaks the protocol can plug in. It is the difference between a drawer of proprietary chargers and a world where everything takes the same cable. ## The problem it solves An AI model on its own is a brain in a jar. It can reason about what you type, but it cannot see your files, query your database, or do anything in your systems unless you connect it. Historically every one of those connections was custom work. Ten tools and three models meant a tangle of one-off integrations, each of which broke when anything changed. It was slow to build, expensive to maintain, and it quietly chained you to whichever vendor you had wired everything to. MCP, introduced by Anthropic and now adopted well beyond it, sets a common language for these connections. You expose your tools and data through an MCP ‘server’, which is just a standard adapter for one system. Any model that acts as an MCP ‘client’ can then use it. Build the adapter for your document store once, and Claude, or any other compliant model, can read from it without further work. The tangle collapses into a set of reusable parts. ## What it actually connects Two things, mostly. Tools, meaning actions the model can take, such as searching your files, querying a database, or raising a ticket. And context, meaning information the model can read to do its job. An MCP server for your practice management system might let the model look up a matter, read the file notes, and draft a next step, all through one standard interface rather than three bespoke ones. The model still only does what you permit, and every action can be logged, which matters when a supervisor later asks what it touched. ## Why it reduces lock-in This is the part that should interest anyone running a firm. When your integrations are bespoke to one vendor’s model, switching models means rebuilding everything, so in practice you never switch, and you have handed a supplier real power over you. When your integrations speak MCP, they are not tied to any single model. The connective tissue you paid to build stays yours. If a better or cheaper model arrives next year, you point your existing servers at it and carry on. In a field moving as fast as this one, the freedom to change your mind without a rebuild is worth real money. We think this matters enough to treat as a design principle. We build integrations to open standards wherever we can, so the value sits in your systems and your processes rather than in a dependency you cannot escape. That is not idealism. It is the same instinct any sensible firm applies to outsourcing: do not let one supplier hold the only key. ## What it is not MCP is not artificial intelligence, and it is not a product you buy. It is a protocol, a set of rules for how software talks to software, in the same family as the standards that let any browser open any website. It does not make a model cleverer. It makes a model connectable. And it does not remove the need for governance. A standard socket makes it easier to plug AI into sensitive systems, which means the questions of who approved this, what data does it touch, and who is watching it become more pressing, not less. Convenience and control have to grow together. ## Why a non-technical leader should care You will not write a line of MCP, and you should not need to. But the decisions made around it shape two things you do own: how much a future model switch costs, and how much of your AI value stays inside your firm rather than trapped in a vendor. When a supplier proposes an AI build, one good question is whether the integrations are proprietary or standards-based. The honest ones will welcome the question. The answer tells you whether you are buying an asset or renting a cage. We build with tools such as Claude precisely because they take this open approach seriously, and because it keeps your options open as the field moves. If you want a build that connects to your systems without chaining you to one vendor, that is exactly how we work. [Book a review](/book-a-review) and we will map what a standards-based integration would look like for your firm, and what it saves you the day you decide to switch. --- # AI for law firms: document review and drafting, done safely Published 15 April 2026 by Spark · Sector playbooks · 5 min read Canonical: https://hellospark.co.uk/writing/ai-for-law-firms-document-review-drafting > AI drafts fast and reviews faster, which is exactly why it needs guardrails in a law firm. Contract review, first drafts and matter intake all speed up. Privilege, confidentiality and a lawyer’s verification are what keep it safe. AI is genuinely good at the two things that eat a fee earner’s day: reading long documents and producing a competent first draft. A contract review that took two hours becomes a 20-minute check of a pre-marked document. A first draft that started from a blank page starts from 80 percent done. The gain is real and large. So is the risk, which is why the interesting question for a law firm is not whether AI helps but how you use it without breaking privilege, confidentiality or your professional duty. The answer is verification and a lawyer’s sign-off on everything that leaves the building. Three workflows carry most of the value. Here is each, and the guardrail that makes it safe. ## Contract review, marked up not decided Point AI at a contract and it will read every clause faster than any human, flag deviations from your standard positions, spot missing provisions, and surface the unusual terms buried on page 40 that a tired reviewer skims past. Against a playbook of your firm’s preferred positions it becomes a first-pass reviewer that never gets bored. The output is a marked-up document with issues ranked and explained, not a verdict. The lawyer reviews the flags, agrees or overrules each one, and owns the advice. Used this way, AI catches more and misses less, while the judgement stays exactly where professional responsibility requires it. ## First drafts from your own precedents The biggest time saving is drafting from a blank page, and the safest way to do it is to ground the model in your own precedent bank rather than the open internet. That is what [retrieval-augmented generation](/writing/rag-explained) does: the AI drafts using your firm’s approved templates and prior matters as its source, so an engagement letter, a standard agreement or a first-cut advice note comes back in your house style, citing your positions, not a hallucinated clause from nowhere. The fee earner edits and finalises. This matters twice over. The draft is better because it is grounded in your work, and it is safer because the model is reasoning over documents you control instead of inventing law. ## Matter intake and conflicts, assembled for the partner New matter intake is mostly structured admin: capture the client and matter details, read the documents that arrive, run the conflict and KYC checks, and open the file. AI can read an incoming instruction, extract the parties and the key dates, draft the matter-opening record and assemble the conflict-check inputs so nothing is keyed by hand. The partner still clears conflicts and accepts the matter, but the assembly that used to delay a file opening by a day happens in minutes. It is the same [document extraction](/writing/ai-document-extraction) pattern the rest of professional services runs on, pointed at instructions and engagement documents. ## Privilege and confidentiality are a design decision This is where law firms are right to be cautious, and where the choice of tooling matters more than anywhere else. Client documents cannot be used to train someone else’s model or leak across a shared service. So the questions to settle before anything is switched on are concrete: where does the data go, who can see it, is it used for training, and where is it stored. The answer should be an enterprise arrangement with no training on your data, tenant isolation, and hosting that satisfies your confidentiality and [data protection](/writing/ai-data-protection-uk-gdpr-jersey) obligations. Privilege is not preserved by a policy paragraph. It is preserved by the architecture, decided before the first document goes near a model. ## Verification, because the model is confident when it is wrong A language model will state a wrong case name or a plausible but non-existent clause with total confidence. That is not a bug you can prompt away, it is how the technology works, and in a legal setting it is the single most important thing to design around. Every citation, every quoted authority, every figure that leaves the firm gets checked by a person against the source. We wrote about why [human-in-the-loop](/writing/human-in-the-loop-ai) oversight has to be genuine rather than a rubber stamp, and legal work is the sharpest example. If the lawyer cannot realistically verify the output in the time they have, you do not have oversight, you have exposure. Build the checkpoint so verification is fast and unavoidable, not optional and skipped under deadline. ## The audit trail protects you and the client Every AI-assisted piece of work leaves a record: what the model was given, what it produced, what the lawyer changed, and who approved the final version, with the prompt and model version logged. For a firm that may one day need to show how a document was produced, or defend the reasonableness of its process, that trail is protection. It is also nearly free if built in from the start and a nightmare to reconstruct later, the same lesson every regulated sector learns the hard way. ## Start with the workflow that hurts most Do not roll AI across the whole firm at once. Pick one workflow, usually contract review against a playbook or drafting from precedents, and build it properly with the confidentiality, verification and audit trail in place. It is worth mapping the work first so you start where the hours actually are. Prove the pattern once, then extend. If document review and drafting are the bottleneck in your practice, a 30-minute [review](/book-a-review) will get you a candid read on where to start and what it is worth. No deck, no hype. Just the workflow, the guardrails and a number. --- # Human-in-the-loop AI: what it means and when a regulator expects it Published 8 April 2026 by Spark · Governance and compliance · 4 min read Canonical: https://hellospark.co.uk/writing/human-in-the-loop-ai > Human-in-the-loop is not a person clicking approve. It is a person who can realistically catch an error in the time they are given. Here is how to tell real oversight from theatre, and how to match it to the risk. Human-in-the-loop is not a person clicking approve. It is a person who can realistically catch an error in the time they are given, with the information they need, and the authority to stop what happens next. If any of those three is missing, you do not have oversight. You have theatre. And a regulator can tell the difference in about ten minutes. The phrase gets used as a reassurance. ‘Don’t worry, there’s a human in the loop.’ But the loop only counts if the human can act. A reviewer who sees forty AI outputs an hour, with no context and no realistic power to reject, is a rubber stamp with a pulse. The [JFSC’s guidance](/writing/ai-governance-jersey-jfsc) is explicit that oversight must be meaningful, and meaningful is a high bar dressed in a plain word. ## Real oversight versus theatre Three things separate the two, and all three have to hold. - **Time.** The reviewer has enough of it to actually assess the output, not just glance at it. If the volume forces a two-second look, the review is decorative. - **Information.** They can see what the AI saw and how it reached the output. A person cannot sensibly approve a decision whose basis is hidden from them. - **Authority.** They can say no, and no sticks. If overriding the system is career-costly or practically impossible, the human is not in control of anything. Theatre is easy to spot once you look for it. The approval rate is 99 percent. Nobody remembers the last time a human rejected an output. The reviewer cannot explain how the system reached its answer. Each of those is a sign the loop exists on the org chart and nowhere else. ## Calibrate oversight to risk Not every AI use needs a human standing over it. Uniform oversight is as wrong as no oversight, because it wastes scrutiny on trivia and starves the cases that matter. We use four modes, chosen by how much damage a wrong output could do. 1. **Approve.** The AI proposes, a human signs off before anything happens. Reserved for high-stakes cases: onboarding decisions, client-facing letters, anything touching a payment or a regulatory return. 2. **Assist.** The AI drafts, a human edits and owns the result. Good for medium-stakes work where a person is already in the flow, like drafting a file note or summarising a document. 3. **Exception.** The AI runs on its own and only escalates the uncertain or unusual cases to a person. Efficient for high-volume, low-stakes work where most outputs are routine. 4. **Kill switch.** The AI runs autonomously, but someone monitors it and can stop it instantly. For low-stakes automation where speed matters and the blast radius of a failure is small and reversible. The art is putting each workflow in the right mode. Most firms find the mistake runs one way: they apply full approve-mode oversight to everything, the reviewers drown, and the scrutiny quietly degrades into theatre on the cases that actually needed it. Match the mode to the harm and the oversight stays real where it counts. ## When a regulator expects it A person is expected in the loop whenever an AI output can reach a client, move money, or feed a regulatory return without a human between the model and the consequence. That is the general shape of it across principles-based regimes like Jersey’s and the more prescriptive tiers in the EU. The higher the potential harm, the closer and more genuine the oversight a supervisor will expect to see. The test a regulator applies is not whether a human was technically present. It is whether that human could and did exercise real judgement. So the burden is on you to show it: an approver by name, a record of what they saw, evidence that rejections actually happen. Without that, ‘there was a human in the loop’ is a claim you cannot back. ## Oversight and audit are one control Human oversight only holds up if it leaves a record. The approver’s name, what they reviewed, whether they changed anything, and when. That is where oversight meets the [AI audit trail](/writing/building-an-ai-audit-trail): the log is what turns ‘we had a human review it’ from a claim into evidence. Build the two together, because separately they are each half a control. In our work, getting oversight right on a single high-stakes workflow is a couple of weeks, and it is the part of the work that keeps a firm defensible as it scales. Prove the pattern once, then reuse the same four modes across every system you build. If you are not sure whether your current oversight is real or theatre, [book a review](/book-a-review). We will look at one live workflow, tell you which mode it should be in, and show you what a supervisor would actually want to see. No deck, no hype. --- # Why most AI pilots never reach production, and how to avoid it Published 31 March 2026 by Spark · Adoption and strategy · 4 min read Canonical: https://hellospark.co.uk/writing/why-ai-pilots-fail > A working demo is not a working system. Most AI pilots stall in the same four places: no owner, no integration, no audit trail, and the wrong first use case. Here is how to close the gap between the demo and the day job. Most AI pilots fail for reasons that have nothing to do with the AI. The model works. The demo impresses. And then the thing never reaches the day job, because a demo is not a system and nobody planned for the difference. In our experience the same four gaps kill pilots again and again, and all four are avoidable if you name them before you start. The trap is that a pilot is designed to prove the technology, when the real question was always whether the firm can run it. Those are different tests. Passing the first tells you almost nothing about the second, which is why so many impressive proofs of concept quietly die on a shelf. ## 1. No owner The most common cause of death. The pilot was run by an enthusiast or a vendor, and when the demo ended nobody senior owned taking it into production. No owner means no budget to finish it, no authority to change how work is done, and nobody accountable when it needs maintenance. A system without a named owner is a hobby, and hobbies do not survive a busy quarter. Put one senior person’s name against it before you build anything. ## 2. No integration The pilot ran in a sandbox, on exported data, with someone copying results back by hand. That is fine to prove a point and fatal as a way of working. If the system does not connect to the tools your team already lives in, the inbox, the document store, the practice management system, then using it is extra work, and extra work loses to the old way every time. Production means the workflow happens where the work already happens, not in a separate window someone has to remember to open. ## 3. No audit trail This is the one that stops regulated firms specifically. The pilot produced good outputs but kept no record of how: no log of the input, the model and prompt version, the output, or who approved it. In a regulated firm that is not a system you can put into production, because the first time a supervisor asks how a decision was reached, you have no answer. Retrofitting an audit trail after the fact is painful and sometimes impossible, which is why we build it in from day one. Our piece on [building an AI audit trail](/writing/building-an-ai-audit-trail) sets out exactly what to capture. ## 4. The wrong first use case Many pilots were doomed at selection. Someone picked the most ambitious, most visible, highest-stakes workflow, because that is the one that would impress. But high stakes means heavy oversight, hard integration, and no tolerance for the errors a first system will make. The pilot collapses under its own risk before it can prove anything. The fix is the opposite instinct: start with the smallest painful workflow, low or medium stakes, where a wrong output is caught before it matters. We made the full case in [where to start with AI](/writing/where-to-start-with-ai). ## The pattern underneath all four Notice what these have in common. None of them is a technology problem. They are all about ownership, plumbing, evidence, and judgement, the unglamorous operational work that a demo is specifically designed to skip. That is the pilot-to-production gap: the distance between something that works once, in a controlled setting, and something a team uses every day and a regulator can inspect. The gap is where the actual value lives, and it is the part most firms are least equipped to cross alone. ## How to avoid it Scope for production from the first day, not the demo. Before you build, name the owner, decide how it integrates, define the audit trail, and pick a first use case small enough to actually finish. Ship it into the real workflow in four to six weeks, measure it, and only then reach for the next. That is not a slower way to run a pilot. It is the difference between a pilot and a system. The other quiet killer is what happens after go-live. A system nobody tends drifts, breaks, and gets abandoned. That is exactly why we stay and run the systems we build: an owner in all but name, keeping them honest, logged, and improving. If your last pilot stalled, or you want the next one to reach production, [book a review](/book-a-review) and we will tell you which of the four gaps is about to bite. --- # The EU AI Act and UK AI rules: what professional services firms actually have to do Published 17 March 2026 by Spark · Governance and compliance · 4 min read Canonical: https://hellospark.co.uk/writing/eu-ai-act-uk-rules-professional-services > The EU AI Act sorts systems into risk tiers and regulates the high-risk ones hard. The UK and Jersey lean on principles instead. If you serve regulated clients, you may sit under both. Here is what that actually requires. Two regimes, two philosophies. The EU AI Act sorts AI systems into risk tiers and regulates the high-risk ones hard, with prescriptive obligations. The UK and Jersey lean on principles, leaving existing regulators to apply proportionate judgement. If your firm serves clients in both, you may sit under both at once. This is what that actually requires, without the alarmism. Start with a calming fact. Most professional services firms are users of AI, not builders of it. That matters, because the heaviest obligations in the EU regime fall on the people who develop and place high-risk systems on the market. As a firm using a vendor’s tool inside your own workflow, your duties are real but lighter, and mostly about governance rather than engineering. ## The EU AI Act in one paragraph The Act ranks AI by risk. A small set of uses is banned outright. A defined group is treated as high-risk and carries obligations around risk management, data quality, documentation, human oversight, and transparency. Most business tools fall into a limited-risk band where the main duty is disclosure, telling people they are dealing with AI. And a large remainder is minimal-risk with no specific obligations. The obligations phase in over time rather than landing all at once, so treat any specific date or clause as something to verify against the primary text, not as gospel from a blog. ## Who is actually caught A UK or Channel Islands firm can be pulled into the EU regime by reach, not just by location. In broad terms, you may be in scope if you provide or deploy an AI system whose output is used in the EU, or if you serve EU-based clients through an AI-driven process. Being outside the EU is not automatic exemption. The practical trigger is where your AI’s effects land, not where your office is. For most professional services work the honest answer is: you are probably a deployer of limited-risk or minimal-risk systems, with a smaller number of use cases that could edge toward high-risk if they materially affect a person’s access to a service. The work is to identify which is which, rather than to assume the whole firm is caught or that none of it is. ## The UK and Jersey contrast The UK has favoured a principles-based, pro-innovation approach, asking existing regulators to apply cross-cutting principles within their own remits rather than passing one omnibus AI law. Jersey is aligned in spirit. The [JFSC’s guidance](/writing/jfsc-ai-guidance-what-it-means) is proportionate and outcomes-focused, landing accountability with the board rather than prescribing controls line by line. The contrast is real but the destination is similar. Whether a rule tells you to keep records and maintain human oversight, or a principle expects you to, you end up building the same handful of controls. In our experience a firm that governs AI well for its own sake satisfies both regimes with one framework, rather than running two compliance projects in parallel. ## What to actually do Five steps cover most professional services firms, and none of them require a legal department. 1. **Inventory your AI.** List every system in use, what it does, whose data it touches, and where the output lands. You cannot tier what you cannot see. A first pass takes about a day. 2. **Tier by risk.** Sort each use into minimal, limited, or high-stakes by what a wrong output could reach. This is the spine of both regimes and it tells you where to spend effort. 3. **Handle disclosure.** For anything client-facing, be ready to tell people when they are interacting with AI. This is often the only hard EU obligation a limited-risk deployer carries. 4. **Build oversight and an audit trail on the high-stakes cases.** A named human who can override, and a record of every material decision. This satisfies the demanding parts of both regimes at once. 5. **Check jurisdiction on EU-touching work.** Where your AI’s output reaches EU clients or users, take a proper view on scope. This is the one place worth a specialist’s read rather than a rule of thumb. That is a four to six week piece of work for a firm of 20 to 250 people, and it doubles as the governance foundation you would want regardless of regulation. It is also why the systems we build tend to survive contact with either regime. One caveat, stated plainly. AI regulation is moving quickly, and the exact tiers, dates, and obligations shift as guidance and case law develop. Treat this as our practical read of the landscape, not as legal advice, and verify any specific point against the primary source before you rely on it. If you want a grounded view of which of your AI uses are caught and what to fix first, [book a review](/book-a-review). We will map your systems to the tiers, flag the EU-touching cases, and put a number on the work. No deck, no hype. --- # Prompt versioning: keeping AI outputs consistent and auditable Published 10 March 2026 by Spark · Technology, explained · 4 min read Canonical: https://hellospark.co.uk/writing/prompt-versioning > The instruction you give an AI is not a throwaway sentence. It is the logic of the system, and quietly editing it changes every output that follows. Treat prompts like code: version them, test them, log the changes. Here is why, and what it looks like in a firm a regulator might inspect. Treat your prompts like code, because that is what they are. The prompt is the instruction that tells the AI what to do, and it is the actual logic of the system. Change a line of it and you change every output that follows, often in ways you did not intend and cannot see until something goes wrong. Yet in most firms prompts live in a document somebody edits on a whim, with no record of what changed, when, or why. That is a control gap, and in regulated work it is the kind a supervisor finds quickly. Prompt versioning closes it. ## Why a prompt is not just a sentence It is tempting to see the prompt as a casual request, the way you would ask a colleague. It is not. In a live system the prompt encodes the rules: what to extract, what tone to take, what to refuse, what to flag for a human. It is the difference between an output you can defend and one you cannot. A small wording change, adding an example, tightening an instruction, telling the model to be more concise, can shift results across thousands of documents. If that shift is invisible and unrecorded, you have no way of knowing why last month’s outputs and this month’s no longer match. ## Consistency is the first casualty Language models do not return the same answer to the same question every time. There is natural variation built in. The prompt is your main lever for pinning behaviour down: a precise, well-tested instruction produces far more consistent results than a loose one. So when someone quietly tweaks the prompt to fix one annoying output, they may be loosening the very control that kept the other thousand consistent. Without versioning you cannot even tell that the tweak happened, let alone measure what it did. The outputs drift, someone eventually notices, and nobody can explain when it started. ## What versioning actually means It is exactly what software teams have done with code for decades, applied to prompts. Every version of a prompt is saved, numbered, and dated. You can see what the prompt said last Tuesday, who changed it, and why. You can compare two versions side by side. You can roll back to the previous one in seconds if a change makes things worse. None of this is exotic technology. It is the discipline of never editing the live instruction in place and never losing the history. - **A stored history.** Every version kept, numbered and dated, never overwritten. - **A reason attached.** A one-line note on why each change was made, so the record explains itself. - **A tie to outputs.** Each output records which prompt version produced it, so you can trace any result back to the exact instruction behind it. - **A way back.** The ability to roll back to a known-good version the moment a change misbehaves. ## Test before you ship The other half of treating prompts like code is testing them like code. Keep a small set of representative examples with the outputs you expect, and run any new prompt version against them before it goes live. Did the change you wanted happen? Did it break anything you were not looking at? This is how you catch the classic failure, where fixing one edge case quietly ruins the ninety-nine ordinary ones. It takes minutes to set up and it turns prompt changes from a gamble into a controlled step. A prompt change that has not been tested against known examples is a change made with your eyes shut. ## Where this meets the regulator A supervisor asking about an AI system will eventually ask the same question in different words: how do you know it does what you say it does, and can you prove what it did on a given day. Prompt versioning is a large part of the answer. It lets you show the exact instruction in force when a particular output was produced, and it lets you demonstrate that changes are controlled rather than made on a whim. That sits right alongside the wider record of inputs, outputs and approvals we set out in [building an AI audit trail](/writing/building-an-ai-audit-trail). Versioned prompts are one of the load-bearing entries in that trail. ## How we do it Every system we build treats the prompt as a versioned, tested, logged component from the first day, not something bolted on when someone asks. The cost of doing it early is close to nothing. The cost of retrofitting it after outputs have drifted and nobody can say why is real, and sometimes the history is simply gone. This is the same principle we apply across our work: build the governance in while it is cheap, so the system survives contact with an auditor rather than collapsing under the first hard question. If you are running AI in production and cannot currently say which prompt produced last month’s outputs, that is worth fixing before someone asks. [Book a review](/book-a-review) and we will look at it with you. When we run a system for you, we keep its prompts, tests and change logs maintained as it evolves. The discipline is dull. The absence of it is expensive. --- # AI for accountancy practices: from bookkeeping to advisory Published 4 March 2026 by Spark · Sector playbooks · 4 min read Canonical: https://hellospark.co.uk/writing/ai-for-accountancy-practices > The lowest-margin work in an accountancy practice is the most automatable: data entry, categorisation, reconciliations, first-draft accounts. Automate it and you free your best people for the advisory work clients will happily pay for. The lowest-margin work in your practice is the most automatable, and that is the whole opportunity. Data entry, transaction categorisation, bank reconciliations, chasing records, first-draft accounts: this is high-volume, low-judgement work that AI and automation handle well, and every hour of it you reclaim is an hour your qualified people can spend on advisory work clients actually value. The point of AI in an accountancy practice is not to do accounting. It is to move your best people up the value chain, from compliance grind to the advice that carries a real fee. In our experience a practice can take 40 to 60 percent of the manual effort out of routine compliance work within a couple of cycles. Here is where the time goes, and where automation gets it back. ## Data entry and categorisation, read not typed The single biggest drain in most practices is turning a client’s shoebox of documents into clean data. Invoices, receipts, bank statements and expense claims arrive as PDFs, photos and spreadsheets, and someone keys them in. AI reads them instead, extracting amounts, dates, suppliers and VAT, and proposes a category based on the description and the client’s history. This is the [document extraction](/writing/ai-document-extraction) engine at the heart of the modern back office. The categorisation learns your treatment of recurring items, so the coding gets more accurate over time and your team confirms rather than types. The tedious part shrinks; the review part stays. ## Reconciliations that surface only the breaks Bank and ledger reconciliation is rule-based work that automation does faster and more reliably than a person working late. The system matches transactions, applies your tolerances, and presents only the exceptions: the unmatched items, the duplicates, the entries that do not tie out, each with a plausible explanation drafted for review. Instead of scrolling a full ledger to find the handful of problems, your team opens a short queue of genuine breaks. The 90 percent that reconciles cleanly runs itself, and the human attention goes to the items that actually need a decision. ## Drafting accounts, letters and queries Once the numbers are in, AI drafts from them. First-draft accounts in your house format, client query letters, the list of missing records, a plain-English summary of the year’s numbers for the client meeting: all built from data you already hold, ready for a qualified person to review and finalise. Drafting from a blank page is where hours quietly disappear, and starting from a competent 80 percent draft changes the economics of the compliance job. The accountant edits and signs. The machine does the typing, not the thinking. ## The human check stays, because the numbers matter Nothing here files a return or issues accounts on its own. Every categorisation, reconciliation and draft is reviewed and approved by a person before it counts. This is not timidity, it is proportion. A miscoded transaction or a wrong figure in filed accounts reaches a client and potentially a tax authority, which puts it in the tier where genuine [human oversight](/writing/human-in-the-loop-ai) is not optional. The automation removes the keystrokes and the assembly; the qualified judgement and the sign-off stay exactly where your professional responsibility requires them. That is what lets you go faster without going careless. ## The advisory dividend is the real return Here is the part that changes the shape of a practice. When compliance work takes half the time, you are not just cutting cost, you are freeing capacity at the top of the pyramid. The senior people who were buried in review can spend that time on cash-flow advice, tax planning, management reporting and the forward-looking conversations clients happily pay a premium for. Advisory is where the margin and the client loyalty live, and the constraint has always been time. Automating the compliance base is how you buy that time back. Measure it properly: track the chargeable advisory hours you free up, not just the compliance hours you save, because that is where the return actually shows up. We set out how in [how to measure AI ROI](/writing/how-to-measure-ai-roi). ## An audit trail, because you are regulated too Every automated step leaves a record: what was extracted, how it was categorised, what was flagged, what a person changed, and who approved the output, with the model and prompt version logged. For a practice answerable to a professional body and to HMRC, that trail is both protection and good discipline. Built in from the start it is nearly free; retrofitted later it is a headache. It is also what lets you adopt AI with confidence rather than crossing your fingers. ## Where to start Start with the workflow that consumes the most junior time, usually data entry and categorisation, and build it properly with the review step and the audit trail in place. It is worth mapping the practice first so you automate in the order that returns the most soonest. Prove the pattern on one client segment, then roll it across the book. If compliance work is crowding out advisory in your practice, a 30-minute [review](/book-a-review) will get you a straight read on which task to automate first and what it frees up. No deck. Just the hours and the number. --- # AI and data protection under UK GDPR and the Data Protection (Jersey) Law 2018 Published 24 February 2026 by Spark · Governance and compliance · 5 min read Canonical: https://hellospark.co.uk/writing/ai-data-protection-uk-gdpr-jersey > Feeding client data to an AI tool is still processing personal data, and the same rules apply. Lawful basis, minimisation, and knowing where your prompts go. Here is what UK GDPR and the Jersey Law actually require. Feeding client data into an AI tool is processing personal data, and every rule you already follow still applies. UK GDPR and the Data Protection (Jersey) Law 2018 do not have an AI exemption. If anything, AI raises the stakes, because you are often sending personal data to a third party, sometimes offshore, and getting back an output whose basis you may struggle to explain. The obligations are familiar. The failure modes are new. The good news is that the two regimes are closely aligned. The Jersey Law tracks the GDPR structure deliberately, so a control built for one generally satisfies the other. You are not running two data protection projects. You are running one, applied carefully to a new kind of processing. ## Lawful basis comes first Before you put personal data through an AI system, you need a lawful basis for that processing, and it has to cover the AI use specifically. Consent gathered for onboarding does not automatically extend to feeding the same data to a model for a purpose the client never saw. Where you rely on legitimate interests, do the balancing exercise properly and write it down. If the data is special category, the bar is higher again. The question to answer, in one line, is: on what basis are we allowed to process this person’s data through this tool, for this purpose? If the honest answer is ‘we hadn’t thought about it that way’, stop and think about it before the data leaves your building. ## Data minimisation, applied to prompts The habit that gets firms into trouble is oversharing with the model. It is tempting to paste an entire client file into a prompt because it is easy. Minimisation says send only what the task needs. If a summary requires three fields, do not send thirty. If a name is irrelevant to the analysis, redact it. Every extra field in a prompt is personal data you have exported, often to a system you do not control. In practice this is a design decision, not a discipline you can rely on staff to remember under time pressure. Build the workflow so it only ever passes the minimum, and minimisation stops being a rule people break and becomes a property of the system. ## Where do your prompts and outputs actually go This is the question most firms cannot answer, and the one a regulator will ask. When someone types into an AI tool, where does that text travel, where is it stored, for how long, and is it used to train the vendor’s models? The answers vary enormously between a consumer chatbot and an enterprise deployment with a data processing agreement and training switched off. - **Residency.** Where is the data physically processed and stored? Transfers out of the UK or Jersey need a lawful transfer mechanism, and ‘the server is somewhere in the US’ is not one. - **Retention.** How long does the vendor keep prompts and outputs, and can you delete them? Indefinite retention of client data by a third party is a problem you own. - **Training.** Is your data used to improve the vendor’s models? For client data, the answer needs to be no, and you need it in writing. - **Sub-processors.** Who else touches the data downstream? Your obligations follow the data even when a vendor hands it on. ## Vendor and contract questions Where an AI vendor processes personal data on your behalf, they are your processor, and you need the same contractual controls you would demand of any material outsourcing. A data processing agreement, clear purpose limits, security commitments, breach notification, and the right to audit. The convenience of signing up with a corporate card does not remove the obligation to have the paperwork a supervisor expects. Treat consumer-grade AI tools with particular care. The free tier that trains on your inputs is not somewhere client data should go, however useful it feels. Pick tools that offer an enterprise footing with the controls in writing, and reserve the consumer versions for non-personal, non-confidential work. ## DPIAs for high-risk use Where an AI use is likely to result in a high risk to people’s rights, a Data Protection Impact Assessment is not optional, it is expected. That covers profiling, automated decisions with legal or similarly significant effects, and large-scale processing of sensitive data. The DPIA is where you set out what the system does, what could go wrong, and the controls that bring the risk down to acceptable. Done well, a DPIA is not a hoop. It is the same thinking as a good AI risk assessment, written in the language a data protection regulator reads. Much of the evidence it needs, what the system does, who oversees it, what is logged, comes straight from your [AI audit trail](/writing/building-an-ai-audit-trail). Build the trail and the DPIA half writes itself. Getting this right across a firm’s AI use is typically a four to six week piece of work. One caveat, stated plainly: this is our practical read, not legal advice, and you should check specific points against the primary legislation and your regulator’s guidance before you rely on them. If you want a candid view of where your client data actually goes when your team uses AI, [book a review](/book-a-review). We will trace one live workflow end to end, flag the data protection gaps, and put a number on the fix. No deck, no hype. --- # An AI readiness checklist for firms of 5 to 250 people Published 10 February 2026 by Spark · Adoption and strategy · 4 min read Canonical: https://hellospark.co.uk/writing/ai-readiness-checklist > You do not need to be AI-ready in the abstract. You need to be ready for one workflow. This honest checklist covers the four things that actually matter: data, governance, skills, and a use case worth building. You do not need to be ready for AI in general. You need to be ready for one specific workflow, and that is a far smaller test. The firms that stall are usually waiting to be ready for everything. The firms that move check four things against one use case, find they are more ready than they feared, and start. This is that checklist, written to be honest rather than reassuring. Treat it as a diagnostic, not a gate. You do not need every box ticked before you begin. You need to know which boxes are empty, because the empty ones tell you what to fix first and what to keep a human in front of while you fix it. ## Data: can the system get to what it needs AI is only as good as what you can feed it. You do not need a data lake. You need the specific inputs for your first workflow to be reachable and reasonably consistent. - The documents or records the workflow needs live somewhere a system can reach, not only in someone’s head or a paper file. - The inputs are consistent enough to work with. Messy is fine. Genuinely unpredictable, with no pattern, is a problem to solve before you automate. - You know where that data may and may not go, especially anything that would leave the island or a controlled environment. Our note on [data protection under UK GDPR and Jersey law](/writing/ai-data-protection-uk-gdpr-jersey) covers the lines that matter. - You can get to the data without breaching a contract or a client confidentiality term. Check before you connect, not after. ## Governance: can you show your working For a regulated firm this is the section that decides whether a pilot ever becomes a system. None of it is heavy. All of it has to exist. - One senior person is willing to put their name against AI adoption, with the authority to stop a system going live. - You have, or can build in a day, a simple register of what AI is already in use across the firm. Most firms find more than they expected once they look. - You can sort a use case into low, medium, or high stakes by what a wrong output could reach, and match the oversight to the tier. - You accept that every material AI decision needs a record: the input, the model and prompt version, the output, and who approved it. Our [AI governance framework](/writing/ai-governance-jersey-jfsc) sets out the whole spine, and it is lighter than you think. ## Skills: enough to govern, not to build You do not need engineers on staff. You need enough understanding to make good decisions and keep an honest hand on the wheel. - At least one person accountable actually understands what the technology can and cannot do. ‘The vendor said it was fine’ is not a control. - The people who will use the workflow are willing to try it and honest enough to say when it is wrong. - Someone will own the system after go-live, because a system nobody tends drifts and breaks. If that owner is not on your staff, we can be that hand for you. - Your team knows the difference between a system that assists a human and one that decides on its own, and knows which one they are being handed. ## One good use case: the box that matters most You can have every other box ticked and go nowhere without this one. A good first use case is frequent, rule-bound, and forgiving of the occasional error because a human checks it before it counts. It is small enough to finish in weeks and painful enough that people will actually use it. If you cannot name one, that is your first piece of work, and it is a conversation, not a project. - It happens weekly or more often, so small savings add up to a real number. - You can describe the logic in a couple of sentences a new joiner could follow. - A wrong output is caught by a human before it reaches a client, a payment, or a regulatory return. - You can put a pound figure on what it costs you today. If you cannot, it is not your first case yet. ## Reading your score If most of the data and use-case boxes are ticked and the governance ones are thin, you are ready to start and should build the governance alongside the first workflow, not before it. If the data is not reachable or you cannot name a use case, fix that first, and it is usually a fortnight of work, not a transformation programme. Perfect readiness is a myth. Enough readiness for one workflow is a Tuesday. If you want the checklist run properly against your firm, [book a review](/book-a-review) and we will tell you honestly which boxes are empty and which one to fill first. --- # Claude vs GPT vs Gemini for professional services work Published 28 January 2026 by Spark · Technology, explained · 5 min read Canonical: https://hellospark.co.uk/writing/claude-vs-gpt-vs-gemini-professional-services > The honest answer is that all three are good, and the differences that matter to a law or trust firm are not the ones the benchmarks argue about. Here is how we compare Claude, GPT and Gemini on the things a regulated firm actually cares about, and where we land. All three are good. That is the honest starting point, and it is worth saying plainly because the internet would have you believe one of them is about to win forever. Claude, GPT and Gemini are each capable of doing serious professional work, and any of them will impress you in a demo. The differences that matter to a law firm, a trust company or a fund administrator are not the ones the benchmark leaderboards fight about. They are about how the model handles a hundred-page document, how your data is treated, and what it plugs into. Here is how we compare them, and where we tend to land for regulated work. One caveat first. These models change every few months, and specific capabilities move with each release. Treat the shape of this as durable and the fine detail as a snapshot. What follows is our working view, not a lab benchmark. ## Reasoning and careful work For the kind of work that carries risk, reading a clause carefully, following an instruction exactly, refusing to guess when it does not know, we favour Claude, and the latest Claude 5 family and Opus 4.8 have reinforced that preference. It tends to be the most willing to say ‘that is not in the document’ rather than inventing a plausible answer, which is precisely the behaviour you want when a wrong output could reach a client or a regulator. GPT is a very strong all-rounder and often the most fluent writer of the three. Gemini reasons well and is improving quickly. The gaps here are narrow and they move. But for cautious, instruction-following work in a regulated setting, Claude’s temperament fits the job. ## Long documents This is where real work lives, because professional services runs on long documents: trust deeds, prospectuses, loan agreements, bundles of correspondence. All three can now take a large amount of text in a single go, which was not true a couple of years ago. Gemini has pushed hard on very large context and is a strong choice when you genuinely need to reason across an enormous volume at once. Claude is reliable and precise across long documents and, in our experience, holds its accuracy well deep into a lengthy input. GPT is capable here too. The practical point is that raw capacity is no longer the constraint it was. What separates them now is accuracy at length, and that is worth testing on your own documents rather than trusting a headline number. ## Data handling and trust For a regulated firm this outranks cleverness, and it is where the decision often actually turns. The questions are the same for all three providers. Is your data used to train the model. Where is it processed. What are the contractual and data protection terms on the business and enterprise tiers. The consumer versions of these tools and the business versions are not the same on any of these points, and the terms genuinely differ between providers and change over time. Anthropic has built much of Claude’s public positioning around safety and careful data handling, which lands well with cautious firms, but you should verify current terms for whichever you choose rather than take any reputation on trust. We cover the underlying obligations in [AI, data protection and UK GDPR in Jersey](/writing/ai-data-protection-uk-gdpr-jersey). The short version: read the terms of the specific tier you are buying, and do not assume the free tier’s rules apply to the paid one, or the other way round. ## What it connects to can decide it Sometimes the deciding factor is not the model at all but what it connects to. GPT sits inside the OpenAI and Microsoft world, so if your firm runs on Microsoft 365 there is a gravitational pull towards it. Gemini is woven through Google Workspace, which matters if that is where your firm lives. Claude is strong on integration through open standards and is a common choice for custom builds where you want to avoid being tied to one vendor’s stack. None of this decides the question on its own, but it is a real cost or saving depending on where your firm already is. ## Where we land For the regulated, document-heavy, careful work that is our bread and butter, Claude is usually our default, because its instinct to stay grounded and admit uncertainty suits work where a confident wrong answer is the worst outcome. That is a preference, not a rule. We reach for the others without hesitation when the job calls for it. - **Claude** for careful reasoning, long regulated documents, and standards-based custom builds where staying grounded matters most. - **GPT** for fluent writing, broad general capability, and firms already living in Microsoft 365. - **Gemini** for very large volumes of text at once and firms built on Google Workspace. The mistake we see most often is choosing a model in the abstract, from a review or a demo, before knowing the actual task. The right way round is to define the workflow first, then pick the model that fits it, and then test it on your own real documents before committing. The differences that show up on your material are the only ones that matter. If you want help matching a model to a real workflow rather than a benchmark, [book a review](/book-a-review) and we will give you a candid steer for your specific work. The model matters less than getting the workflow and the governance right, which is the part we care about most. --- # Document extraction: turning PDFs, contracts and statements into structured data Published 20 January 2026 by Spark · Sector playbooks · 5 min read Canonical: https://hellospark.co.uk/writing/ai-document-extraction > Document extraction is the unglamorous workhorse behind almost every useful AI system in a professional firm. It turns the PDFs, contracts and statements clogging your inbox into structured data. Here is how it works and where the value actually is. If you automate one thing in a professional firm, automate document extraction. It is the unglamorous workhorse sitting underneath almost every AI system worth building: the step that turns the PDFs, scanned contracts, bank statements and forms clogging your inbox into clean, structured data a computer can actually use. It is not exciting. It is where the money is. Get extraction right and onboarding, reconciliation, review and reporting all speed up, because every one of them starts with reading a document. Get it wrong and nothing downstream works. Typically a well-built extraction step handles 80 to 90 percent of documents with no human touch, and routes the rest to a person. ## What it actually does Extraction takes an unstructured document, a thing built for a human to read, and pulls out the specific data points you care about into fields a system can store. A passport becomes a name, a number and an expiry date. A bank statement becomes a list of transactions with dates and amounts. A contract becomes parties, dates, values and key clauses. Old approaches used rigid templates and broke the moment a layout changed. Modern AI reads the document more like a person does, understanding that the total is the total whether it sits top-right or bottom-left, which is why it copes with the messy variety of real-world paperwork that template systems never could. ## How it works, without the jargon There are three steps. First, if the document is a scan or a photo, the text is recognised and lifted off the page. Second, an AI model reads that text, and increasingly the layout and images too, and identifies the fields you asked for. Third, and this is the part firms skip at their peril, each extracted value comes back with a confidence signal. The model tells you not just what it read but how sure it is. That confidence score is the hinge the whole system turns on, because it lets you send the easy 85 percent straight through and hold back the 15 percent that needs a human eye. ## Accuracy, honestly Extraction is very good and it is not perfect, and anyone who tells you otherwise is selling something. A clean, typed document extracts near-flawlessly. A crumpled receipt photographed at an angle in poor light does not. The mistake is chasing 100 percent accuracy from the model itself, which is neither achievable nor necessary. The workable target is high accuracy on clear documents plus honest confidence scoring on the rest, so the system knows what it does not know. A model that is 90 percent accurate and always flags its uncertainty is far more useful than one that is 95 percent accurate and confidently wrong on the other 5 with no warning. You design for the errors you can catch, not the accuracy you wish you had. ## Why human review is a feature, not a failure The confidence score turns human review from a bottleneck into a design choice. High-confidence fields flow straight through. Low-confidence fields land in a short review queue where a person confirms or corrects them in seconds, and every correction is a data point that improves the setup over time. This is how you get speed and reliability at once, and in a regulated setting it is also how you stay compliant, because [genuine human oversight](/writing/human-in-the-loop-ai) is exactly what the JFSC and every professional body expect. The human is not there because the machine failed. The human is there by design, spending their attention only on the documents that actually need it. ## Extraction versus RAG It helps to know what extraction is not. Extraction pulls specific fields out of a document into structured data. Retrieval, the technique behind [RAG](/writing/rag-explained), does something different: it finds and quotes the relevant passage from a large body of documents to answer a question. You use extraction when you know exactly which fields you want, every time, in the same shape, which is the case for onboarding forms, statements and invoices. You use retrieval when the question varies and you need the system to go find the right paragraph. Most real systems use both, extraction to structure the incoming documents and retrieval to answer questions across them, and knowing which job you have is half of designing it well. ## Where the value actually sits The same extraction engine, tuned to different documents, powers the highest-value workflows across every sector we serve. In trust and corporate services it reads KYC documents for client take-on. In fund administration it turns custodian statements and price files into a validated NAV. In law it lifts the parties and dates out of an instruction. In accountancy it reads the shoebox of invoices and receipts. Build it once, properly, with the confidence scoring and the review queue working, and you have the foundation every other automation stands on. That is why it is the first thing we build and the piece we take most care over. ## What it takes to build A production extraction system tuned to your document types, with the confidence thresholds, the review queue and the audit trail in place, earns its keep faster than almost anything else because everything downstream depends on it. The work that matters is not the model, it is the tuning: testing against your own real documents, setting the confidence thresholds where they belong, and designing the review step so it is fast and reliable rather than an afterthought. If your team is still keying data off PDFs by hand, that is the workflow to fix first. A 30-minute [review](/book-a-review) will get you a candid read on which documents to start with and what automating them is worth in hours per week. No deck, no hype. Just the workhorse, done properly.