Sector playbooks
Document extraction: turning PDFs, contracts and statements into structured data
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.
By Seb Lawson · 20 January 2026 · 5 min read
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 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, 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 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.