Specialism
AI document processing for back office teams
Invoices, forms, statements and certificates arriving in formats you do not control, read by a person and retyped into a system. It is the most common manual process in business, and one of the most automatable.
Why now
What changed since the template-based era
Many teams tried document capture years ago, found it brittle, and concluded the category does not work. The underlying approach has genuinely changed.
Traditional optical character recognition reads characters at coordinates. It needs a template per layout, and it breaks the moment a supplier redesigns their invoice or a new customer sends a form built differently. Maintaining templates for hundreds of counterparties is a job nobody wants, which is why most of those projects were quietly abandoned.
A model reads a document closer to the way a person does. It can find the invoice total wherever it sits, on a layout it has never encountered, because it understands what a total is rather than where one was last time. The variety that killed template systems is exactly what this approach handles well.
That does not make it infallible, and the honest framing matters: extraction is probabilistic. A printed total is read near-perfectly; a handwritten annotation is not. So the engineering question is never accuracy in the abstract — it is what the system does about the cases it is unsure of.
| Stage | What happens | Deterministic or AI |
|---|---|---|
| Intake | Email, scanner, portal or folder monitored continuously | Deterministic |
| Classification | What kind of document is this, and whose is it | AI |
| Extraction | Pull the fields the downstream system needs | AI |
| Validation | Formats, ranges, arithmetic, reference-data matching | Deterministic — must be exact |
| Confidence routing | Below threshold or failing validation goes to review | Deterministic |
| Human review | Confirm or correct a pre-filled record beside the original | People |
| Write-through | Post to the finance, practice or operations system | Deterministic |
Validation is the part that makes it safe. A model should never be the last check before a number lands in your ledger. Line items must sum to the stated total, dates must be plausible, supplier identifiers must match something real. These checks cost almost nothing and catch the failures that would otherwise be expensive and silent.
Expectations
Do not aim for zero human review
It is the most common scoping mistake, and it usually destroys the business case. Most of the saving arrives early: turning a reviewer's job from reading a document and typing into another screen, into confirming a record that is already filled in with the source alongside. That is the bulk of the time, and it arrives with modest accuracy.
Squeezing out the last few per cent of review costs disproportionately and removes the check that keeps bad data out of your systems. A sensible target is a high straight-through rate on clean documents with a review queue that stays small — not an empty one.
Sector-specific versions of this work appear on the insurance agency, accounting firm and healthcare practice pages, the charge documents and remittance files described under AI automation for medical billing companies, the order packets and recorded instruments handled under AI automation for title companies, the grant and funder reporting case under AI automation for nonprofits, the certificate, permit and authorisation chasing under AI automation for funeral homes, and the logistics case is covered under Long Beach. For how this gets wired into surrounding systems, see workflow platform consulting. Where a security review is required we document against the NIST AI Risk Management Framework.
Questions
What teams ask about extraction
How accurate is extraction in practice?
Accurate enough to be worth it, and anyone quoting you a single percentage without seeing your documents is guessing. Accuracy varies by field and by document quality — a printed total is near-perfect, a handwritten note is not. That is why the design question is never accuracy alone but what happens to the cases the system is unsure about.
Is this different from the OCR we already tried?
Materially, yes. Traditional OCR reads characters at coordinates, so it needs a template per layout and breaks when a supplier redesigns their invoice. A model reads the document more like a person — it can find the total wherever it sits, on a layout it has never seen. That is precisely the failure mode most template systems die of.
What happens when it gets something wrong?
It should be caught before it matters. Extracted values are validated against expected formats, ranges and reference data, and cross-checked where the document allows — line items summing to the stated total, for instance. Anything failing validation or falling below the confidence threshold goes to a review queue with the source document attached.
Do we need to reach zero human review?
No, and chasing it usually destroys the business case. The large saving comes from turning a reviewer's job from reading and typing into confirming a pre-filled record — that is most of the time back. Eliminating the last few per cent of review costs disproportionately and removes the check that keeps errors out of your systems.
Free consultation
Tell us what is costing you the most time
Describe the process. You get a straight read on whether it is worth automating, roughly what it would take, and what we would tackle first — on a call that costs nothing.
- Under an hour, free, and no obligation
- Fixed price per stage, agreed before any work starts
- You own everything we build, documented and in your accounts
- If you do not need us, we will say so on that call
The form is the fastest way to reach us — we reply from a real person, usually within one business day.