All use cases
Document and data extraction

Anything that arrives as a document and leaves as a field

There is a category of work that consists of reading something and typing what it says into a system. It is the most common AI problem in any company, and the reason it stays unsolved is that generic tools extract to their schema rather than to yours, and none of them will tell you how often they are wrong.

How it runs today

The document arrives as a PDF, a scan or a photograph. Somebody opens it, finds the eight fields that matter, types them into the system, and files the original somewhere it will never be read again. Everything about the document that was not one of those eight fields is now lost, and the eight are only as good as the afternoon they were typed on.

What the system does
  1. 01We extract into the schema your systems already use, so the output posts where it belongs instead of into a spreadsheet somebody has to re-key.
  2. 02Every value keeps a pointer to the page and position it was read from, so a disputed figure can be traced back rather than argued about.
  3. 03The checks that a person does without thinking about them are built in: totals that have to add up, dates that have to be in order, references that have to exist.
  4. 04We build a test set out of your own documents, agree what a correct extraction looks like on it, and measure against it before anything goes live.
  5. 05It keeps being measured in production, because a document format changes and nobody sends a notice when it does.
What you get
  • Structured records in the shape your systems expect
  • Every field traceable to the place it was read from
  • The documents that failed a check, separated, with the check that failed
  • An error rate you can quote, measured on your own material rather than a benchmark
Where it stops

It does not quietly guess. A field it cannot read with confidence comes back marked rather than filled in, because a wrong value that looks like a right one is worse than a gap.

Worked example

In detail: invoice parsing

Invoices are the version of this problem most companies recognise. They arrive from hundreds of suppliers, each with a layout of its own, and somebody types them into the finance system one at a time.

What arrives

A PDF attached to an email, a scan of a paper invoice, a photograph taken on somebody's phone, or a structured file from a supplier who already sends one. The same supplier changes its layout without telling anyone, and a system that was tuned to that layout quietly starts being wrong.

What comes out

Supplier and their tax identifier, invoice number, issue and due dates, currency, net, tax and gross, the purchase order it belongs to, the payment details, and every line with its description, quantity, unit price and tax rate. Each value carries the page and the position it was read from.

What gets checked

Lines are added up against the stated totals. Tax is checked against the rate that applies. The invoice is matched to its purchase order and to the goods receipt, and the prices are compared against what was agreed. Duplicates are caught on supplier, number and amount together, which catches the same invoice arriving twice in two formats.

What happens next

Invoices that pass every check post into the finance system with their reference and their source document attached. Anything that fails goes to a person with the failure named: a total that does not add up, a supplier not on file, a line priced above the order, a duplicate of something paid last month.

The same shape works for delivery notes, insurance claims, lab reports, bank statements and contracts. What changes is the fields and the checks; the pattern of extract, verify, post and escalate does not.

Where this shows up
  • Contracts reduced to the clauses that carry an obligation or a date
  • Invoices, statements and filings into the schema your systems expect
  • Decades of archive made searchable field by field

Tell us the decision you want to fix

One call, no deck. We will tell you whether this is something we can build for you, and if it is not, we will say that too.

EU AI Act compliantGDPR compliantOur cloud and models, yours, or on-premiseBuilt in Europe