Skip to content

AI solutions

AI for documents

Documents without manual entry: invoices, contracts, statements and certificates become records in your system — fields, amounts and dates included.

Reply to your brief
within 24 hours
First call
30 minutes, no commitment
First working version
in 4 weeks

Process

What happens, and when

The work is split into stages, and each one has a named result. You see it yourself rather than reading about it in a report.

  1. Days 1–5

    Your real intake

    We take your real intake rather than samples: photos shot at an angle, scans of mixed quality, PDFs that turn out to be a picture inside. We count the document types and the fields people retype by hand today, and set part of the pile aside — quality is measured on it later, and it is never shown during tuning.

    You get — Types, fields and a pile held back

  2. Week 2

    The target record

    We describe the target record: which fields are mandatory, in what shape your system accepts them, what to do with a date written three ways and a total with and without tax. This is also where “uncertain” gets defined, and where an uncertain field goes — into a review queue, not into the database.

    You get — The record described, with uncertainty thresholds

  3. Weeks 3–4

    Field extraction

    Extraction is built in steps: the batch is split into documents, each document is typed, fields are pulled out, and every value is tagged with the place on the image it came from. We measure on the pile set aside, field by field — an overall accuracy figure hides the one field that is always wrong.

    You get — Extraction with accuracy measured field by field

  4. Before launch

    Wired to your system

    It connects to your system and review goes on: the operator sees the draft beside the scan and confirms it. Thresholds are set so that documents the pipeline will not take are handed over whole — an honest refusal costs less than a quiet error.

    You get — A review screen writing into your system

  5. The first month and beyond

    Operator corrections

    Operator corrections are the most valuable labelled data you will ever get: every one of them is folded into the pile quality is measured on, and the thresholds are revisited. A new document type, or a supplier who changed their form, shows up as a rise in refusals long before it shows up as a complaint.

    You get — A growing test set and revised thresholds

What the work covers

  • Your real intake, not samples

    an inventory of the document types and of the fields typed in by hand today, taken from the actual flow — phone photos, creased scans, PDFs with no text layer

  • The target record

    the fields without which no record is created, one format for dates, totals and tax, the counterparty list names are matched against, and the test that marks a document as a duplicate

  • Pulling out the fields

    splitting the batch into documents, typing each one, extracting values tied to their place on the scan — and settling whether a scan goes to the model whole or as de-identified text

  • The uncertainty threshold

    every field gets its own threshold, below which a value queues for an operator instead of reaching the database; a document the pipeline cannot identify goes to a person whole

  • Accuracy field by field

    measured on the part of your own intake that was set aside and never shown during tuning — one field at a time, because an average across a document says nothing about the field the whole thing was for

  • Wired to your system

    a review screen with the scan and the draft record side by side, writing through an API or an export, and operator corrections folded back into the set quality is measured on

What we need from you

  • A month or two of real documents, crooked scans and rare forms included, rather than ten good examples
  • An operator on your side for the first weeks — to confirm the disputed fields and work through the refusals
  • A decision on whether a scan goes out whole or only as de-identified text — it changes both the design and the price

What people usually ask

Contact

Send a description of the task

A reply with the scope, the timeline and a budget estimate comes within 24 hours.

The first call is 30 minutes, with no commitment on your side.

  1. 01

    You describe the task

    Five questions in the form, or a plain email — whichever suits you.

  2. 02

    We answer within a day

    With the scope, the timeline and a budget estimate, based on what you told us.

  3. 03

    We talk for 30 minutes

    To clear up whatever is unclear. It commits you to nothing.