Build 01 · Document extraction

Any document. One workflow.

A document comes in. Data gets pulled out. Where the model is confident, it moves straight into the process. Where it isn't, a person sees only the fields in question, in a review window built for that, not the raw file. This page is the full detail behind the first build on our list.

All builds

Scored per field, not per document

Field names and bar lengths are illustrative. Real thresholds are set per document type against your own tolerance for review.

KYC & onboarding

Evidence work only

Identity documents, proof of address, and screening results in a dozen formats from a dozen jurisdictions. The agent reads them, checks them against each other, and builds the file. A person clears the case.

ID and address documents read and cross-checked for name and date mismatches
Screening evidence assembled into the case file with the source of every match

Invoices & AP documents

Highest volume

Line items, tax codes and PO references pulled from any layout, not just the templates a vendor happened to standardise on. What comes out feeds the match, not a spreadsheet someone still has to check.

Header and line-item fields extracted regardless of vendor layout
Extracted data pushed straight into the P2P match, no re-keying step

Contracts

Strong fit

Renewal dates, obligations and pricing terms buried in free text across hundreds of contracts. The agent reads them once and keeps a system of record updated as amendments land.

Key terms and renewal dates extracted and tracked against a calendar
Amendments matched to the original clause, not filed as a new document

Anything with a workflow behind it

Ask us

Claim forms, correspondence, applications, handwritten notes, a mix of structured fields and free text on the same page. If there's a document and a process it should feed, the pipeline is the same. We build it to your document, not the other way round.

01 How it works

A validation window on the fields that need one, not a queue of documents to check.

Confidence split · illustrative, not a measured result

High confidence · moves straight through Below threshold · routed to the validation window

1 Any format in

A scan, a photograph, a native PDF, a structured form. The pipeline reads the document as it actually arrives.

2 Extract and score

Every field comes out with a confidence score, not a single pass-fail on the whole document.

3 A person checks the low-confidence fields

Only the fields the model flagged, shown next to the source document. Nothing else needs a second look.

4 It moves forward

Confirmed data is pushed into the target system, and the correction becomes training signal for the next document like it.

02 What the research says

These are not our numbers. The strongest one is forty years of peer review.

We have not published a client result yet. What we can show you is the published error rate of the thing you are doing now, which is the only baseline that matters when deciding whether extraction is worth it.

Data entry error rate, by method meta-analysis · 93 papers

READING OFF A DOCUMENT

6.57%

OPTICAL SCANNING

0.74%

SINGLE KEYING

0.29%

DOUBLE KEYING

0.14%

Systematic review
& meta-analysis,
93 papers, PMC

Day-to-day manual entry share of fields, experienced staff

TYPICAL

1 – 4%

BEST CASE, VERIFIED

0.5 – 1%

UNDER TIME PRESSURE

18 – 40% OBSERVED

Aggregated published
benchmarks, 2026

Automated extraction accuracy vendor-reported. read the note.

STRUCTURED FIELDS

99.5 – 99.9%

Vendor and integrator
aggregations, 2026

Treat that last row carefully, including when we say it. A 99% accuracy claim at character level is a completely different number from 99% at record level, and vendors rarely say which one they mean. At record level, 99% still means one in a hundred documents needs correcting. The meta-analysis above is peer-reviewed across 93 papers. The accuracy row is vendor-aggregated and should be read as direction of travel, not as a promise.

Never evaluate extraction on clean PDFs. Every tool looks good on a well-formatted invoice. Test it on the faxed copy, the handwritten form, the one where the total is buried in a table footnote. That is where the differences show up, and it is how we scope. Send us your worst documents, not your best ones.

03 Document fit

Not every document is worth a pipeline.

Our assessment, not a measurement. High variability is not a reason to say no. No downstream system is.

VOLUME

LAYOUT
CONSISTENCY

CLEAR TARGET
SYSTEM

SAFE TO
AUTO-POST

OUR READ

Invoices & AP documents

START HERE

Claim forms & applications

STRONG

KYC & onboarding files

EVIDENCE ONLY

Contracts

GOOD, SCOPED

One-off correspondence

WE SAY NO

STRONG CASE PARTIAL WEAK COUNTS AGAINST

Where we say no

Low-confidence fields never post without a person seeing them, whatever the deadline. Identity and compliance decisions get an evidence file, not an automatic pass. And every extraction keeps a trail back to the source document, so nothing goes forward unaccounted for.

04 How we engage

Three steps. You can stop after the first one.

01 FREE

Discovery

We sit with whoever actually clears the exception queue and watch a normal week. Where one of our builds fits, we demo it on your own samples. You get a written assessment either way, including the case for not building anything. If a scheduled job fixes it cheaper, we tell you that and we are done.

02 FIXED SCOPE

Build

One process, scoped and priced before we start. Built against your systems and your rules, not a demo environment. You see it running on your own data, not on ours.

03 YOU OWN IT

Handover

Code, documentation and runbook handed to your team. No licence, no lock-in, no retainer you have to keep paying to keep it working. Keep us on if you want more built.

Build pricing is set per engagement and quoted in writing before any work starts.

Send us your worst documents.

The faxed copy, the handwritten form, the one with the total buried in a footnote. Discovery is free, and we'll show you field-level results on them.