Paperwork turned into structured, checkable records.
Documents arrive in whatever shape the sender chose. The job is to turn that into a fixed structure your systems and your finance team can both trust.
Eight stages, one of which is a person.
Extraction alone is not the deliverable. Validation, the confidence threshold and the review queue are what make the output safe to post into a finance or operations system.
Corrections made by reviewers are captured, so the same failure pattern can be addressed during optimisation rather than repeated indefinitely.
- 01Ingest
Documents arrive from email, upload, a shared folder or a scanning workflow.
- 02Classify
The document type is identified so the correct field set and rules apply.
- 03Extract
Agreed fields are read into a fixed structure, with the source location retained.
- 04Validate
Format, arithmetic and cross-reference checks against your own records.
- 05Score
Each field carries a confidence indication; the threshold is set by you.
- 06Route
Clean items continue; low-confidence or incomplete items go to the review queue.
- 07Review
A person confirms or corrects, and the correction is captured for tuning.
- 08Post
The approved record is written into the system of record with an audit entry.
Structured output only makes sense once the fields are agreed.
We define the field set per document type with the team that uses the data — not from a generic template.
Typical document types and fields
Indicative examples. Your actual field set is confirmed during scoping and written into the scope of work.
Supplier invoices
Fields
Supplier, invoice number, date, line values, VAT, total, PO reference
Purchase orders
Fields
PO number, requester, cost centre, items, approved value
Application forms
Fields
Applicant details, requested service, declared information, attachments
Delivery documentation
Fields
Reference, consignee, quantities, dates, exceptions noted
Contracts and agreements
Fields
Parties, term dates, renewal notice, values, key clauses flagged
Identity and licence documents
Fields
Document type, holder name, number, expiry — handled per your policy
Operational correspondence
Fields
Sender, subject, request type, referenced records, required action
Internal forms
Fields
Submitter, department, request category, approval path
The threshold is a commercial decision, and it is yours.
A stricter threshold sends more items to review and catches more errors. A looser one moves faster and accepts more risk. There is no universally correct setting — it depends on the value and consequence of the document.
We set it with you, monitor how much reaches the queue in the first cycles, and adjust deliberately rather than quietly.
We do not publish accuracy percentages. Extraction quality depends on your document quality, and any figure quoted before seeing your paperwork would be marketing, not measurement.
What is stored, and what is deliberately not.
Identity documents, personal data and commercially sensitive contracts are handled under an agreed policy: retention period, access list, storage location and deletion route are all stated before processing starts.
Where a document class is too sensitive to process, we exclude it from the pipeline entirely rather than applying a weaker control to it.
Data and security approach →Start with one document type.
Working on your real paperwork gives an honest read on quality, exception volume and reviewer effort.