From documents to data you can check.
Document intelligence turns invoices, contracts, forms, certificates, reports and correspondence into structured records: what kind of document it is, which fields it contains, and how it differs from the order, agreement or template it should match.
The aim is not to take people out of document work but to send them only the cases that need judgement — with every extracted value traceable to where it came from.
01Where it helps
Recurring documents whose contents someone types into another system.
The best candidates arrive in volume, follow a recognisable type, and feed a decision or a system further on.
Potential use case Examples to assess, not existing deployments.
Accounts payable intake
Classify supplier documents, extract header and line-item fields, and match them to orders and deliveries before approval.
Contract review support
Extract parties, dates, renewal and liability clauses, and flag where a contract departs from your standard template.
Forms and applications
Check that required fields are present and consistent, and route incomplete submissions back with the gaps listed.
Certificates and compliance records
Capture certificate numbers, issuers and expiry dates, and list what is missing or about to lapse.
Correspondence routing
Sort incoming letters and emails by type and urgency, with the key references extracted for the receiving team.
02How it works
Classify, read, extract, validate — then decide who needs to look.
- InputIntakeEmail, upload, scan or API
- ModelClassifyDocument type decides the schema
- ToolOCR & layoutText, tables, reading order
- ModelExtractInto the type's schema
- CheckValidateTypes, ranges, cross-checks
- CheckConfidence gate↳ below threshold: exception queue, then a reviewer
- OutputRecord + provenanceEach value linked to its page and region
- Input
- Model
- Tool
- Check
- Output
Define types and schemas
For each document type, agree the fields, their formats and the rules that make a record valid. The schema is the contract with every system downstream.Handle the source
Born-digital PDFs are read directly; scans and photos go through OCR with layout analysis so tables, columns and reading order survive. Pages too poor to read reliably go straight to a person.Classify before extracting
Identifying the type first means each document is read against the right schema, and unfamiliar types are caught instead of being forced into the wrong one.Extract with provenance
Every value is stored with the page and region it came from, so a reviewer can see the source beside the extracted field.Validate every record
Field checks (dates, totals, identifiers) and cross-checks (lines sum to the total; the supplier exists in master data) run before anything is accepted.Learn from corrections
Reviewer corrections are recorded, used to improve extraction and to show where the process is weakest.
03Validation in practice
What a schema with rules looks like.
An invoice-like example. Real fields and rules are agreed for each document type and the systems the data feeds.
Sample data Illustrative fields and rules only.
| Field | Rule | If it fails |
|---|---|---|
| Document number | Present, and in the issuer's known format | Exception queue |
| Issue date | A valid date, not in the future | Exception queue |
| Issuer | Matches an approved record in master data | Held for review |
| Line items | Quantity × unit price equals each line total | Mismatched lines flagged |
| Total | Equals the sum of lines plus tax | Automatic posting blocked |
04Business documents and engineering drawings
This page is about business documents. Our products read engineering drawings.
General document intelligence works mostly with text, forms and tables, and is delivered as scoped work. Ramerion’s two products are specialised workflows that interpret engineering drawings into 3D models and structured engineering data, with engineers reviewing the results.
05Limits and security
Some documents will always need a person.
Limitations
- Handwriting, low-resolution scans, stamps over text and unusual layouts lower accuracy.
- Extraction can be confidently wrong — which is exactly why validation and thresholds exist.
- A new supplier template or form version can degrade results until it is added to the test set.
Security implications
- Documents often hold personal and financial data; access, retention and deletion rules cover source files, extracted data and logs alike.
- Text inside a document is processed as data: instructions embedded in it cannot change how it is handled.
- If an external OCR or model service is used, what it receives and keeps is part of the agreed data boundary.
06How success is measured
Field by field, against documents people have checked.
Measured on a reviewed sample of your own documents — see evaluation and governance.
- Field accuracy
- Precision and recall per field and per document type, against reviewed ground truth.
- Straight-through rate
- Share of documents accepted without correction — always reported with the error rate among them.
- Exceptions
- Volume, reasons and time to resolve.
- Classification
- Correct type assignment, including recognising documents that fit no known type.
07Related
Related services.
Next step
Show us the documents that slow your team down.
A representative sample you are permitted to share — redacted where needed — and the fields you need from it are enough to judge feasibility and likely accuracy.