Capability 02 · Automation
A document arrives.
The fields are read, checked, and ready to review.
An extraction flow for teams that retype data from PDFs, certificates, and emails into the ERP — and catch the mistakes weeks later in the ledger.
The situation
Someone is retyping the same fields again.
Supplier invoices, certificates, and order documents arrive as PDFs and email attachments. Each supplier formats them differently.
A person opens the document, finds the supplier, the numbers, the dates, and types them into the system. Then they check the account code and move to the next one.
It works at low volume. It quietly breaks when the volume grows — and the errors only surface later, in the ledger or in a customer complaint.
The challenge
Reading the document is easy. Trusting the result is not.
A model can pull text from almost any document. The hard part is knowing when it is right.
A field that looks confident but is wrong is worse than no answer — because the team stops checking, and the error slips through.
So the flow needs more than extraction. It needs validation against real records, a confidence threshold, and a clear rule for when a human must look.
The capability
AI reads the fields. Rules decide what's trusted.
The workflow reads each document and extracts the fields. It validates them against your suppliers, accounts, and amounts, and scores how confident it is.
Confident, validated fields move forward. Anything unclear is held — with the original document shown beside it — for a person to confirm.
The team reviews exceptions, not every line of every document.
How the flow works
Before / after
A person opens each PDF, reads the fields, types them into the ERP, checks the codes, and fixes the mistakes when they surface weeks later.
Accuracy depends on attention on a busy day.
Fields are read, validated, and scored before anyone looks. Confident cases are ready to approve; unclear ones are held with the source.
The ledger sees validated data, not a fast typist’s best guess.
Controls
- Validation against system records before posting
- A confidence threshold decides what needs review
- The original document is shown beside every field
- Missing or unmatched data stops the flow
- No document posts without approval
- Untrained formats are routed to a separate queue
- Every extraction is logged
- The final write goes to the agreed system
Where this applies
What changes
Questions about extraction
How do we know the extracted data is correct?
Reading a document and trusting the result are two separate steps. The model reads the fields; rules decide which values are accepted, which are flagged, and which go to a person. Nothing reaches your systems before it has passed those checks, so the accuracy level is something you set rather than something you hope for.