Document understanding for a distribution company
Sales documents arrived in every format imaginable and were re-keyed by hand. An LLM-based extraction system with human review now handles the routine volume — with people approving what matters.
The situation
Orders and contracts arrived by email, Zalo, and paper — scanned, photographed, typed in inconsistent formats, in Vietnamese and English. Staff re-keyed everything into the company’s systems by hand. The cost was hours per day; the risk was the quiet kind: transposed figures and missed lines discovered only when a customer complained.
The approach
Discovery mapped the document journey end to end and set a deliberate constraint: the system would extract and propose, but humans would approve. AI that touches order data must be auditable.
The pipeline extracts structured fields from incoming documents using an LLM, validates them against the company’s schemas and confidence thresholds, and routes anything uncertain into a review queue where staff confirm or correct in seconds rather than minutes. Approved data is written back into the existing system; every extraction keeps an audit trail from source document to final record.
Because documents arrive in Vietnamese as well as English, extraction quality is monitored per document type, and the system is improved continuously against real volume — not a demo set.
The outcome
Routine documents no longer wait for manual re-keying; staff time moved from typing to checking.
A growing share of documents passes straight through; the rest flow through a fast, supervised review queue.
Errors are caught at review, with the source document one click away — instead of surfacing as customer complaints.
Where it stands now
The system runs in production under ongoing improvement: new document types are added as the business identifies them, and extraction quality is reported monthly in plain numbers.