™AI implementation · Regulated manufacturing
An AI assistant for supplier document intake
Incoming supplier documentation arrived in every format imaginable and was re-keyed by hand. We built an assistant that reads it and proposes structured fields for a person to confirm.
The challenge
A manufacturing operation received supplier documentation as PDFs, scans, spreadsheets and email text, with no consistent layout between suppliers. Two staff members spent a significant share of their week re-typing the same fields into an internal system.
The work was tedious, and because it was tedious it was error-prone. Errors surfaced weeks later, usually during a review, when tracing the original document was difficult.
Approach
We started by mapping the existing process with the people doing it, which surfaced something useful immediately: three of the fields being captured were no longer used by anyone downstream.
We then agreed an explicit rule, the assistant would never write directly to the system of record. Every extraction would be proposed and confirmed by a person.
What was built
A document intake application that accepts a file or a forwarded email, uses a language model to extract the agreed fields, and presents them side by side with the source document for confirmation.
Low-confidence extractions are flagged rather than hidden, so reviewers know where to look. Every confirmed record keeps a link to the source file and a record of what was changed during review.
Technology and process
- A language model for extraction, with the source document retained and referenced rather than discarded.
- Deterministic validation rules applied after extraction, formats, ranges and required-field checks.
- Role-based access, with review and approval separated.
- An audit record of every extraction, edit and approval.
Outcome
Re-keying was replaced by review, which staff described as substantially faster and considerably less tiring. We do not have published figures for the time difference.
The more valuable change was traceability: a question about a figure can now be answered by opening the source document attached to the record, rather than searching an inbox.
What we learned
- Mapping the process before building removed three fields from the work entirely, the cheapest improvement in the project.
- Keeping a person in the approval path was not a limitation; it was what made the tool acceptable to use in a regulated setting.
- Flagging uncertainty openly built more trust than a higher apparent accuracy rate would have.
Client identity and commercially sensitive detail have been withheld. Outcomes are described qualitatively where measured figures are not available for publication.