Your AI Read the Wrong Revision
Hypothetical scenario: a small fabrication shop quotes a bracket from drawing revision B. The customer sends revision C with a different material. An agent retrieves B, calculates beautifully, and prepares the wrong purchase request. Excellent arithmetic. Expensive scrap.
The problem is not necessarily model intelligence. It is document authority. A folder called “latest final” is not a control system, however confidently everyone says it.
Approval Belongs to a Version
A forward deployed AI engineer works with estimating, purchasing, and production to define which drawing is released, who can replace it, and which jobs a change affects.
Store the part number, revision, release status, effective date, and source document identifier together. Search only released revisions applicable to the job. The agent extracts proposed material and quantity changes with page references. An authorized engineer checks the drawing and approves the resulting bill of materials. AI does not release parts for manufacture.
Bind that approval to the exact source revision and proposed ERP change. If either changes, require another review. Otherwise, yesterday’s permission becomes today’s blank cheque.
Reject Stale Writes
For an ERP API that supports it, use the conditional request mechanism defined in RFC 9110. Send the record’s previously read entity tag with the update. The server checks that version before applying the change, preventing an older proposal from overwriting newer data.
This guards the target record, not every related drawing. The integration must separately verify the approved source revision. If either check fails, stop and refresh the review. Never silently overwrite a conflict.
The reusable agentic asset is a revision resolver, evidence extractor, approval binding, and guarded ERP connector. Test it against superseded drawings, ambiguous part numbers, changed quantities, and approvals arriving after a revision. Reuse those tests across shops; configure each shop’s release rules.
Count Decisions, Not Documents
Illustrative assumptions: 80 monthly revision reviews reduced from 20 to 10 minutes release 13.3 hours. At an assumed $60 per hour, that is $800 of capacity before implementation and operating costs, not realized profit.
Measure stale proposals blocked, review time, incorrect purchases, and actual scrap cost. Reading documents faster is useful. Buying the right material is the point.