Will AI invent numbers in procurement work?

Yes, it can. Not every model invents every time. The July 2026 PAIR-20 run was a set of single dated runs, not averages, so read the result as "this happened" rather than "this always happens."

In that run, the fabrication problem showed up twice, both times in tasks that were built to test blank-field discipline and cross-document checking.

In the supplier-risk-register task, Copilot's structured-brief run read Dravec as if its backup contract had expired, even though the case file did not say that. The invented issue then survived into the compressed VP-ready summary. In the MRO catalog task, Copilot wrote "Likely 100ct" for Atlas's unstated box quantity. The case file said the quantity varied by item and needed confirmation.

Those are not style problems. They are hard facts being added to a procurement deliverable.

A fabrication is worse than a miss. A missed risk can be found by a reviewer who checks the source file. An invented expiry or invented pack quantity sends the reviewer looking for evidence that never existed. In a working team, that can waste time, misdirect a supplier call, or put a bad number into a recommendation.

The risk appears where the table looks useful. Both examples were the kind of tidy output people want from AI: a risk register and a merged catalog. The dangerous cell was not loud. It looked like one more filled-in field.

The control is source-status discipline. Ask the model to mark missing fields as missing. Then check that it actually did. If a source file leaves a contract date, backup status, pack size, or denominator blank, the answer should say "confirm with supplier," not fill the gap with a confident-looking number.