Can you trust AI's arithmetic in procurement files?

AI arithmetic in procurement files is useful, but it is not a final control. The danger is that the table looks complete: line items are named, currencies are converted, and the answer has two decimal places. That finish can make a wrong number feel more checked than it is.

PAIR-20's July 2026 T03 task tested a synthetic freight award with three quotes, three currencies, and different Incoterms. The correct move was to rebuild each quote to the same landed-cost scope: buyer cost from supplier warehouse to cargo cleared at Hamburg port. The case answer for TranStar was EUR7,940.

All four platforms saw that the headline TranStar quote was incomplete. All four rebuilt the stack under the structured brief. None reached the case answer. ChatGPT and Gemini landed EUR112 low. Claude landed EUR87 low. Copilot received a partial verdict because its table and prose contradicted each other.

The visible structure is still valuable. A table that separates freight, terminal handling, customs, insurance, surcharge estimates, currency conversion, and Incoterms scope gives a buyer something to audit. That is better than a single unexplained recommendation.

The final number needs its own check. In the July run, agreement on the shape of the calculation did not produce agreement on the answer. The platforms were wrong by different routes, which is exactly why a clean total should not be read as proof.

Treat disagreement as a worklist. If two passes produce different landed costs, do not average them and move on. Trace the line items that changed. For a working procurement team, the useful AI output is the calculation map; the buyer still has to verify the assumptions that move the award.