Can AI rank bids correctly once the math is done?

Sometimes, but the July 2026 PAIR-20 run shows the last step can still fail. In T05, the task was to compare three consulting proposals on an all-in basis. The headline spread looked larger than the real decision spread; after hidden costs and scope adjustments, the all-in gap across the three bids was $750 (case figure).

That kind of comparison is unforgiving. If the all-in spread is that tight, the ranked table is not a decoration. It is the recommendation.

In the July run, ChatGPT found the right components and calculated the all-in math correctly, then ranked Pinnacle first in the summary table even though Pinnacle had the highest all-in total. The arithmetic underneath and the table above it disagreed. Gemini had a different presentation problem: it showed Bridgepoint's correct all-in figure in the narrative, but listed the effective blended rate as base-fee-only, stripping out the costs that made the comparison honest.

Correct math is not the same as a correct decision. A model can extract the fee, add the scope adjustment, include travel and expenses, and still put the wrong supplier at the top. The failure is not in the calculator step. It is in carrying the result into the recommendation.

The table needs its own review. Procurement teams often review the working and then scan the final ranking. Reverse that habit. Read the ranked row against the all-in total, then read the all-in total against the cost build-up.

Small spreads make presentation errors material. A $750 case spread leaves no room for a table contradiction. If a model ranks the highest all-in bid first, a buyer can forward the wrong recommendation while every number in the working still looks clean. For bid ranking, verify the final order separately from the math.