What is spend analysis (and can AI do it)?

Spend analysis is the work of pulling purchase data from ERP, P2P, and payment systems, then cleansing, classifying, and examining it so the procurement team can see where money goes. The classification step — assigning each transaction to a category, supplier, business unit, and contract — is the heaviest manual effort and the place where errors compound: a misclassified line hides savings from every report built on top of it. The AI question splits cleanly in two: classification and arithmetic.

AI can draft the classification layer. Category assignment is pattern work: reading line descriptions, matching suppliers to segments, flagging outliers. A structured brief that defines the category tree, names the known edge cases, and shows a few already-classified examples gives the model something to match against. The draft still needs review — a misclassified tail-spend line may not matter, but a misclassified strategic supplier changes the savings story.

The totals need their own check. In the July 2026 PAIR-20 T03 run, all four platforms rebuilt a landed-cost stack and every final number differed from the EUR 7,940 case answer. ChatGPT and Gemini were EUR 112 low. Claude was EUR 87 low. Copilot was EUR 48 high, and its table and prose contradicted each other. Spend analysis totals face the same risk: the output looks complete, the number has two decimal places, and the answer can still be wrong.

Normalize before you aggregate. A spend cube built on mixed units, currencies, or delivery terms will produce variances that are artifacts, not findings. The same principle shows up in PO price variance work: a carton price and a case price cannot be compared until they sit on the same basis. Clean the dimensions before trusting the sum.