Make an AI vendor prove the claim you are actually buying.

Approve the vendor, approve with data-handling conditions, dual-run, or decline. The owner is usually the operational lead plus security. Legal and procurement review contract and residual model risk.

This is not a generic software bake-off. The load-bearing claims are usually about training data, retention, evaluation on your task, and whether you can leave.

Vendor Selection freezes commercial and implementation criteria before the second demo. This page is the AI-specific card: model claims, evaluation method, data path, and output provenance. Use both when the buy is an AI product from a named vendor.

Champion can take the leading vendor. Challenger attacks unreproduced evals and lock-in. Scout looks for current incidents you can verify. Inspector marks a vendor leaderboard as marketing. Judge can refuse a strong winner when your task was never tested.

A generated paragraph in the demo is not evidence the product works on your data.

Connectors to the vendor’s platform are not part of Argumont today. The Decision record still has to hold the Human Decision if Legal cannot accept the data terms.

Champion and Challenger test the case, Inspector checks the evidence, and Judge records the Verdict for the Decision Owner.

Approve vendor A for an internal summarization task, approve with no-training terms, dual-run A and B on a named sample, or decline. Vendor win-rate slides stay claims. If the Verdict favors A and Legal blocks training-on-customer-data, the override stays beside that Verdict.