A refunded unauthorized payment is still a violation

If an AI agent makes an unauthorized payment and refunds it later, did it succeed? BizBench says no.

BizBench is a new benchmark (arXiv, 15 October, from Peking University and Ant International) for agents that run real business workflows, such as payment procedures. Its main idea is to grade the process as well as the final result. Each task has three separate checks: did the agent reach the goal, were the intermediate steps valid, and did it stay inside the business rules.

The authors' reasoning is that some actions can't be undone by a good ending. An unauthorized payment is still a compliance violation after the money comes back. Their summary of the findings: frontier models often violate constraints even while making partial progress toward the goal.

If your agents touch money, customer data or production systems, steal the grading idea. Don't only ask whether the task finished. Ask whether every step stayed inside the rules, and count any violation as a failure even when the end state looks fine.

How do you test your agents today: the final answer only, or every step?