Tests for AI agents that move money.
Did the agent pay the right party the right amount, avoid being scammed, stay within its limits, and tell the truth afterwards? We answer each question from the ledger, not by asking another model.
Four ways an AI agent goes wrong with money
AI agent payment testing that grades what actually settled: who was paid, how much, how often, within what limits, and whether the agent told the truth.
Why AI agents pay the wrong payee or the wrong amount, how a duplicate payment happens after a retry, and how a ledger-graded test catches each one.
How fake sellers, lookalike listings and hidden instructions trick AI agents into paying, and how a scam test measures whether an agent notices in time.
Testing whether an AI agent keeps to its budget, its spending limits and its instructions, and whether its report of what it spent matches the ledger.
Why the ledger is the grader
A transcript records what an agent said it did. The ledger records what it paid: to whom, how much and how often. Grading against the ledger turns most of a payment test into a comparison anyone can check, and it gives the same answer every time the same agent runs the same task.
A result belongs to one version of one agent. When the model, the prompt or the tools change, test again.