Why are mismatch tasks a good test for business agents?

**Why not just test lookups and summaries?**

Lookups show whether an agent can fetch a field. Mismatch tasks show whether it can decide which fields matter, in what order, and what to do with the result.

**What does a mismatch task exercise that other tasks do not?**

- **Hypothesis testing.** The agent must propose a cause and then look for evidence that would rule it out. - **Source authority.** Often one record is correct and the other is stale. The agent must know which, and the policy usually says so. - **Action limits.** Fixes sit behind permissions, such as who may post an adjustment or reverse a payment. - **Escalation.** Some cases should not be fixed by the agent at all, and stopping is the right answer.

**Is there one right answer?**

When a case is generated, its root cause, correct fix and acceptable alternatives can be recorded alongside it. That makes automated scoring possible without relying only on a model grader.

**Does this replace real-world testing?**

No. Generated cases are synthetic and informed by how businesses work. They suit controlled, repeatable evaluation. A buyer should still inspect how well the cases match their own processes and check for disclosure risk before relying on them.