Why is this workflow a better test for agents than question answering?

**Why use loan booking instead of a question-answering benchmark?**

Because the result is a set of changed records. After an agent runs, you can inspect the account, schedule, fees, collateral links and servicing handoff and compare them with the approved terms. A fluent final message tells you little by comparison.

**What makes it hard?**

Several things at once. Steps depend on order: the account must exist before a schedule attaches to it, and collateral must be linked before release to servicing. Terms arrive in mixed forms, such as structured fields, a credit memo and an email from the credit officer. Some cases should not be booked at all until a condition is cleared.

**Does it test judgement or only data entry?**

Both. The agent has to decide when to proceed, when to raise an exception and who to hand it to. A task pack can include cases where the correct action is to stop and escalate.

**Does it work across domains?**

The pattern applies to bank lending, equipment finance, insurance premium funding and vendor finance. Each changes the product rules and the systems involved, and a release can be scoped to one of them.

**Is the data real?**

No. It is generated, and informed by direct observation of how businesses operate. Inspect grounding and disclosure risk in any sample.