What makes a booking task worth building an agent evaluation around?

Use this checklist to judge whether a booking task will give you a discriminating signal rather than a trivial pass rate.

- The approval has both structured terms and free-text conditions the agent must read. - More than one record must be written, with links that can be checked afterwards. - Some files should be held or returned, not booked. - Policy limits apply, such as authority levels, product eligibility or pricing exceptions. - A wrong write has a visible downstream consequence, such as a bad schedule or early disbursement. - The correct end state can be defined in advance and compared automatically. - The audit trail, including reasons and approver references, is part of the expected output. - Variation exists across products, such as term loans, asset finance and premium funding, so the agent cannot memorise one template.

If a candidate task meets most of these, it is likely to expose differences in reading, tool use and judgement. If it meets only the structured-terms point, it is probably a data entry test and will saturate quickly.