What does a first pilot look like for a professional services engagement request environment?

The environment lead at a team building agent simulations for professional services firms drafts a pilot request for engagement requests such as ENG-2290, where a prospective client asks for tax advisory work.

They ask for one request type and four exception families: missing authority to act, conflict-check hit, incomplete entity details and withdrawn request. They specify that tasks should be delivered as JSONL specifications with Parquet snapshots, and that the reset call must be deterministic.

On receiving the sample, the lead's first action is not a training run. A test engineer loads the snapshots, runs the consistency checks, and executes a scripted reference solution on every task. Two tasks fail because the expected queue references a conflict-review team that is missing from the queue table. The lead sends the findings back with the task ids.

**Consequence:** the next version fixes the queue table and adds a schema check, and the team can decide, with evidence, whether to widen to a second request type or deepen the exceptions in the first.

**What it teaches:** ask for the exception families and interface you need, test the reference solution before any agent, and keep a defect list keyed by task id. Licence and scope decisions follow once the sample has passed your checks, and are agreed directly.