What should my team prepare to start an escalation pilot?

A pilot goes faster when the generated material already matches how your agent sees the world. Useful inputs from your side:

- **Tool definitions.** Names, arguments and return shapes of the calls your agent uses, for example `get_case`, `escalate_case` or `record_decision`, or the OpenAPI or function schemas behind them. - **Role and grade names.** The titles and levels you want in the matrix, so tasks refer to `senior_credit_manager` rather than a generic label. - **Authority rule shapes.** How your limits are expressed: by amount, by exposure, by product, by risk grade. You do not need to share real figures, since limits can be generated. - **Status model.** Your case states and permitted transitions, such as `IN_REVIEW`, `ESCALATED` and `RETURNED`. - **Format preferences.** JSONL, Parquet or both, and any naming conventions for identifiers. - **Harness details.** How your tests call an agent, collect traces and record results, so grader hooks can be shaped to fit. - **Reset needs.** Whether you need per-task resets, seeded simulated reviewers or both.

Share interface material that you are comfortable sending, and avoid including real customer records. The pilot's inputs and outputs are agreed in scoping, and the delivered material is generated, so review it for grounding and disclosure risk before wider use.