How should I scope a first pilot for conditional-approval material?

A useful first pilot is narrow. Pick one domain, one decision type and a short list of condition types, then grow from there.

**Choose the process.** Conditional loan approval through to settlement, or supplier onboarding through to first order, are typical choices. A single process lets you judge fit against something you know well.

**Choose the depth.** Decide whether you need a flat dataset for training, a task pack with ground truth for evaluation, or a simulated environment with tool calls and reset. The environment takes more scoping, so a task pack is often a sensible first step.

**Choose the exceptions.** Name the three or four exceptions that hurt most in your process, such as stale evidence, a waiver outside authority or a lapse between steps.

**Agree the limits.** Output is generated, so it will not reproduce your internal systems exactly. Treat it as a way to exercise behaviour, and expect to calibrate against your own environment. Licence terms, delivery format and volume are agreed for each engagement, and nothing here sets them.

**Define success before you start.** Write down what result would justify going further, such as clear separation between a naive and a careful baseline, or a useful list of failure modes.

A pilot should end with a decision: widen, adjust the scope, or stop.