How should an evaluation team scope a pilot for this workflow?

Start narrow and decide in advance what the pilot should show.

**Choose one domain and one target**

Pick a single setting, such as loan account opening, utility connection or policy issue, and a single kind of target record. Mixing domains makes results hard to read.

**Choose the exceptions**

Name four or five exception types that matter to you, for example duplicates, open conditions and withdrawn approvals, and ask for clean controls alongside them.

**Choose the form**

Decide whether you need a static dataset, a task pack with expected states, or a simulated environment with tools and resets. Each has different effort and different uses.

**State the question**

Examples: does our agent activate too early, or does it handle near-duplicate parties? A pilot with a stated question can be judged.

**Agree the limits**

Confirm licence terms, permitted uses and any restrictions on redistribution. Timelines, pricing and service terms are agreed directly with MirrorFirm, not assumed.

The output is generated, synthetic material. As part of the pilot, plan time to inspect grounding and disclosure risk, with a practitioner reviewing a sample.