How should we scope a first pilot for set-up agents?

A useful first pilot is small and specific. The aim is to learn whether generated set-up material exercises your agent in ways you care about, not to cover every product.

**Pick one domain and one path.** For example, one product line with a handful of record types: approval, customer, account, service and task. Avoid starting with several domains at once, because you will not know which difference explains a result.

**Define the case mix.** Include clean cases, near-matches, incomplete approvals and a few cases where the right answer is to refuse or escalate. Ask for a small held-out set that your team does not use for tuning.

**Agree success measures first.** Write down what you will look at: final-state accuracy by failure mode, honesty checks, realism ratings from someone who knows the process, and how easily the release loads into your harness.

**Settle the limits.** Ask what the release can and cannot be used for, how many regenerations are possible and how fixes are handled. Scope, quantity and timing are agreed with MirrorFirm for your case; do not assume them.

**Plan the review.** Have a practitioner read a sample, and have security or governance staff review disclosure risk under your own process.

At the end, decide whether to widen scope, change the case mix, or stop. A pilot that ends with a clear no is still a useful result.