What does a first pilot for an intake agent look like in practice?

A model lab's applied team is building a back-office agent and wants to test one question: can it decide whether a new loan application is complete and route it correctly?

They scope a pilot around Aster Financial, MirrorFirm's sample organisation. The scoping note names one product, personal loan applications, three channels (web form, broker email and branch note), and eight policy rules covering identity documents, income evidence and referral to a credit officer. It asks for a mix of clean, incomplete, duplicated and escalated cases, a tool contract in JSON Schema, and reference outcomes.

The lab's evaluation lead reads the small sample. Case `APP-2187` is a broker email with payslips attached but no identity document. Case `APP-2203` is a second submission from the same applicant after a typo correction. They trace decisions to rules, run a baseline model, and list ambiguities: one rule lacks an exception path, and one reference accepts only one of two valid queues.

They send the list back. The next version addresses those items, and the lab then decides whether to widen scope to a second domain or to more cases of the same type.

**What it teaches:** judge the sample by tracing decisions to rules and finding ambiguity, then let those findings set the next scope. The output is generated throughout.