What does scoping a first sufficiency pilot look like for a lending environment?

An environment lead at a team building agent benchmarks wants a first look at sufficiency decisions in lending. They ask MirrorFirm for a sample based on Aster Financial, the sample organisation, covering personal loan applications.

The scoping conversation settles four things. - **Case type:** personal loan applications such as APP-2187, a single product. - **Requirements:** proof of identity, income evidence and a statement of expenses, with income evidence varying by employment type. - **Tools:** `list_documents`, `read_document`, `get_requirements`, `send_request`, `set_status` and `escalate`. - **Outcomes:** sufficient, one gap, several gaps, resolvable from held records, and a conflict that should be escalated.

The sample release holds cases in JSONL with the requirement table, held records and an answer key, plus a reset entry point with seeds. APP-2187 is a self-employed applicant who supplied one payslip, which is the wrong evidence for that employment type. The key expects a request for tax returns or accountant-prepared statements, not another payslip, and a status change to Awaiting Information.

The lead runs two agents against the sample, regenerates a case from its seed to confirm reset works, and has a lending reviewer read a spread of cases for plausibility. Findings go back as change requests: tighten one requirement, add a variant with an expired identity document.

**What it teaches** - A pilot works best when it fixes one product, one table and one tool surface. - Review findings can feed the next sample as specific changes. - Reset and seed checks belong in the first sample, not later.