What does a first pilot review loop look like for a lender's intake agents?
A head of evaluation scopes a personal loan intake pilot around Aster Financial, reviews the first sample, requests changes and sets baselines before extending.
The head of evaluation at a lender asks MirrorFirm for a pilot slice on personal loan intake. Aster Financial, MirrorFirm's sample organisation, supplies the setting: an intake team, a credit officer and a document policy for unsecured loans.
The brief is one page. Case type: unsecured personal loans. Required evidence: identity, two recent payslips, a bank statement and, for self-employed applicants, a tax return. Decisions: proceed, request, escalate. Defects: expired identity, payslips from different employers and a bank statement with missing pages.
The first sample holds a small set of cases, with gold labels in a separate file. One is APP-2187, a self-employed applicant who supplied identity, a bank statement and payslips from a previous job. The gold decision is request: the tax return is required and the payslips do not apply to a self-employed applicant. The status moves from RECEIVED to AWAITING_INFO.
Review finds the case useful but too easy to spot, since the file names reveal the defect. The head asks for neutral file names and a second defect in the same file, then checks that the labels reproduce from the policy.
Extension is considered only after baselines run: always-proceed and always-ask agents should both score poorly, and a reviewer should agree with the gold labels.
**What it teaches:** a pilot works as an inspection loop. The first sample exists to find what to change.
Treat the first sample as an inspection loop for finding what to change, not as proof that the data is good.
- Audience:
- Evaluation teams
- Content type:
- Scenario
- Buyer question:
- Scope and getting started
Content release: r-a220eb6e099b243b
Related questions
- How do environment reset, seeds and versioning work for a sufficiency environment?
- What licence terms and limits should I expect for a sufficiency release?
- How should I scope a sufficiency dataset, task pack or environment?
- What does a useful first pilot look like for a team building a utility onboarding agent?
- What should I ask for in a sufficiency sample or pilot?
- What should a first handoff package contain and how do we integrate it?