Consumer lending, decided in hours: inside an agentic underwriting build
How an underwriting agent assembles the credit file, applies policy, and drafts decisions — cutting time-to-decision by 60% while keeping a human on every approval.
The problem
A mid-size consumer lender was drowning in its own growth. Every application meant the same ritual: chase income documents, re-key data from pay stubs and bank statements, walk the file through a policy checklist, write up the decision. Underwriters spent roughly 70% of their day assembling files and 30% actually underwriting. Time-to-decision stretched past three days — and applicants who wait three days shop elsewhere.
The hypothesis
Most of the assembly work is not judgment — it's retrieval, extraction, and rule application. Our hypothesis: an agent could own the file build end to end, apply the credit policy deterministically, and hand the underwriter a decision-ready file with a drafted recommendation, so human time concentrates on the 20% of files that genuinely need judgment.
The build
We shipped the system in three layers:
- Document AI intake — extraction models read pay stubs, bank statements, and IDs, normalize the fields, and cross-check income against stated figures.
- Policy engine + agent core — credit policy expressed as explicit, versioned rules; the agent walks each file through the checks, gathers missing items by drafting applicant requests, and computes exceptions.
- Decision drafting with a hard human gate — the agent writes the recommendation and rationale, citing every source document. No decision is issued without an underwriter's sign-off; declines always route to a human first.
Design choice that mattered: we kept the policy logic outside the model. The LLM reads, assembles, and drafts — but pass/fail against credit policy is deterministic code the risk team can audit line by line. Regulators don't accept "the model decided."
Rollout
We ran four weeks of shadow mode — the agent processed live files in parallel while underwriters worked normally, and we measured agreement. Disagreements were reviewed weekly with the credit team; most traced to ambiguous document quality, which shaped the exception rules. Only after agreement stabilized above threshold did drafted decisions reach the queue.
Results
Time-to-decision dropped from days to hours for clean files. Underwriter throughput tripled without adding headcount, and — the part risk leadership cared about most — every decision now carries a complete, source-linked audit trail that used to exist only in an underwriter's head.
What we'd tell you
- Start with the file build, not the decision. The assembly work is where the hours are, and it's the least controversial to automate.
- Shadow mode is not optional in lending. Agreement data is what earns the risk team's trust.
- Keep declines human. It's the right call ethically and the safe call under fair-lending scrutiny.
More from the field.
Early-warning credit risk
The 6 a.m. advisor brief
Tax-aware rebalancing
Bring us a hypothesis. Leave with a system.
Tell us what's eating your team's time. We'll give you an honest read on whether AI is the right tool — and if it is, a scoped v1 with a timeline and cost.