Engagement snapshot
- Mandate
- Increase underwriting throughput without weakening compliance controls.
- Timeline
- 12 weeks from evidence mapping to production launch.
- Team shape
- Head of risk, 2 back-end engineers, 1 ML engineer, and 4 underwriting reviewers.
The problem
Underwriters manually compared onboarding forms, web presence, sanctions data, processing intent, and historical risk signals before approving merchants.
What we built
Created an evidence-gathering agent that prepared structured risk memos, highlighted inconsistencies, and triggered manual review for high-risk or ambiguous merchants.
Operating context
Merchant onboarding volume was growing faster than the underwriting team. The platform needed a way to standardize evidence review without automating approvals that still required policy judgment.
Key constraints
- The system could never make final merchant approval decisions on its own.
- Sanctions and watchlist checks had to be deterministic and logged separately from model reasoning.
- Risk ops needed a memo format they could challenge, edit, and export into their existing review record.
What we built
Evidence orchestration layer
Collected KYB records, website scans, sanctions results, and onboarding declarations into one review bundle.
Policy-shaped memo drafting
Used model reasoning to organize evidence into the underwriting team standard rather than inventing a new workflow.
Manual decision boundary
Kept all binding actions and final approval states behind the human reviewer, with structured disagreement capture.
Delivery path
Policy decomposition
Translated underwriting policy into deterministic checks, weighted evidence classes, and human-only decision points.
Reviewer shadowing
Observed live underwriting to understand where memos were useful and where analysts still needed raw evidence.
Production tuning
Measured escalation quality, memo edits, and approval reversals to tighten prompts and evidence sequencing.
Why it mattered
The most valuable shift was consistency. Junior underwriters had a better first pass, senior reviewers spent less time reconstructing the file, and compliance retained a clean audit trail.
Implementation notes
- Agentic evidence collection worked only after deterministic checks and sources were locked down.
- Reviewers wanted less prose and more explicit evidence blocks than the first draft provided.
- Escalation precision mattered more than maximizing auto-cleared merchants.