Cut underwriting from 30 minutes to 5 minutes per file.

BH Capital Funding is a founder-led capital firm in NY running a high-volume underwriting operation in business finance and lending. The work was already flowing — the constraint was how fast files could move from intake to a lender submission.

Underwriters were doing high-judgment work buried under low-value work. Every file meant pulling data from documents by hand, checking it against lender criteria, and assembling a submission package — roughly 30 minutes per file before a human even made a decision.

Volume capped out at how fast people could read, so throughput was limited by reading speed. Approvals stretched into days, and too many files were declined for reasons that were knowable up front.

The mechanical steps — extraction, scoring against criteria, lender-fit routing — were rules-based and repeatable; only the final decision needed a person. Automating the front of the workflow would free underwriters for judgment and let cleaner files reach the right lenders.

We mapped the underwriting workflow end to end, then built an AI underwriting engine wired into BH Capital's CRM. It ingests a file, extracts and structures the data with AI-assisted parsing, scores it against a custom rules engine, routes it to the right lenders, and assembles the submission package automatically — with everything writing back to the CRM so the pipeline stays current. The stack runs on Python, the OpenAI API, Postgres, n8n, and HubSpot.

The engine runs inside the team's existing CRM workflow, so underwriters review and decide where they already work. Handoff included a walkthrough of the intake-to-submission flow so the team could operate and rely on it day to day.

This is a finance firm trusting an outside team with applicant financials and its own underwriting logic. We treated that accordingly: the engine automates the mechanical steps — extraction, scoring, lender-fit routing — while a human stays on every approval decision, and the whole flow runs inside infrastructure the firm already controls.

Files that took 30 minutes now take about 5, approvals that took days now land in 2–3 hours, and cleaner, lender-aligned submissions dropped the decline rate by roughly 35%. Underwriters spend their time on decisions, not data entry.

After a 12-week initial build, the engagement continues as ongoing optimization — tuning the rules engine and lender-fit routing as volume and criteria evolve.