Key Results

50%Faster underwriting review
1,000Loans processed per month
~40%Conversion lift
$6MIncremental profit

Background

After the appraisal workflow and agent portal work, MegaLoans turned to the loan application itself: the manual verification that sat between a borrower applying and a processor issuing a pre-approval.

The problem

  • Borrowers uploaded pay stubs and bank statements by hand to prove income and assets.
  • Processors requested missing documents by hand and created conditions by hand. Conditions are the open items a loan has to clear before it can move forward.
  • All of this slowed origination and initial quotes. When income wasn’t verified correctly up front, it came back later as rework and inaccurate quotes.

My role

I owned the strategy and requirements for an AI agent that verifies income, employment, and residency, and I built early prototypes to test the approach. Engineering and data science built and trained the production system.

How the agent works

  • It pulls data directly from the source. Income and employment come from payroll providers through Argyle, assets and bank data from Plaid, and identity and residency from Persona. Borrowers no longer have to upload documents for these checks.
  • It starts from GSE guidelines. Every file is evaluated against standard GSE guidelines, with our own exception rules layered on top for cases the guidelines don’t settle.
  • It has three outcomes. If the agent can clear income, employment, and residency on its own, it issues an automated pre-approval. If it can’t, the file goes to a person in one of two ways: someone checks the agent’s finding, or someone completes that step by hand.

Guardrails

  • Monitoring on exception handling, so edge cases get routed to a person instead of failing quietly.
  • A feedback button that lets users rate how the agent handled a task. The team used those ratings to track accuracy and decide what to fix first, especially when document formats changed or new edge cases appeared.
  • An AI security review before launch.

Results

  • Underwriting review time fell 50%, which freed up processor capacity.
  • Processing capacity grew to 1,000 loans a month.
  • Conversion across the point-of-sale flows rose about 40%, mostly from faster pre-approvals.
  • The combination of lower costs and new revenue produced $6M in incremental profit.

What I learned

  • Design the handoff before the automation. The most important decision was giving the agent three outcomes. With only pass or fail, every edge case would have landed in one manual queue.
  • Feedback keeps an agent working. Without a way for users to flag bad results as they happened, changes in document formats would have gone unnoticed until they showed up in the metrics.
  • Leading an AI build takes a different skill. Much of my job was turning vague edge cases into rules the system could act on, and deciding when a person should handle the file instead.