Premier Member Editorial: The Economics of Loan QC Are Changing With AI
Pranay Shetty is Co-Founder & CEO of Sei AI. Sei AI builds AI agents for banks, lenders, & servicers.
Quality control in mortgage lending has reached an inflection point. Between agency guidelines, federal regulations, state-specific rules, and investor overlays, a single loan file can require validation against hundreds of discrete checkpoints. The complexity isn’t just technical—it’s expensive and increasingly unsustainable at scale.
The Scope of the Challenge
Consider what QC teams face daily. Fannie Mae’s Selling Guide exceeds 1,200 pages with quarterly updates. Freddie Mac’s Seller/Servicer Guide runs similarly long. HUD 4000.1 spans 900+ pages. Each contains thousands of individual requirements that interact in non-obvious ways.
Take self-employment income. For a borrower with 25% ownership in an S-Corp, validation requires parsing two years of personal and business returns, completing Form 1084 or Form 91, analyzing business cash flow, verifying ownership percentages against K-1 data, and applying declining income trending rules if revenue dropped year-over-year. Miss one step, and you have a post-purchase defect sitting in your pipeline.
TRID adds another layer. Tolerance tracking across 10%, zero, and unlimited buckets. Redisclosure triggers for APR variances. Waiting period resets for product changes. Multiply this across recording fees and transfer taxes in 50+ jurisdictions, and manual tracking becomes a liability.
Then there’s the state regulatory patchwork—high-cost loan thresholds that differ from federal HOEPA, prepayment penalty restrictions that vary by jurisdiction, licensing supervision ratios, and extended rescission periods. A loan compliant at the federal level can still trigger state-specific violations.
The Cost Reality
Industry benchmarks paint a clear picture: pre-close QC runs 90-120 minutes per loan, post-close runs 80-140 minutes. Analyst capacity maxes out at 8-12 loans daily. Defect rates on initial review range from 15-35%. Fully loaded costs land between $75-200 per loan.
For a lender originating 5,000 loans monthly, that’s north of $1 million annually in QC costs—before accounting for defect remediation, repurchase reserves, and investor penalties.
The defects driving these costs are predictable: income calculation errors, asset documentation gaps, appraisal deficiencies, missing disclosures, and AUS data mismatches. Predictable means automatable.
Where AI Fits
AI systems can now parse unstructured documents and validate extracted data against guideline logic programmatically. This shifts QC from document-by-document human review to exception-based workflows where analysts focus on judgment calls rather than checkbox verification.
What AI handles well: document classification, calculation verification, tolerance tracking, AUS findings reconciliation, and guideline citation for identified defects. What still requires humans: compensating factor evaluation, complex income scenarios, appraisal disputes, and fraud pattern recognition in ambiguous cases.
Companies like Sei AI have automated these validation workflows end-to-end, processing loans against agency handbooks and state regulations without manual intervention on deterministic checks.
The Competitive Math
Lenders operating with 25% defect rates and $150 per-loan QC costs are competing against operations achieving sub-10% defect rates at $50 per loan. In today’s margin environment, that gap is existential.
Manual QC at scale isn’t just expensive—it’s a constraint on growth. When every incremental loan requires incremental analyst hours, capacity becomes the bottleneck. AI removes that constraint.
The lenders solving this problem now will define the cost structure that the rest of the industry competes against.
(Views expressed in this article do not necessarily reflect policies of the Mortgage Bankers Association, nor do they connote an MBA endorsement of a specific company, product or service. MBA NewsLink welcomes submissions from member firms.)
