Automated extraction
OCR and NLP read applications, financial statements, underwriting notes, KYC files, contracts, and regulatory filings to extract key financial metrics.
Headline result
~10%
More realized losses concentrated in the riskiest rating band during back-testing.
Operating model
Hybrid
AI creates the quantitative baseline; underwriters retain final approval authority.
Decision controls
Gated
Read accuracy, confidence, policy, and eligibility checks prevent weak inputs from becoming ratings.
Challenge
Manual commercial loan approval was slow, variable, and difficult to scale.
SmartBank assigns every commercial borrower an internal risk rating that drives approve/reject decisions and loan pricing. The legacy process relied heavily on individual underwriters manually interpreting financial statements, disclosures, and policy rules.
Applications could arrive through a call center, in person, or online. From there, underwriters entered application details into internal systems, reviewed financial statements by hand, converted key metrics into predefined bins, and applied policy weights before assigning a final internal rating on a 1-5 scale.
Defaults are rare in a healthy commercial book, often around 1%, so the most valuable risk signal is concentrated in a small subset of failed loans. That made the process especially sensitive to data quality, expert judgment, and consistent interpretation.
Manual data entry
Manual document review
Variable scoring
Headcount-limited scale
Dense disclosures
Slow approval cycles
Before
Traditional approval workflow
01
Applications arrive through in-person, call center, or online channels.
02
Underwriters key application data into internal systems by hand.
03
Financial statements and disclosures are read and interpreted manually.
04
Metrics are placed into predefined bins, weighted by policy, and converted into an internal rating.
05
Lower ratings are approved at different pricing levels; higher ratings are rejected.
Solution
Dual Risk Rating turns documents, metrics, and policy into an auditable approval workflow.
Risk score mapping
A machine learning model predicts risk from extracted metrics and historical loan performance, then maps the output to the bank's 1-5 rating scale across obligor risk and facility risk.
Validation gates
Read accuracy, completeness, model confidence, and policy checks hold or route cases before weak inputs can silently become ratings.
Underwriter review
Underwriters review the AI-generated rating, chat with the application and supporting documents, drill into sources, and apply a qualitative overlay where judgment requires it.
Confidence routing
Low-risk, low-dollar loans can be auto-approved while high-risk, high-value, borderline, or low-confidence cases are escalated for human review.
Model validation
Methodology, assumptions, performance, limitations, versioning, and monitoring are documented for model risk review and audit readiness.
Human + AI clarity
Every model rating can be traced back to the document level.
The underwriter does not receive a black-box score. The review dashboard exposes the extracted values, source documents, page references, confidence levels, policy checks, and exception flags used to produce the rating.
The same idea supports the small-data problem in commercial lending: because default cases are rare, the system learns from curated, context-rich examples rather than depending only on volume. Each review and override becomes feedback for future extraction, scoring, monitoring, and retraining.
Governance
Controls are built into the approval workflow, not added after deployment.
AI risk
Operational control
OCR or extraction errors
Read-accuracy thresholds, data-completeness gates, and manual fallback for low-confidence documents.
Model bias
Fairness testing, prohibited-variable removal, approval-pattern monitoring, and regular audit sampling.
False approvals or rejections
Human final authority, escalation protocols, qualitative overlays, and exception review for borderline cases.
Model drift
Loan performance monitoring, periodic retraining, model versioning, and controlled deployment.
Audit and regulatory review
End-to-end lineage, immutable logs, methodology documentation, and collaboration with model review teams.
Impact
Sharper risk assessment with responsible underwriter control.
In back-testing against historical loan performance, model-derived ratings showed stronger risk separation than manual underwriter ratings. The riskiest model band captured roughly 10% more realized losses than the historical riskiest band.
The result is a faster, more consistent rating process that preserves sound credit judgment, regulatory compliance, and human accountability.
- Sharper loss capture in the riskiest rating band
- Faster approval cycles through automated extraction
- Consistent and repeatable scoring logic
- Higher volume without proportional headcount growth
- Traceability from rating to source documents
- Immutable audit trail for decisions and overrides