.png)
A mid-sized lender processing 35,000 applications monthly cut underwriting time from 4.2 hours to 14 minutes with a production AI scoring engine.
.png)
A mid-sized US consumer lender processing 35,000 loan applications per month was running underwriting through a rules-based scoring system built in 2016 and a team of 24 credit analysts. Average decision time was 4.2 hours. With a 40% volume growth target set for the next 12 months and no appetite to scale the analyst team proportionally, the business needed a fundamentally different approach.
Verttx conducted an AI Readiness Audit, built a gradient boosting predictive risk scoring engine trained on 2.24 million historical applications, and integrated it directly into the existing loan origination system. The system went live 11 weeks after the initial discovery call and now processes 92% of applications automatically.
The rules engine was flagging 71% of applications for manual review — not because they were genuinely high-risk, but because 47 fixed threshold conditions lacked the nuance to distinguish borderline candidates from credit-impaired ones. Analysts were reviewing volume rather than complexity, averaging 22 minutes per application on work a well-calibrated model could handle in milliseconds.
Decision consistency was the more serious problem. Approval rates varied by 18.3 percentage points between analysts reviewing comparable risk profiles — a pattern flagged in two consecutive compliance audits and creating regulatory exposure under ECOA's disparate treatment provisions. The 40% growth target made the urgency concrete. More volume was arriving whether the process was ready or not.
Verttx conducted a four-week AI Readiness Audit before agreeing any development scope. The audit confirmed 2.24 million usable historical applications with complete repayment outcomes, and surfaced three data quality issues — including a silent break in debt-to-income calculation methodology from a 2019 schema change — that were corrected before model training began.
We selected XGBoost over neural network alternatives for its accuracy on tabular credit data, training stability across three product lines with different risk profiles, and SHAP explainability outputs required for FCRA-compliant adverse action notices. The model trained on 312 engineered features across credit bureau signals, behavioural application patterns, product-specific risk indicators, and macroeconomic context variables — achieving a Gini coefficient of 0.71 on the held-out test set, 34 points above the previous rules engine.
A real-time scoring API returns a risk tier, probability of default, and FCRA-compliant adverse action reasons within 340 milliseconds per application — connected directly to the existing loan origination platform with no core system changes. A three-week shadow deployment processed 91,400 live applications before cutover. A monthly fairness monitoring pipeline using the BISG methodology runs in production, producing a compliance-ready report covering disparate impact ratios across protected characteristics.
Within the first 30 days, 92% of applications were decided automatically. The analyst team moved from reviewing 24,850 applications per month to 2,800 genuinely complex cases — the work that actually requires human judgement. Average decision time fell from 4.2 hours to 14 minutes. Decision variance fell from 18.3% to 2.1%, resolving the ECOA exposure that had appeared in two consecutive audits.
Six months after launch, monthly volume had grown from 35,000 to 49,200 without a single additional hire. Cost per underwriting decision fell from $18.40 to $3.87 — a reduction of 79%.
The full model codebase, feature pipeline, scoring API, fairness monitoring infrastructure, and all training documentation were transferred to the lender's engineering team at handover, no proprietary framework, no retained access, no dependency on Verttx to keep the system running.
We had been told for two years that our data was not good enough to support a model like this. Verttx did the audit, proved it was, and had a production system live in eleven weeks. The consistency problem that appeared in two compliance audits is gone. — Chief Risk Officer, US Consumer Lending Company
.png)
Underwriting time dropped from 4.2 hours to 14 minutes. Decision variance fell from 18.3% to 2.1%, eliminating a compliance exposure that had been flagged in two consecutive internal audits. The system automated 92% of decisions from day one, and the lender scaled monthly application volume from 35,000 to 49,200 in six months without adding headcount. Cost per decision fell from $18.40 to $3.87 — a reduction of 79%.
.png)
.png)
