The lender did not need a chat interface. It needed better judgment embedded into daily routing: which leads to prioritize, which loans to move, and where throughput was being lost.
High-volume lead pools made small ranking errors expensive because sales capacity was fixed.
Operations teams needed recommendations that improved conversion without starving downstream teams or creating queue imbalances.
The model had to work inside a regulated environment where explainability, auditability, and operational control mattered.
Propensity ranking
We trained a selection model that scored likelihood of conversion using historical lead, borrower, channel, and interaction signals, then calibrated the output for operator review.
Throughput optimizer
A second optimization layer modeled the loan flow as a constrained operation, identifying which queues and handoffs had the largest effect on crossing time.
Human-in-the-loop controls
The system shipped with cutoffs, review bands, override reasons, and monitoring so managers could tune capacity without losing the model's signal.
- Benchmarked the model against the existing lead selection process before changing routing behavior.
- Introduced recommendations in a review workflow so managers could see the score, rationale, and operational effect.
- Extended the work from lead ranking into loan-throughput optimization after the first model proved lift.
Propensity lift improved by roughly fifty percent against the prior baseline.
Qualified lead volume nearly doubled because the team spent capacity on better-ranked opportunities.
The optimizer reduced crossing time from thirty-one to twenty-nine days by focusing on operational constraints, not generic automation.
Not every AI win is an agent. In high-volume operations, a well-calibrated ranking model and a sober queue optimizer can create more value than a flashy interface.