The Agentic Bank

Auto Decisioning Agent

⬡ Ignition Decisions auto-loan applications from dealer and direct channels.
◆ Supervised Specialist

Scores the borrower and the collateral together … credit, affordability, the vehicle's loan-to-value … and returns a tiered decision and rate to the dealer portal in seconds. Commits approvals directly; stages declines with reason codes for an oversight agent to re-derive before the adverse-action notice.

Memory

Working The application, bureau, collateral value, and pricing in progress.
Episodic The dealer's prior application quality; the borrower's history.
Semantic Auto credit policy, LTV + advance rules, ECOA/Reg B, rate sheets.
Store Feature store + decision ledger

Orchestration

pipeline MCPA2A

Harness · Managed Agents … fast session; decision rationale captured for the record.

Tools

{ } Credit bureau pull API { } Vehicle valuation feed API ›_ Pricing + policy engine Code exec { } Dealer portal (decision return) API Adverse-action staging A2A

Evals & guardrails

  • Fair-lending guardrail: disparate-impact testing across segments + dealers.
  • Declines NEVER auto-commit … staged with reason codes for an oversight agent to re-derive and sign off.
  • Dealer-fraud watch: anomalous application patterns route to Financial Crime.
  • Champion/challenger before pricing or scorecard changes go live.

Offline reflection

Replays funded-loan performance by dealer + tier to spot drift in approval quality and flag dealers whose books underperform for an oversight agent to review.

Frontier edge

  • World-model simulation: simulates the deal's funding and likely-default path before returning a tier, pricing the collateral risk into the dealer decision in real time.
  • Causal reasoning: counterfactual LTV / affordability analysis, so a strong borrower on a depreciating vehicle is priced for what drives the risk.
  • Continual learning (eval-gated): online updates to dealer-quality signals from funded-loan performance, gated by disparate-impact testing before any pricing self-edit ships.

A sample run

Trigger Dealer submits a Saturday-afternoon application for a used SUV.
  1. 1Pull the bureau; value the vehicle and compute LTV.
  2. 2Run affordability + pricing against policy; land an approval tier.
  3. 3Return the decision and rate to the dealer portal with conditions.
Output Approval returned to the dealer in 90 seconds at a tier-2 rate; one thin-file applicant staged as a decline pending an oversight agent's re-derivation.

In numbers

38,000
Applications decisioned / day
median 84s
Dealer decision time
55%
Straight-through approval

Handoffs

Across ⇢ Financial Crime → Fraud for dealer-channel anomalies

More on the Auto Finance desk