The Agentic Bank

Card Underwriting Agent

⬡ Ledger Line Decisions credit-card applications and assigns lines.
◆ Supervised Specialist

Pulls the bureau, scores affordability, applies the credit policy and decides approve / decline / counter with a credit line. Commits clean approvals directly; stages every decline with reason codes for an independent oversight agent to re-derive before the adverse-action notice issues.

Memory

Working The application, bureau pull, and affordability calc in progress.
Episodic The applicant's existing relationship and prior applications.
Semantic Credit policy, ECOA/Reg B rules, line-assignment matrix.
Store Feature store + decision ledger

Orchestration

pipeline MCPA2A

Harness · Managed Agents … session per application; structured note-taking captures the decision rationale for the adverse-action record.

Tools

{ } Credit bureau pull API ›_ Affordability / income model Code exec { } Credit policy engine API Adverse-action staging A2A

Evals & guardrails

  • Fair-lending guardrail: disparate-impact testing on every model + policy change.
  • Declines NEVER auto-commit … staged with reason codes for an oversight agent to re-derive and sign off.
  • Reason-code completeness check so every decision is explainable (Reg B).
  • Champion/challenger before any scorecard goes live; immutable decision ledger.

Offline reflection

Replays funded accounts' later performance against the decision to detect policy drift … surfacing where the cut-off may be mis-calibrated for model-risk review.

Frontier edge

  • Causal reasoning: counterfactual affordability analysis ('what if income shifts, or this line is halved?') instead of a flat correlational score, sharpening line assignment.
  • Formal action-gating: declines are cryptographically staged and replayable, so the system provably cannot issue an adverse action without an independent oversight agent re-deriving the decision the law requires it to justify.
  • Continual learning (eval-gated): consolidates funded-account performance into the scorecard offline, gated by disparate-impact tests before any self-edit goes live.

A sample run

Trigger New credit-card application from an existing checking customer.
  1. 1Pull the bureau; compute affordability against verified deposit cash-flow.
  2. 2Apply credit policy; land an approve recommendation with a line.
  3. 3Assemble reason codes and the decision rationale for the record.
Output Auto-approves with a $7,500 line; a near-miss applicant is staged as a decline with reason codes, held for an oversight agent to re-derive and issue the adverse-action notice.

In numbers

44,000
Applications decisioned / day
61%
Straight-through approval rate
47s
Avg. decision time

Handoffs

Across ⇢ Risk → Model Risk Management for scorecard validation

More on the Card Services desk