The Agentic Bank

Real-Time Fraud Agent

⬡ Reflex Scores and intervenes on transactions in-flight.
◆ Always-on Specialist

Sits on the authorization path and decides, in milliseconds, whether to approve, step-up, or block, balancing fraud loss against false declines of legitimate spend. Emits a plain-language rationale for every block, retained for audit.

Memory

Working The transaction + the cardholder's live session signals.
Episodic The cardholder's spend pattern and prior fraud events.
Semantic Current fraud typologies and scam scripts.
Store Feature store + behavioural profiles

Orchestration

MCP

Harness · Managed Agents … ultra-low-latency loop; async rationale generation.

Tools

{ } Authorization stream API { } Device + behavioural signals API { } Customer step-up (push/SMS) API

Evals & guardrails

  • Continuous false-decline rate monitoring … over-blocking is a tracked harm.
  • Champion/challenger on every model update; shadow-scored before promotion.
  • Fairness checks across customer segments.

Frontier edge

  • On-device/edge inference: the scoring head runs at the authorization node in single-digit milliseconds, so the approve/step-up/block call never leaves the payment edge.
  • Counterfactual decisioning: scores 'is this loss-prevention worth declining a real customer at the till?' against a live elasticity model, not a static rule, to hold false-declines down.
  • Self-improving fleet: confirmed scam scripts learned at one edge node propagate to the whole population overnight via an eval-gated update, never a hot self-edit on the auth path.

In numbers

8,500
Auth decisions / sec
−34%
Fraud loss vs. baseline
0.6%
False-decline rate

Handoffs

More on the Fraud Detection desk