◆ Autonomous Monitor
Watches cash-flow patterns per customer cohort and composes a nudge on a signal … an upcoming overdraft, an idle balance that should earn interest, a recurring subscription. Every nudge is frequency-capped, suppressible, and audited for fairness so it never tips into dark-pattern territory.
Memory
Working The triggering signal + the customer's recent cash-flow window.
Episodic Which nudges this customer engaged with or dismissed.
Semantic Product fit rules, suitability constraints, communication-preference policy.
Procedural Learned messaging that converts without annoying, per segment.
Store Feature store + engagement history
Orchestration
swarm MCP
Harness · Managed Agents … event-driven, always-on per customer cohort.
Tools
{ } Transaction + cash-flow analytics API ⌕ Product catalogue + eligibility Retrieval { } Customer comms (push / in-app) API
Evals & guardrails
- UDAAP + fair-treatment review: no manipulative framing, suitability checked per offer.
- Frequency caps and an absolute opt-out the agent must honour.
- Champion/challenger on message variants; engagement vs. complaint rate tracked.
- Fairness checks so nudges aren't skewed across protected segments.
Offline reflection
Offline consolidation of engagement outcomes refines which nudges help (opened, acted on) versus which annoy … pruning the latter from the playbook.
Frontier edge
- ▲Proactive / anticipatory: predicts the next financial pinch (an overdraft three days out, a lapsing CD) and acts before the customer notices.
- ▲Causal reasoning: counterfactual 'would this nudge change the outcome, or just annoy?' analysis, so it only fires when it moves the customer's position.
- ▲World-model simulation: simulates the customer's cash-flow forward under several nudge options to pick the one that helps most without tipping into dark-pattern territory.
In numbers
210,000
Overdrafts pre-empted / month
2.1%
Nudge opt-out rate
Handoffs
Hands to → Customer Service Triage Agent