The dispute agent a Chief Risk Officer would actually approve.
An autonomous payment-dispute & chargeback resolution system built on Lyzr - six specialized agents that resolve a dispute end-to-end, know when to stop, and leave a replayable audit trail a Chief Risk Officer would approve.
Risk-score is the visible fork: clear the 0.95 gate and it settles & files itself; miss it (or trip a ceiling / sanctions rule) and it stops for a human. Every step is sealed to a hash-chained audit ledger.
Disputes are where agents win the argument outright
RPA breaks
Robotic automation shatters the moment a screen or document format changes - and it can't reason about ambiguous evidence.
Chatbots can't act
A conversational bot can advise a customer, but it can't pull records across five systems, score fraud, or file a chargeback.
Agents reason, act, and escalate
A dispute is irreducibly multi-step, multi-system, deadline-bound, and human-in-the-loop - the textbook agentic workflow.
Six specialized agents, one orchestrator, a hash-chained audit
Everyone can demo a dispute bot. This one ships to production.
Glass-box scoring
Every decision shows the confidence factors that built it - inspectable, tunable, defensible.
Tiered autonomy
Act / brief-a-human / mandatory sign-off. The agent knows when to stop - at 0.95, at the ceiling, on a sanctions hit.
Hash-chained, replayable audit
Every action, tool call, model, and guardrail event logged and sealed with a SHA-256 hash chain - replay (and tamper-test) any decision for a Reg E / SR 11-7 / RBI examiner.
Data sovereignty
Runs in the bank's own VPC with zero data egress, native PII redaction and groundedness.
Watch a dispute resolve - and watch it stop.
Run a live case through six agents, see it auto-resolve at 0.95 confidence, then watch a $1,310 friendly-fraud case correctly halt to a human - with a hash-chained, replayable audit trail.
Open the live consoleSynthetic concept demo. The engine below executes for real; connectors & the dispute book are simulated - every console tool call is badged exec or sim accordingly.
- Decision engine + 0.95 autonomy gate (deterministic factors)
- Filing rule engine: network windows, 3DS liability shift, small-dollar file floor, Reg E clock
- Network-correct reason codes (Visa VCR vs Mastercard 48xx)
- Idempotent GL journal (localStorage state write)
- SHA-256 hash-chained audit ledger (verify / tamper-test)
- LLM narration via OpenRouter when a key is set (labeled)
- Core banking, CRM, OFAC, fraud-feed, VROL/Mastercom connectors (results seeded)
- The dispute book - synthetic, ~3:1 auto/escalate mix authored to exercise every outcome
- Cost model ($37 manual / auto / escalated) + handle-time latency budget
- “Pilot to production” timeframe is an estimate