← Banking · 02

Pix is irreversible. The agent has fifty milliseconds.

Real-time scoring, step-up authentication, dispute drafting. One loop, sub-50ms, with the human in the escalation path.

Proven pattern

Real-time fraud agent scores incoming/outgoing instant-payment transactions on device, behavioral, and network signals; orchestrates step-up authentication; on confirmed fraud, hands off to a dispute-resolution agent that opens the ticket, assembles evidence, and routes to the correct queue with an SLA-bound response. Pix processes more transactions than any other instant-payment system globally; BCB Normative #491 (effective 2025) requires device-registration controls, real-time intelligence sharing, and adaptive defenses.

Hard KPIs · operator-disclosed where possible
30%+
false-positive reduction at parity recall vs. incumbent rules
50%+
median dispute-resolution cycle-time reduction
<50ms
p99 per-transaction inference latency

Buyer · Head of Fraud + COO + Chief Risk Officer.

Operator references

Provenance tiered explicitly. Audited > Operator-disclosed > Vendor-published. Treat all vendor-tier figures as directional.

Itaú Unibanco (Brazil)
Multi-agent architecture inside Itaú Intelligence; Pix on WhatsApp at 100% of customers, transactions ~20s faster, 90% repeat usage. MIT-CSAIL alliance for fraud research.
Operator-disclosed
Bradesco (Brazil)
BIA Customer at 87% resolution / 25M+ interactions on the Bridge platform; MentorIA collections agent analyzes 18,000 calls/day with 22% conversion lift.
Vendor-published
Nubank (Brazil/México/Colombia)
GPT-4o customer copilot resolves 50–55% of Tier-1 inquiries; 70% chat response-time reduction; 2M+ chats/month. Composite-tool ReAct architecture.
Operator-disclosed
Banrisul · Sicredi · Banco do Brasil
All combining behavioral monitoring + MFA + AI fraud detection + user education for Pix.
Operator-disclosed
Why this beats incumbent rule engines and isolated ML

Rules cannot reason about novel social-engineering scripts. Classifiers cannot draft a regulator-compliant dispute response. The agent loop wins because it orchestrates: real-time scoring + step-up + intervention + dispute drafting + audit. Nubank's composite-tool / ReAct pattern is the published reference.

Why this is hard

Sub-50ms SLAs constrain architecture choices. BCB Normative #491 control-mapping is not optional. Brazilian-Portuguese voice and chat fluency is a real differentiator over global vendors trained on Iberian Spanish or English.

Mexico · LATAM specifics

SPEI is not Pix. Faster than ACH but not as instant or as volume-dense, with a different fraud surface (more authorized-push-payment, less account-takeover at scale). The pattern still extends; the rules and regulator (Banxico) differ. SPEI-CoDi for QR-driven retail flows is the highest-growth fraud surface in Mexico.

Reference architecture

Each node maps to a regulatory anchor.

  1. 01
    Transaction ingress: Pix / SPEI / SPEI-CoDi rail
  2. 02
    Device + behavioral + network signal extraction
  3. 03
    Real-time risk score with policy citation
  4. 04
    Step-up authentication or pass-through decision
  5. 05
    On confirmed fraud → dispute agent assembles evidence pack
  6. 06
    Routing to fraud queue with SLA + customer notification
  7. 07
    Audit log mapped to BCB Normative #491 / Banxico controls
Anti-positioning

Not A chatbot. A "fraud AI." A system that auto-blocks then sends a customer to a help center.

But Production infrastructure on the rail: terminals, APIs, observability, audit logs, human escalation by SLA.

What didn't work initially

The Klarna reversal is the design constraint. Early deployments fail when borderline cases are auto-blocked without an SLA-bound human path; CSAT collapses, the regulator reads the social-media complaints. The handoff to a human is itself a first-class metric.

A working session, not a sales call.

Two hours with a partner. Your incumbent stack, your data posture, and the regulatory surface against a sovereign reference architecture for banking.

Request a working session →

A working session, not a sales call.

Two hours with a partner. We map your AI spend, data exposure, and governance posture against a sovereign reference architecture. You leave with a memo. We leave with a decision.

By invitation.