Risk Agents in Banks: The Bottleneck Was Never the Model

When I talk to technology and compliance teams at fintechs and payment institutions in Brazil, the conversation almost always starts in the same place: which model to use, which agent architecture makes sense, how to orchestrate decisions in real time.

That is the wrong starting point.

Not because the model does not matter. It does. But what will block any Risk Agent deployment in production, especially inside a Brazilian financial institution, is not the quality of the reasoning. It is the quality of the data that feeds that reasoning. More precisely, it is whether that data meets the standards the Banco Central do Brasil expects when it reviews automated decisions.

Febraban Tech 2026 runs August 24 to 26 at Anhembi in São Paulo under the theme "Intelligent Agents, Human Leadership." Dozens of demonstrations will show agents making autonomous decisions: detecting fraud, approving credit, monitoring transactions in real time. What will rarely appear in those demonstrations is what makes those decisions auditable by a Brazilian regulator. That is the infrastructure gap that matters.

What a Risk Agent Actually Needs#

A risk agent operating in payment flows over PIX, or in merchant onboarding for a subacquirer or payment service provider, needs five specific inputs for each decision it makes.

Continuous identity validation. KYC is not an onboarding event. It is a state. The CPF that was clean at registration may have a restriction today. The document validated six months ago may have been revoked. The beneficial ownership structure that was KYB-cleared at contract time may have changed. An agent that decides using stale identity data is not making a real regulatory decision. It is processing a snapshot that may no longer reflect reality.

Transactional history with behavioral context. Not just whether a CPF or CNPJ exists in a database. The behavioral pattern: typical transaction volume, operating frequency, counterparty network, usual hours and channels. Without this context, the model generates alerts based on absolute thresholds rather than deviations from an individual's own pattern, and the false positive rate makes the operation unmanageable for any compliance team at scale.

Real-time regulatory list screening. COAF registers, OFAC lists, PEP databases, international sanctions. A false negative here is not a poor metric. It is a regulatory liability. Under Lei 9.613/1998, the base AML law in Brazil, and under BCB Circular 3.978/2020, financial institutions are required to maintain continuous monitoring of suspicious activities and report them to COAF within defined timeframes. An agent querying cached data introduces a window of regulatory exposure that cannot be justified in an inspection.

Bot detection and device integrity signals. In high-volume digital flows, whether merchant onboarding batches or retail PIX transactions, the difference between a human applicant and a malicious automated process needs to be detectable before the decision runs. Bot activity that reaches the decision layer has already bypassed the control point where it was cheapest to stop it.

A complete decision audit trail. Every block, every approval, every alert needs an associated record that includes: the data source queried, the source's last update timestamp, the rule version applied, the confidence output of the agent's assessment, and the final decision with its timestamp. The Banco Central do Brasil does not ask which model you use in an inspection. It asks: where did this information come from? Who validated it? When was it updated?

This is not a wishlist. This is the minimum viable compliance stack for operating regulated financial flows in Brazil.

The Regulatory Layer That Cannot Be Skipped#

Brazil's AML and KYC framework is built across multiple regulatory layers that interact with each other in ways that matter for automated systems.

At the base, Lei 9.613/1998 establishes the anti-money laundering obligation and defines COAF, the Conselho de Controle de Atividades Financeiras, as the financial intelligence unit responsible for receiving, analyzing, and disseminating suspicious transaction reports from obligated entities. Financial institutions, subacquirers, and payment institutions operating under BCB authorization are all obligated parties under this law.

BCB Circular 3.978/2020 operationalized the monitoring requirements: regulated institutions must implement policies, procedures, and internal controls for the prevention of money laundering and terrorist financing, maintain records for at least five years, and communicate suspicious operations to COAF within defined timeframes. The circular is explicit that monitoring systems must be calibrated to the risk profile of the institution's customer base, not applied uniformly.

The BCB's 2025-2026 regulatory agenda explicitly includes artificial intelligence governance, according to Chambers and Partners, including guidelines for the ethical and secure use of AI within regulated institutions and active monitoring of domestic legislative initiatives related to AI. The Banco Central also created a Center of Excellence in Data Science and Artificial Intelligence specifically to develop these governance guidelines.

This regulatory context has a direct operational implication for any institution deploying Risk Agents: every automated decision that affects a regulated flow, whether approving a merchant, flagging a transaction, or blocking a withdrawal, is subject to the same documentation standards that apply to human decisions. The agent is not exempt from traceability requirements because it is automated. In many respects, the traceability expectations are higher, because the volume of automated decisions makes individual review impossible, and the regulator needs to be able to sample any decision and reconstruct it from the record.

The institutions that will integrate Risk Agents successfully are not the ones with the most sophisticated models. They are the ones that can demonstrate, in an inspection, exactly what data the agent used, where that data came from, and how current it was at the moment of the decision.

When the Data Layer Is Missing: What Breaks in Production#

Working with subacquirers and payment service providers in Brazil that process significant volumes, the failure pattern when a Risk Agent is deployed without adequate data infrastructure is consistent: it performs well in test environments where data is controlled and sanitized. It breaks in production, where the data reflects the actual market.

One scenario that recurs: the agent blocks a legitimate transaction because the CPF of the account holder matches an entry on a sanctions list by name, a homonymy false positive. The compliance analyst has to review it manually. When this happens at low volume, it is manageable. When a subacquirer is processing thousands of transactions daily across a merchant network, the manual review queue grows faster than the team can clear it. The bottleneck was not the agent's decision logic. It was the absence of automatic false positive resolution at the data layer, using secondary identifiers that were already in the registry and would have distinguished the sanctioned entity from the customer.

A second scenario: the agent approves an onboarding that should have been flagged because the regulatory list query ran against cached data that was 48 hours stale. The event that triggered the restriction had occurred in that window. The data latency was the failure, not the model's reasoning. The model reasoned correctly from the data it received. The data was wrong.

"Human leadership" over risk agents cannot function when the data layer does not give humans something reliable to review. A supervisor looking at a flagged transaction needs to see the data state at the time of the decision, not the current state after the agent already acted. When that audit trail does not exist, the supervisor cannot evaluate whether the decision was correct. They can only observe the outcome.

This is why the market is beginning to price the data and governance layer separately from the model. Quantifind, a risk intelligence platform focused on AML, KYC, and sanctions screening with explainability as a core capability, raised $200 million in June 2026. FIS, one of the largest financial technology providers globally, announced a partnership with Anthropic to build a Financial Crimes AI Agent. Neither of these investments is primarily a bet on model quality. They are bets on the infrastructure that makes model outputs actionable and defensible in regulated environments.

What "Human Leadership" Really Means for Risk Agents#

The Febraban Tech 2026 theme, "Intelligent Agents, Human Leadership," is operationally precise in a way that gets obscured in the broader conversation about AI autonomy.

Human leadership over an agent system does not mean a human approves every individual decision. At the volumes that make automation worthwhile, that is operationally impossible and eliminates the efficiency the agent was deployed to create. Human leadership means the system is designed so that when a human does intervene, whether in a dispute, an inspection, an edge case, or an escalation, they can understand exactly what happened and why, and they can act on that understanding.

For a risk agent in a regulated environment, that requires three operational properties.

Decision reconstructibility: the ability to reproduce any past decision exactly, with the data state that existed at the time of the decision, not the current state. A compliance officer responding to a COAF inquiry cannot work from the current data. They need to show what the agent knew when it acted.

Documented escalation logic: the agent needs to know not just what to decide but when not to decide. When to route to a human analyst, what information to surface in that routing, and at what confidence threshold the decision requires human confirmation rather than automated execution. This is not a limitation of the agent. It is the mechanism by which human leadership is operationalized.

Consistency between automated and human decisions: an agent that applies criteria inconsistently, blocking similar transactions at different rates based on factors that cannot be documented, fails regulatory review. Consistency means the same decision criteria, verifiable from the audit record, regardless of whether the decision was made by the agent or by an analyst following the agent's rules.

None of this is achievable if the data feeding the agent is stale, inconsistent, or unverifiable at query time. The governance structure starts at the data layer.

Infrastructure in Production: What Gu1 Brings to Febraban Tech 2026#

At Balcão 25 in the Fintech Lounge at Febraban Tech, August 24 to 26, we will demonstrate Risk Agents operating on the infrastructure described above.

Real-time KYC, integrated with the Brazilian regulatory ecosystem through primary data sources, with identity state that updates continuously rather than at onboarding only. AML monitoring with rules calibrated to Brazilian market risk profiles, with alert workflows connected to COAF reporting requirements. Bot detection integrated into high-volume merchant onboarding flows. And on each decision, a complete audit trail: queryable, exportable, auditable by any party that needs to review it.

This is not a pilot environment. Gu1 operates in production with 34 financial institutions across 6 countries in Latin America, processing over 20 million transactions monthly for a single Brazilian client. The platform holds ISO 27001, SOC 2, GDPR, and PCI DSS certifications and operates under a contractual SLA of 99.5% availability. The demo runs on the same infrastructure.

What the demonstration will show is not the sophistication of the underlying model. It will show what a Risk Agent can actually do when the data it needs is available, current, and auditable: KYC state that reflects today's registry, not last quarter's snapshot. AML alerts built on behavioral context, not only absolute thresholds. Regulatory list queries that run in real time against live sources, not cached data. And on every decision, the trail that a BCB inspection or a COAF inquiry can follow.

If you are attending Febraban Tech and working in compliance, fraud prevention, or merchant onboarding at a financial institution in Brazil, the team is at Balcão 25 from August 24 to 26. The conversation will be operational.

---

Aryanne Reis leads Gu1's operations in Brazil, covering the full cycle from commercial engagement through compliance onboarding and day-to-day risk operations with subacquirers, PSPs, and fintechs.