The OCC’s New AI Framework: What It Means for Your Financial Institution

Fraud

By Jeff Scott

6 Nov, 2026

Breaking down the newest framework on artificial intelligence in financial services and what it means for risk, compliance, fraud and disputes teams. 

If you run fraud at a financial institution and you have agents anywhere near a decision, you need to pay attention. 

On Nov.  4 the Office of the Comptroller of the Currency (OCC) issued Bulletin 2026-54, a supervisory framework for what it calls Super Intelligence  (SI), its term for artificial intelligence and agentic AI. It is written for examiners, not for financial institutions. The OCC says it sets no supervisory expectations, no enforceable standards and no prescriptive requirements. 
 
Read it anyway. 
 
It is the script examiners will use, and it fills a gap that has been open since April, when the revised model risk guidance (Bulletin 2026-13) put generative and agentic AI expressly out of scope and told financial institutions to manage them under “broader risk management” with no specifics.
 
The supervisory framework, which the OCC promises to update regularly, has 16 components grouped into six pillars: purpose, risk, control, oversight, resilience and assurance. It covers topics such as credit decisioning, payments, fraud detection, customer service, operations and risk management. While the framework serves as guidance to the broader banking industry, it also includes a note that "examiners may use the framework for supervising community banks' SI use," while instructing examiners to "tailor their examinations to the bank's size, complexity, and risk profile."
 
Three ideas in the framework change how a fraud team should think about agents.
  1. Risk is assessed by what the agent can do, not how smart it is.
    The framework says classification “may be based on potential consequences, transfer of agency, and the degree of human oversight, rather than the technical sophistication” of the system. Its own example: A simple rules-based agent authorized to initiate transactions can present more risk than a sophisticated model limited to drafting. In its illustrative tiers, fraud-hold release sits at Tier 1 Critical alongside autonomous payment agents and real-time credit decisioning. Anti-money laundering alert disposition and agent-initiated account holder communications are Tier 2. 
  2. The decision is what gets examined, not the model. 
    The documentation section asks whether a financial institution can reconstruct a material decision from “inputs, outputs, context, retrieved sources, tool calls, approvals, overrides, and actions.” Likewise, the third-party section names “opaque provider chains” as a risk and warns that shared providers and infrastructure can create “concentration or correlated risk across otherwise separate use cases.” This means you need one record, even when more than one vendor touched the decision.
  3. Oversight has to be real.
    The human-in-the-loop section says review should be “more than a nominal click,” warns that bulk approval “could reduce review quality and contribute to alert fatigue,” and asks the financial institution to evaluate “how many agents, accounts, processes, alerts, or actions one person can credibly monitor.”

What should you do? 

These ideas change how we assess AI agents in financial services. If you run compliance, risk, fraud operations, treasury management or disputes, here’s what I would do next: 

Put your first agents inside the workflows you already control

The framework defines decision support with a mandatory human checkpoint as Tier 3, and direct action with its own permissions as Tier 1 or 2. An agent that works one case, takes only the actions that queue allows, and hands its result to an analyst passes the bar by inheriting every control you already have. The trigger defines who can invoke it, the case defines the data, the queue defines the actions, the analyst is the human in the loop, and the case record is the audit trail. Boxes checked across the board. That is where the best vendors are building first, and it is the version of agents your examiner can say yes to.

Take an inventory of every moment and action a model or vendor touches

The purpose-and-permitted-use section asks financial institutions to document “what an [AI] system or agent is designed to do, what it is not designed to do, who or what may invoke it, the data and systems it is permitted to access, the actions it could take,” and to treat activity outside that scope as a control failure “even when the resulting output or action appears reasonable.” You own every hold, pause, disabled login, and dispute outcome, and your partners should be able to tell you exactly where their decisions land.

Ask every vendor where the evidence lives when it triggers an action in your environment. 

The third-party section tells financial institutions to define “the scope of authority, identity, permitted tool use, data use and escalation limits for provider agents that interact with bank systems and data.” The right answer is in your case record, next to the action and the outcome. A vendor who has built for that transparency will welcome the question.

Underneath all of this is one design principle: The alert, the action and the record belong in one place, whoever you work with. Fraud intelligence that is connected from detection, action, resolution, and learning gives you one story to tell an examiner. A fraud strategy spread across disconnected tools means telling it from scratch every time.

We're here to help

That is why Q2 is uniquely positioned to help. We natively own the tools across detect, act, resolve and learn, in one connected fraud intelligence platform, so the system is designed to provide the decision trail an examiner will ask for. 

Financial institutions that design for this now, and vendors that build for it now, will show up to the next exam with answers instead of a remediation plan.