The OCC just showed its hand on AI
The OCC’s AI framework offers key lessons for banks, asset managers, pensions, and insurers.
On Nov. 4, the Office of the Comptroller of the Currency (OCC) released a 22-page framework showing how its examiners will evaluate AI in banks. A footnote discloses that the document was prepared with AI assistance, then reviewed, edited, and approved by a human. That is the message in one line: use the technology, and keep a person accountable for it.
Two facts frame everything that follows. First, this is an examiner resource, not a rule; the OCC says it sets no new expectations, and banks are not expected to use it to assess themselves. Second, it is aimed mainly at generative and agentic AI, the systems that the April 2026 model risk guidance left out of scope.
A note on language: following a recent executive order, the OCC now calls these technologies “Super Intelligence.” We use “AI” here because that is still what your teams call it.
This is Part One of two. It covers what the framework says and what you can do about it now. Part Two tests those points against a real, widely reported incident.
Five AI governance takeaways for people who own the risk
1. It is not a rule, but it is the OCC’s exam script
The framework is written almost entirely in “may” and “could.” Do not let the soft verbs mislead you. It sets out 16 components across six pillars: purpose, risk, control, oversight, resilience, and assurance. Examiners now have a common structure for scoping reviews and asking questions.
In practice: If you lead internal audit, read it as a preview of the questions and document requests you may see in your next exam cycle.
2. What the system can do matters more than how smart it is
This is the framework’s central idea. Risk classification may rest on potential consequences, the transfer of agency, and the degree of human oversight, not on technical sophistication. The OCC’s own example: a simple rules-based agent that can initiate transactions could present more risk than a sophisticated model limited to internal drafting.
The framework illustrates this with four tiers. The top tier includes autonomous payment agents, real-time credit decisioning, and fraud-hold release. The second tier is where many teams will be surprised. It includes AML alert disposition, agent-initiated customer communications, and coding assistants that change internal code or data access.
In practice: If you lead enterprise risk, re-rank your AI inventory by authority to act and by whether a mistake can be undone. Expect a few tools you considered routine to move up the list.
3. AI lives inside the risks you already manage
Examiners are told not to treat AI as a separate risk category, and banks are not required to define a separate AI risk appetite. Instead, AI can introduce, amplify, or modify operational, compliance, and credit risk. That is good news; you do not need a new risk taxonomy.
There is a catch. The same document says boards and senior management may establish an AI risk appetite that addresses autonomy, permitted actions, and customer impact. And because the April model risk guidance left generative and agentic AI out of scope, someone still has to own them before an examiner asks who does.
In practice: If you are the chief risk officer, be ready to answer one question plainly: how much autonomy are we comfortable with, and who decided?
4. AI controls on paper are not controls
Three examples from the text make the point.
-
- Purpose. Activity outside a system’s approved purpose may be treated as a control failure, even when the result looks reasonable.
-
- Human review. Approval should be more than a nominal click. The framework asks how many agents, alerts, or actions one person can credibly monitor.
-
- Agent access. Agent identities and credentials may be governed in a manner comparable to privileged users.
In practice: If you support the CISO or lead data governance, treat every agent like a privileged account with a written job description: what it may touch, what it may do, and when it must stop.
5. Plan for “working but wrong,” and be ready to prove it
Traditional resilience planning asks what happens when a system goes down. The framework adds a harder question: what happens when the system stays up and does the wrong thing? Its answers include kill switches, tested fallback processes, and a separate workstream to find and make whole every affected customer.
Then comes proof. Records may need to support the reconstruction of a material decision. Internal audit should be independent of both the team that builds a system and the team that validates it, with access to underlying evidence and not only management summaries. And when a vendor builds or hosts the tool, accountability generally remains with the bank.
In practice: If you lead internal audit, add the backstops to your plan. Has the kill switch been tested, by whom, and when?
Where your existing frameworks fit
A certificate is not an exam answer. The OCC framework names no control standard; it borrows vocabulary from Treasury’s AI Lexicon and stops there. The mapping below is ours, not the regulator’s.
| Framework | What it gives you | What it leaves open |
| Treasury Financial Services AI RMF | Roughly 230 control objectives written for financial services, built on the NIST structure | Voluntary; you still must show the controls operate for each high-consequence use case |
| NIST AI RMF and its generative AI profile | A common structure to govern, map, measure, and manage AI risk | Not certifiable, and not specific to banking |
| ISO/IEC 42001 | A certifiable AI management system: roles, lifecycle, impact, risk, and applicability assessment, internal audit | Shows the management system exists; does not show that a given agent’s permissions are right |
| SOC 2 and ISO 27001 | Security assurance over a vendor’s environment | Little on AI behavior, such as hallucination, drift, or what an agent may do inside your systems |
| AIUC-1 | Provides independent technical assurance over the security, privacy, safety, reliability and governance of AI agent | Should sit within a broader AI governance, model risk management, and regulatory compliance framework |
Think of these frameworks as the scaffolding and the OCC’s questions as the test. A management system certification tells an examiner you have a program. The framework asks something narrower: what can this specific system do, what happens when it is wrong, and can you prove the controls work?
The same logic applies to vendors. A vendor’s ISO/IEC 42001 certificate is useful diligence evidence. It does not transfer accountability, and it does not settle what that vendor’s agents may do inside your environment.
Three things you can do today that will make your AI governance better tomorrow
None of these needs a budget cycle. Each one produces evidence you can show.
-
- Count and rank. List every AI use, including features your vendors have added to products you already own and tools employees adopted on their own. For each material use, write one paragraph: what it may do, what it may not do, and when a person must step in. Then rank the list by authority to act, not by sophistication.
- Pull the plug on purpose. Pick your highest-ranked use case and run a short exercise. Can you stop it? Can the business keep running on a manual process? Can you identify every customer it touched? While you are there, check how many approvals each human reviewer clears in an hour.
- Read three contracts. Take your three most important AI vendors and look for four things: limits on what the vendor’s agents may do inside your systems, a definition of an AI incident with a notice period, who bears the cost of errors and customer harm, and advance notice when the vendor adds or changes AI features.
If you are not a bank, read this anyway
Risk practices that start in bank supervision have a way of traveling. Model risk management and third-party oversight both began as banking disciplines and became common practice well beyond it. Asset managers, pension organizations, and insurers should pay particular attention. While the OCC framework applies to banks, its core questions around AI authority, human oversight, resilience, third-party risk, and accountability are relevant across financial services.
Banking is not the first sector to see AI rules; the European Union and several states moved earlier. What is new here is a detailed examiner playbook for AI that acts on its own. And nothing in its core questions depends on a bank charter. What may the system do? What happens when it is wrong? Who can stop it? Who is accountable? Can you prove it?
For banks, asset managers, pension organizations, and insurers, those questions increasingly form a common AI governance challenge. If your organization is deploying AI agents, your auditors, insurers, board, and regulators are likely to ask the same five questions. Banks simply received them in writing first
Related services
Our solutions are tailored to each client’s strategic business drivers, technologies, corporate structure, and culture.