In This Article
- What the OCC super intelligence framework says
- The four AI risk tiers
- Where Microsoft 365 Copilot sits
- Why agents change the question
- What examiners will look for
- Matching each tier to Microsoft 365 controls
- Banks, credit unions and mortgage companies
- How ABT puts the tiers to work
- Frequently Asked Questions
The OCC's new examiner framework for AI, issued November 4, 2026, opens with the upside. It says AI and AI agents "can facilitate banks serving customers faster and more consistently," catch more fraud, and free skilled staff for higher-value work. Then it gives examiners a way to sort every AI use at a bank into one of four risk tiers.
That sorting is the useful part of the OCC super intelligence framework for any financial institution running Microsoft 365. A Copilot answering an employee's policy question and an agent releasing a fraud hold are both "AI," and the OCC treats them very differently. Knowing which tier each of your uses falls into tells you where to spend control effort, and where you can let people move fast.
This is the first article in our series on AI agent security and shadow AI for credit unions, banks, and mortgage companies. ABT manages Microsoft 365 tenants for more than 750 financial institutions, and the series maps each risk to the settings we configure in those tenants. This article covers what the framework says, how the tiers work, and which Microsoft 365 controls support each tier.
What the OCC Super Intelligence Framework Says
On November 4, 2026, the OCC published Bulletin 2026-54, "Super Intelligence: Supervisory Framework", addressed to national banks, federal savings associations, and federal branches and agencies. The framework document itself is written for examiners.
"This supervisory framework does not establish supervisory expectations for banks. It is intended as a resource to assist examiners in their supervisory activities."
So the framework sets no new rules. It tells you how an examiner is likely to think. The bulletin says examiners should tailor their work "based on the materiality of the SI use case," and its note for community banks says examinations should fit "the bank's size, complexity, and risk profile." The same bulletin says examiners "should not treat SI as a separate and distinct risk category." AI risk gets folded into the model risk, technology, compliance and operational risk work examiners already do.
The term "super intelligence" needs a word of explanation. Following Executive Order 14434, as the bulletin's footnote explains, the OCC uses "Super Intelligence" for the technologies "previously encompassed by the term 'Artificial Intelligence'" in federal law. It is the same AI your staff already use. The rest of this article says AI, and keeps the OCC's "SI" only inside quotes.
The framework defines the four terms that carry the most weight, the first two in its glossary and the two oversight terms in section 2.3:
Super Intelligence (SI)
A machine-based system that, for human-defined objectives, "can make predictions, recommendations, or decisions that influence real or virtual environments." In practice, any AI system.
Agentic SI
AI "capable of independently making decisions, interacting with their environment, and optimizing processes without direct human intervention." An agent acts, where a chatbot answers.
Human-in-the-loop
"Review and approval before an output or action is used or executed." A person approves each action.
Human-on-the-loop
"Active monitoring with authority and practical ability to intervene." The system runs, and a qualified person watches and can halt it.
The framework also fills a gap. In April the agencies revised their model risk guidance, and the OCC's announcement said generative and agentic AI models "are not within the scope of this guidance." The new framework says it "is more applicable to generative and agentic SI," which puts Copilot, chat assistants and agents squarely in view. We covered the Treasury's AI risk framework in our guide to the Treasury AI risk framework.
The Four AI Risk Tiers
The OCC super intelligence framework organizes its guidance into sixteen components across six pillars: purpose, risk, control, oversight, resilience and assurance. The part most institutions will use first is section 3, an illustrative method that scores each AI use against seven factors and rolls them into a composite tier.
- Decision consequence: does the system inform a person, decide, or act directly?
- Reversibility: can a wrong output be caught and undone, or is it final once executed, like "a payment sent, a disclosure made, an account closed"?
- Customer and financial impact: one transaction or a whole portfolio.
- Autonomy level: human-in-the-loop, human-on-the-loop, or fully autonomous.
- Data sensitivity: regulated, confidential or personally identifiable data.
- Explainability: can the output be traced to a reason good enough for an adverse action notice?
- Provenance: built in-house, fine-tuned from a third-party model, or bought as a vendor service.
The bulletin sums up the logic: classification may rest on "potential consequences, transfer of agency, and the degree of human oversight rather than the technical sophistication of the SI used." The framework's own example makes it concrete: "a simple internal productivity tool with no customer or financial data exposure sits at Tier 4, whereas a fully autonomous, customer-facing payment agent sits at Tier 1, regardless of how well it performs in testing."
Section 3.2 then lays out four illustrative tiers. The examples in the right-hand column are the OCC's own.
| Tier | Profile (OCC, illustrative) | Example uses (OCC, illustrative) |
|---|---|---|
| Tier 1: Critical | Direct action, low or no reversibility, high financial or customer impact, autonomous or partially autonomous | Autonomous payment agent; real-time credit decisioning; fraud-hold release; market access |
| Tier 2: High | Direct action or high-weight decision support, human-on-the-loop, moderate-to-high impact | AI-assisted underwriting; agent-initiated customer communications; AML alert disposition; code assistance for internal code or data access changes |
| Tier 3: Moderate | Decision support with mandatory human-in-the-loop, limited data sensitivity, reversible | Internal document summarization for credit memos; agent-assisted research |
| Tier 4: Low | No customer or financial data exposure, no direct action, fully human-controlled use | Internal productivity tools; code assistance; general employee Q&A |
Notice that code assistance appears twice. Plain code assistance is Tier 4. Code assistance that changes internal code or data access is Tier 2.
Same tool, different consequence, different tier. That is the whole framework in two rows.
Where Microsoft 365 Copilot Sits
Most institutions start with Copilot answering questions, drafting email and summarizing meetings. The OCC's Tier 4 examples are "internal productivity tools" and "general employee Q&A," which describes that early work well.
The profile matters as much as the label, though. Tier 4 assumes "no customer/financial data exposure." Copilot answers with whatever the person asking can already open. Once it summarizes loan files or member records, the data sensitivity factor moves the use up, and the OCC's own Tier 3 example is "internal document summarization for credit memos." An agent that sends customer communications on its own is a Tier 2 example.
Key Takeaway
Copilot is several uses, each with its own tier. Policy Q&A and drafting that touch no customer data sit at Tier 4, and summarizing credit files sits at Tier 3 with a person reviewing the result. An agent that acts for a customer sits at Tier 2 or higher. Rank the use, then pick the controls.
That ranking also answers the shadow AI question. A loan officer pasting a borrower file into a consumer chatbot is running a Tier 3 use with Tier 4 controls, or none. Our article on shadow AI compliance risk in banking covers how that happens and how to see it.
Why Agents Change the Question
The framework's first section makes a point that cuts against intuition: "a simple rules-based agent authorized to initiate transactions could present greater risk than a sophisticated model limited to internal drafting assistance." Risk follows what the system can do, more than how clever it is.
"The breadth of agent permissions and tool access may be treated as part of the attack surface." (OCC framework, section 2.11)
Section 2.11 goes on to say prompt injection can "cause an agent to misuse connected tools without a conventional network breach," and that broad credentials could let a manipulated agent "cause fraud, data exposure, or operational disruption at machine speed." One answer is identity: agent identities and credentials "may be governed in a manner comparable to privileged users."
That is a familiar discipline. You already limit, monitor and review your administrator accounts. An agent with access to your SharePoint libraries and the ability to send email deserves the same treatment. We explain the identity side in our article on AI agent identities and the injection side in our guide to prompt injection and Copilot.
What Examiners Will Look For
The framework lists considerations, written with "may" and "could" throughout. Five of them translate directly into evidence you can have ready.
The inventory "may capture the use case, owner, provider, permissions, deployment status, monitoring, and fail-over arrangements." The framework adds that activity outside the approved purpose "may be treated as a control failure even when the resulting output or action appears reasonable."
The person approving an AI output may need the expertise and authority "to evaluate, reject, modify, pause, or halt" it, and that review triggers, evidence and procedures "may be defined so that approval is more than a nominal click."
"Prompts, outputs, embeddings, and logs could become repositories of sensitive information and may warrant access controls and risk-based retention periods."
Containment tools the framework names include "least-privilege permissions, and kill switches," because an agentic system "could remain available while taking inappropriate or harmful actions."
"Inputs, outputs, context, retrieved sources, tool calls, approvals, overrides, and actions" may provide the evidence needed to reconstruct a material decision.
Our AI governance assessment checklist walks through building the inventory, and our article on auditing AI governance covers testing it.
Matching Each Tier to Microsoft 365 Controls
Here is the practical part. Most of what an examiner would look for at Tiers 2 through 4 is already in Microsoft 365, and ABT configures it in the tenants we manage. The mapping below is ABT's working model, built from the framework's components and Microsoft's documentation. The tiers are the OCC's; the product mapping is ours.
| Tier | What the framework points to | Microsoft 365 controls that support it |
|---|---|---|
| Tier 4 | Purpose, permitted use, data boundaries | Copilot with enterprise data protection; sensitivity labels and data loss prevention in Microsoft Purview; tightened SharePoint permissions so Copilot sees only what each person should |
| Tier 3 | Human-in-the-loop review, prompt retention, traceability | Microsoft Purview Audit for Copilot interactions; Purview retention policies for Copilot prompts and responses; a named reviewer for each use |
| Tier 2 | Agent identities governed like privileged users, least privilege, human-on-the-loop | Microsoft Entra Agent ID for every agent; Conditional Access that blocks risky agent identities; scoped permissions; an owner who can pause the agent |
| Tier 1 | Containment, kill switches, tested fallback, vendor accountability | Everything above, plus documented containment and fallback, and vendor contracts that spell out how the system is monitored and stopped |
Four of those controls carry the most weight, and Microsoft documents each one:
- Enterprise data protection. Microsoft states that with Copilot, "your data isn't used to train foundation models" (Microsoft Learn). That speaks to the vendor data-use question in section 2.7.
- Audit. Copilot interactions are logged "as part of Audit (Standard)," and the records include "references to files, sites, or other resources" Copilot accessed (Microsoft Learn). That supplies part of the trail section 2.14 describes; approvals and overrides in your own workflows need their own record.
- Retention. Purview retention policies "include user prompts and responses for Microsoft 365 Copilot and Copilot Studio" (Microsoft Learn), which supports the retention point in section 2.7.
- Agent identity. Microsoft Entra Agent ID is generally available, with a Conditional Access template "to block sign-ins from risky agent identities" (Microsoft Learn). That gives agents the identity foundation section 2.11's privileged-user treatment starts from.
Copilot's memory feature deserves its own look before you rely on it in a Tier 3 use. We cover how it is stored and governed in our guide to Copilot memory for banks and credit unions. For agents built in Copilot Studio, our article on agent authentication covers the settings that keep a Tier 2 agent inside its lane.
Find the AI already in use before you tier it
The ABT AI Readiness Assessment answers the questions an examiner would start with:
- Which AI services your people have used in the last 30 days, and who used them
- Whether data loss prevention and sensitivity labels protect member and borrower data
- How Conditional Access, MFA and admin access are set up in your tenant
- A draft AI use policy and a 30/60/90-day plan your board can review
Banks, Credit Unions and Mortgage Companies
The OCC framework speaks to national banks and federal savings associations. CSBS has published its own AI framework for state examiners, NCUA's 2026 priorities point credit union examiners at vendor oversight, and the tiers give every charter a practical way to rank AI uses.
National banks and federal savings associations
Bulletin 2026-54 is what your examiners will carry. They may use it at community banks, tailored to size, complexity and risk profile. The OCC says it intends to update the framework "regularly."
State banks and mortgage companies
On September 16, 2026, CSBS released its own AI Supervisory Framework, "a discretionary tool" for examiners of state-chartered banks and state-licensed nonbank financial institutions. Independent mortgage companies are licensed by the states, so this is the AI framework their state examiners may draw on.
Credit unions
NCUA's 2026 supervisory priorities say examiners "will also assess third-party risk-management practices" where lending, servicing or collection functions are outsourced. The OCC tiers give a credit union a ready-made way to rank its own AI uses. Our guide to NCUA's AI resources covers the rest.
One more point applies to every charter. Many higher-tier AI uses arrive inside a vendor's platform, such as a core, lending or fraud system. The framework is direct about who carries that: "accountability for managing the bank's risk generally remains with the bank." Ask each vendor where its AI features sit on the four tiers and how you would stop one.
The OCC super intelligence framework rewards institutions that know what their AI can do. Inventory each use, rank it by consequence and autonomy, and put the matching control in place. For the Copilot and agent uses most institutions run today, the core controls sit in Microsoft 365, and they do their job once they are configured in your tenant and kept that way. ABT does that work in the tenants it manages.
How ABT Puts the Tiers to Work in Your Tenant
ABT manages Microsoft 365 tenants for more than 750 financial institutions as a Tier-1 Microsoft Cloud Solution Provider. Getting Copilot right at Tier 4 and Tier 3 comes down to the tenant underneath: permissions, sensitivity labels, Purview audit and retention, and Conditional Access. That is the work we do every day.
Microsoft Entra Agent ID is generally available. It gives each AI agent its own identity, separate from the person who built it, and ships with Conditional Access templates, including one that blocks sign-ins from risky agent identities. For an institution that wants to show an examiner its agents are governed "in a manner comparable to privileged users," this is where that evidence starts. ABT configures these policies in the tenants we manage, alongside the Conditional Access baseline already in place for your people.
For the security layer, M365 Guardian hardens the tenant and watches for configuration drift, so the controls you match to each tier stay set. Our guide to examiner-ready Copilot controls lists the settings in detail.
Set up Copilot for Tier 4 and Tier 3 work before the first agent arrives
ABT reviews your permission model, sensitivity label coverage, admin center settings and Copilot governance gaps, and documents the findings in writing before go-live.
Frequently Asked Questions
It is an examiner resource the OCC published on November 4, 2026 in Bulletin 2026-54. It helps examiners evaluate how national banks and federal savings associations use AI, which the OCC now calls super intelligence, and it scales the review to how consequential and autonomous each use is.
The bulletin says the framework "does not establish supervisory expectations for banks" and is a resource for examiners. It still shows how an examiner is likely to approach your AI uses, so it is a useful guide to the evidence worth having ready.
Tier 1 Critical covers uses like an autonomous payment agent or fraud-hold release. Tier 2 High covers AI-assisted underwriting and agent-initiated customer communications. Tier 3 Moderate covers internal summarization for credit memos with a person reviewing. Tier 4 Low covers internal productivity tools and general employee Q&A. The OCC calls the table illustrative.
It depends on the use. Drafting and general employee questions match the OCC's Tier 4 examples. Summarizing credit files matches its Tier 3 example and calls for a person to review the output. An agent acting on behalf of customers matches Tier 2 or higher.
The OCC framework is addressed to national banks, federal savings associations, and federal branches and agencies. State-chartered banks and state-licensed nonbanks such as mortgage lenders have the CSBS AI Supervisory Framework, a discretionary tool for state examiners. Credit unions can use the OCC tiers as a practical model for ranking their own AI uses.
Section 2.11 says agent identities and credentials may be governed in a manner comparable to privileged users, and treats the breadth of agent permissions as part of the attack surface. In Microsoft 365 that means a Microsoft Entra Agent ID for each agent, Conditional Access for agent sign-ins, scoped permissions, and an owner who can pause the agent.