What the OCC’s supervisory framework means for banks
Executive summary
The OCC’s new AI (Super Intelligence) supervisory framework puts the burden on banks to demonstrate that governance and controls match the risk of each AI use. Banking leaders should prioritize complete inventories, risk-based controls, effective human oversight, third-party governance and defensible evidence to prepare for examinations and support responsible AI adoption.
Needed: A documented, well-supported position
The burden of demonstrating sound AI governance has been placed squarely on the shoulders of financial institutions as a result of a new supervisory framework issued by the Office of the Comptroller of the Currency (OCC) that bank examiners will use to evaluate how national banks and federal savings associations use AI.
The framework, issued on Nov. 4, is neither a rule nor a self-assessment tool. It defines what examiners will review and how they will calibrate that review to the risk of each use of AI, which the OCC now refers to as “Super Intelligence” (SI). This is consistent with the principles-based approach the banking agencies have applied to model risk, third-party risk and Bank Secrecy Act/Anti-Money Laundering systems for more than a decade.
Under this approach, the burden of demonstrating sound governance rests with the institution. There is no prescribed set of controls. Each bank must be able to articulate how it assesses the risk of each AI use, why the controls applied to that use are appropriate, and how it has determined that those controls operate effectively. Institutions that do this well also benefit: a documented, well-supported position establishes the basis on which examiners evaluate the program.
Banks already anticipate this scrutiny. In Grant Thornton’s 2026 Cyber & Risk Survey, 60% of banking respondents identified AI governance and the use of AI as a top source of future risk or regulatory scrutiny.
Drawing on prior supervisory guidance, it’s possible to anticipate how examiners will apply the framework, where it extends current practice and the actions banks should prioritize now.
“Banks are adopting AI quickly and with clear intent,” said David Head, Grant Thornton’s Head of the Banking Industry. “What will distinguish institutions is whether they can demonstrate that governance has kept pace with adoption, use by use. The OCC’s framework gives boards and management direction as to where that scrutiny will fall.”
Key considerations for leadership
- The supervisory question has shifted from “Does the institution have AI governance?” to “Can the institution demonstrate that its governance is appropriate for each use?”
- Supervisory scrutiny will be driven by what an AI system is permitted to do, not by its technical sophistication. AI risk is therefore a business design decision, and accountability belongs with the business.
- Institutions outside the OCC’s supervisory perimeter should anticipate that the framework will become a common reference point for auditors, partner banks and other regulators.
Regulatory precedent
Historically, banking agencies have responded to new technologies and emerging risks in a relatively consistent manner. They issue principles-based guidance rather than prescriptive requirements, direct examiners to scale their work to each institution’s size, complexity and risk profile, and place responsibility on the institution to demonstrate that its program is commensurate with its risk.
This has been the case for key releases such as the OCC examiner guidance on credit scoring models in 1997 (which was rescinded earlier this year), the interagency statement on model risk management for BSA/AML systems in 2021, and revised model risk guidance earlier in 2026. The AI framework follows the same approach.
Lessons learned from prior guidance
- Principles become expectations through the examination cycle. Supervisory standards are defined over time through examination findings and peer comparisons. Institutions that established and documented their own standards early were better positioned to influence how those standards developed.
- A complete inventory is foundational. Model risk and third-party risk programs that encountered supervisory difficulty frequently began with an incomplete inventory. The same risk exists for AI, much of which is embedded in vendor products.
- Compliance-driven checklists are not sustainable. Programs designed to satisfy an assessment tool required redesign when the tool was retired. Programs grounded in the institution’s actual risk profile remained durable.
- Reliance on third parties tends to outpace oversight of them. This pattern emerged with core processing and fintech relationships, and it is recurring with AI.
“Bank examiners will ask what the institution decided, why and how it knows its controls work,” said Tariq Mirza, Grant Thornton Partner and Regulatory Compliance Leader. “Institutions need a documented methodology and supporting evidence so they frame that discussion instead of being benchmarked against what examiners have seen at their peers.”
How we can help you
INDUSTRY
SERVICES
Overview of the framework
The framework applies to AI used in the ordinary course of business, including credit decisioning, payments, fraud detection, customer service, operations and risk management. The SI terminology reflects a September 2026 executive order directing federal agencies to adopt that term. In practice, it encompasses the generative, agentic and traditional AI that institutions already manage.
Two features of the framework will shape its application:
- AI is addressed within existing risk categories. The OCC treats AI as a source of new risk, or as a factor that may change the level of existing risk, across operational, compliance, credit, third-party and other risk categories. Institutions are not expected to establish a separate AI risk appetite.
- Supervisory review is calibrated to each use. The depth of supervisory review depends on the materiality of each AI use and on the institution’s size, complexity and risk profile. Higher-consequence uses will receive greater scrutiny.
The OCC states that the framework does not establish new supervisory expectations, is not intended for institutional self-assessment and will be updated as the technology and supervisory experience change. It nonetheless represents the clearest indication of how examiners will approach AI and should be treated accordingly.
A strategic opportunity
The absence of prescribed controls gives each institution latitude to define how it assesses AI risk, which controls apply at each risk level and what evidence it maintains. An institution that enters an examination with a clearly articulated risk appetite statement, a well-defined risk assessment methodology, a reconciled inventory and supporting evidence for its highest-risk uses provides examiners with an established basis for testing. Institutions that defer this work are more likely to be evaluated against practices examiners observe at peer institutions.
Grant Thornton’s 2026 Cyber & Risk Survey indicates that the conditions the framework targets are already present across the banking sector. Banking respondents identify AI governance as a leading source of future regulatory scrutiny, point to third-party relationships as a principal cybersecurity challenge, and report that fragmented risk functions increase their exposure to fraud.
60%
of banking respondents cite AI governance and the use of AI as a top source of future risk or regulatory scrutiny, compared with 57% across industries
71%
cite third-party risk as a top cybersecurity challenge, compared with 61% across industries
47%
say poor risk management integration increases fraud exposure, compared with 34% across industries
Each survey finding corresponds to a supervisory focus area under the framework:
- Institutions already anticipate scrutiny of AI governance; the framework now defines how that scrutiny will be applied.
- Third-party relationships, where much of the industry’s AI is embedded, are an area banks find difficult to secure.
- Where cybersecurity, fraud, AML and compliance functions operate separately, AI risk is likely to fall between them, which is precisely the accountability gap the framework requires institutions to close.
Where the framework builds on current practice
Most elements of the framework align with programs institutions already maintain: model risk management, third-party risk management, operational resilience, cybersecurity, data governance and internal audit. But to align with the framework, each program must be extended to address AI.
| Area | Prevailing practice | Supervisory focus under the framework |
|---|---|---|
| Use-case approval and scope | AI tools approved through technology or model governance, often with a general description of intended use | A specific, documented purpose and a defined set of permitted actions for each AI use. Operating outside the approved scope may constitute a control failure, regardless of output quality. |
| Risk assessment of AI uses | Models tiered by materiality, often based on model type or complexity | Each use assessed by what it is permitted to do: degree of autonomy, reversibility of its actions, and customer and financial impact. Favorable testing results do not reduce the inherent risk rating. |
| Human review and approval | A designated reviewer or approval step | Reviewers with the authority, expertise, capacity and information necessary to change the outcome. High-volume approvals without substantive review may not be regarded as an effective control. |
| Validation and ongoing monitoring | Validation prior to deployment and on a periodic cycle | Continuous monitoring where outputs vary between runs, drift over time or change when a vendor updates the underlying model. |
| Operational resilience | Planning for outages and recovery | Planning for AI systems that remain available but perform incorrectly, including detection, containment, manual fallback and customer remediation. |
| Third-party risk management | Due diligence and contract review at onboarding | Governance of vendor-operated AI agents within bank systems equivalent to that of privileged users; contract provisions addressing AI incidents, model change notification and remediation; and assessment of concentration among common providers. |
| Documentation and audit trail | Documentation of design decisions and approvals | The ability to reconstruct any material AI-driven action from its inputs, outputs, systems accessed, approvals and overrides. |
Source: Grant Thornton analysis of the OCC Supervisory Framework for Bank Use of Super Intelligence (November 2026).
Illustrative example: Financial crimes compliance
An institution deploys a generative model to draft case narratives that investigators review, edit and approve. The institution subsequently integrates the same model with its case management system and permits it to close lower-risk alerts without human review. The model is unchanged, but the risk profile is materially different. The second design executes an action, removes a human decision point and may be difficult to remediate if suspicious activity is missed. The OCC’s illustrative risk tiers place AML alert disposition in the high-risk tier. Because a design decision drove the change in risk, accountability must rest with the business, not solely with technology or model risk functions.
“In financial crimes compliance, the line between AI that assists an investigator and AI that acts without supervision can be crossed with a single configuration change,” said Kyle Daddio, Partner, Forensics & Financial Crimes and AML & Sanctions Practice Leader. “When a model moves from drafting a narrative to making decisions like closing an alert, the institution has changed its control environment and its suspicious activity reporting risk. That change needs to go through the same approval, testing and accountability as any other change to a high-risk process because examiners will treat it that way.”
Priority actions for banks
The following practices reflect what we expect examiners, auditors and boards to look for. They are organized in line with how many institutions already structure risk management.
AI inventory and use-case approval
Effective governance begins with a complete view of where AI is deployed and what each system is permitted to do. At many institutions, AI has been adopted incrementally across business units and vendors, and a growing proportion is enabled through routine vendor product updates rather than formal approval processes. In our experience, vendor-embedded AI is a common gap. To align with the framework, institutions will need:
- A single enterprise inventory, reconciled across business lines, technology, procurement and model risk, which includes AI embedded in vendor products
- A documented purpose and defined permitted actions for each use, with an accountable business owner
- A mechanism to identify changes in scope and the reasons behind them in addition to changes in performance
Risk assessment and control standards
The OCC bases control strength on the consequences of failure and the transfer of decision-making authority. For banks, that makes AI risk partly a business-process design decision: changing what a tool is permitted to do can change its risk even when the underlying model remains the same.
Each AI use should be assessed based on what it is permitted to do and the consequences of error: the extent to which decision-making has shifted from people to the system, whether its actions can be traced, reproduced or reversed, and the associated customer, financial and data exposure. That assessment should determine the required level of control.
Two shortcomings are common: tiering by technology type, which understates the risk of simple systems that execute actions, and applying model materiality scales that were not designed for systems that act. A portfolio view is also essential. Multiple moderate-risk uses that depend on the same provider or data source can create a concentration that no individual rating reveals. To align with the framework, institutions will need:
- A risk assessment methodology aligned with the existing risk taxonomy and appetite statements, rather than a standalone AI scale
- Minimum control requirements defined for each risk tier to ensure a consistent response
- An aggregate view of shared providers, models and data sources
Human oversight, monitoring and escalation
Many institutions rely on human review as a key AI safeguard. To be effective, reviewers need the authority, expertise, capacity and information to change the outcome. Where review volumes allow only cursory evaluation, the control may be documented but ineffective in practice.
Oversight must also operate at the system level. Institutions need the capability to detect an AI system that remains operational but is performing incorrectly, suspend it, revert to manual processes and remediate affected customers while maintaining critical operations. To align with the framework, institutions will need:
- Documented authority to override or suspend each system, with reviewer capacity assessed against actual volumes
- Reviewer training focused on effective challenge of the specific uses they oversee
- Suspension, fallback and escalation procedures that have been tested, not merely documented
Third-party and vendor AI
Many institutions source AI from vendors that often operate within bank systems rather than simply returning outputs. A vendor-operated agent with access to bank systems should be governed with the same rigor the institution applies to its own privileged users.
Contracts executed before AI became part of the service seldom address what an institution requires when a provider modifies its model or an incident occurs.
To align with the framework, institutions will need:
- Identity, access and monitoring controls for vendor-operated agents equivalent to those applied to internal privileged access
- Contract provisions that define AI incidents and require notification of model changes, cooperation in remediation and access to necessary evidence
- Visibility into which critical processes depend on the same AI provider
Documentation, audit trail and independent testing
For higher-risk uses, an institution should be able to trace what the system received, what it produced, which systems it accessed, who approved each step and where overrides occurred, even when the model cannot explain its reasoning.
With examiners now directed to focus on material financial risk, institutions need to demonstrate that controls operate effectively for the uses with the greatest potential for harm.
To align with the framework, institutions will need:
- Logging and retention sufficient to reconstruct a material action end to end
- An indexed evidence file for each higher-risk use linking its purpose, risk rating, controls, testing and results
- Independent testing by internal audit, separate from the functions that develop and validate the systems
Implications across the industry
By institution profile
The framework does not establish an asset threshold, and examiners are directed to tailor their approach. Starting points and principal challenges will vary by institution profile.
| Community banks | Regional and midsize banks | Large and complex banks | |
|---|---|---|---|
| Typical starting point | AI adopted primarily through core processing, fraud and customer service vendors | AI deployed in fraud, AML and servicing, supported by established model risk and third-party programs | Mature model inventories, three lines of defense and operational resilience programs |
| Principal challenges | Identifying AI enabled by vendors; documenting specific purpose statements; securing change-notification rights | Risk assessment of priority uses; reviewer capacity at actual volumes; drift monitoring | Validating systems with variable outputs; resilience for agents that execute actions; decision reconstruction; board reporting aligned to the inventory |
| Likely examiner question | “What AI is in use, and who approved it?” | “Is human review effective at actual volumes?” | “Walk us through this high-risk decision and demonstrate that the backstop has been tested.” |
Across the three lines of defense
Because AI risk is addressed within existing risk categories, the framework places demands on all three lines simultaneously. The first line is accountable for the purpose, risk assessment and day-to-day oversight of each use. The second line must integrate AI into risk appetite, model risk, third-party risk and monitoring without establishing a parallel structure. Internal audit must be able to reconstruct material decisions and test backstops independently of the functions that develop and validate the systems. Gaps and overlaps in accountability are frequent issues: vendor-embedded AI without a clear owner and agents that each line assumes another is monitoring.
Beyond the OCC’s supervisory perimeter
Although the framework applies to OCC-supervised institutions, state-chartered banks, credit unions, fintechs and technology providers serving banks will also be affected. Until other agencies issue comparable guidance, auditors, partner banks, boards and other supervisors are likely to use it as a benchmark. Institutions that align now will be better positioned as other regulators respond.
“Being regulator-ready means an examiner can follow a single decision from policy to outcome and see, in reproducible evidence, that every control along the way performed as designed,” said Erin Lentz, Risk Advisory Services Partner and Fraud & Financial Crimes Practice Leader. “Institutions that have navigated evolving requirements most effectively tested their own evidence before the examination and treated accountability as a business responsibility.”
Questions for boards and executive management
Leadership should be able to answer the following with confidence.
AI inventory and use-case approval
- Do we have a complete view of where AI is in use, including within vendor products, and what each system is permitted to do?
- Would we detect a system operating outside its approved purpose?
Risk assessment and control standards
- Is our most rigorous oversight directed at AI that executes actions, rather than AI that only informs decisions?
- Do we understand where multiple uses depend on the same provider, model or data source?
Human oversight, monitoring and escalation
- Do our reviewers have the authority to override or suspend an AI system, or has approval become a formality?
- Do our reviewers have the expertise and capacity to oversee the AI uses for which they are responsible?
- If a system continued operating but began producing errors, how quickly would we detect it, and could we contain it?
Third-party and vendor AI
- Do our contracts with AI providers provide the visibility and rights we would require in the event of an incident?
- Are vendor agents operating within our systems subject to the same access controls as our own personnel?
Documentation, audit trail and independent testing
- Could we reconstruct a material AI-driven decision from the records we maintain today?
- Have our highest-risk uses been tested by a function independent of development and validation?
Any question that cannot be answered with confidence should be addressed before the next examination cycle.
Contacts:
Head of Banking Industry
Grant Thornton Advisors LLC
David serves as Head of Banking Industry and has over 25 years’ experience in professional services and banking.
Charlotte, North Carolina
Industries
- Banking
- Healthcare
- Manufacturing
- Transportation & Distribution
- Not-for-profit & Higher Education
Partner, AI & Transformation
Grant Thornton Advisors LLC
Vlad Anichkin is a Partner in Grant Thornton’s Advisory business, helping organizations harness the power of AI and digital transformation to achieve measurable growth, efficiency, and innovation.
Jacksonville, Florida
Industries
- Life Sciences
- Retail & Consumer Brands
- Technology
Partner, Forensics & Financial Crimes Leader
Grant Thornton Advisors LLC
Erin is a partner, Forensics & Financial Crimes Leader practice located in Charlotte, NC with over 12 years of experience.
Charlotte, North Carolina
Service Experience
- Advisory Services
Partner, Forensics & Financial Crimes
AML & Sanctions Practice Leader
Grant Thornton Advisors LLC
Kyle Daddio is a Partner, Forensics & Financial Crimes, where he leads the firm’s AML and Financial Crimes Compliance practice.
Chicago, Illinois
Service Experience
- Regulatory compliance
- Cyber & Risk Advisory
Partner, Regulatory, Risk & Compliance
Grant Thornton Advisors LLC
Tariq Mirza leads Grant Thornton’s Regulatory Center of Excellence and Compliance Risk Advisory Services.
Arlington, Virginia
Industries
- Banking
Service Experience
- Advisory Services
Content disclaimer
This Grant Thornton Advisors LLC content provides information and comments on current issues and developments. It is not a comprehensive analysis of the subject matter covered. It is not, and should not be construed as, accounting, legal, tax, or professional advice provided by Grant Thornton Advisors LLC. All relevant facts and circumstances, including the pertinent authoritative literature, need to be considered to arrive at conclusions that comply with matters addressed in this content.
Grant Thornton Advisors LLC and its subsidiary entities are not licensed CPA firms.
For additional information on topics covered in this content, contact a Grant Thornton Advisors LLC professional.
Ready to talk? We’re ready to listen.
Request a meeting and a member of our team will be in touch to see what we can do to meet your needs.
Want to submit an RFP? Please submit your request through our RFP submission page.
Share with your network
Share