What will OCC examiners check when banks let AI agents act?
A new US examiner framework says control intensity should follow consequences, autonomy and human oversight—not model sophistication alone. It is guidance for supervisors, not a new rule or a finding that any bank system is safe.
The short version
- What happened
- The OCC's 22-page framework gives examiners 16 components across six pillars: purpose, risk, control, oversight, resilience and assurance.
- Who it affects
- Customers should have a route to challenge and correct AI-assisted decisions or transactions; a human label without authority and evidence is not meaningful oversight.
- What remains unproven
- Nonbinding examiner resource, not a new rule or enforceable standard
Keep in mind
Nonbinding examiner resource, not a new rule or enforceable standard
No examination outcomes, incident denominator or bank-level findings
No quantified performance threshold for any control
US supervisory context; other jurisdictions use different legal frameworks and terminology
The framework's effect on customer harm and bank resilience has not yet been measured
Source analysis · 3 linked sources. Primary analysed source date: 4 November 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 9 min The evidence, methods, implications and limitations
The direct answer: examiners are being told to follow consequences and delegated agency
The Office of the Comptroller of the Currency has published its first dedicated framework for examiners reviewing how national banks, federal savings associations and federal branches or agencies use artificial intelligence. The document uses the federal government's new term “Super Intelligence”, but its footnote ties that term to technologies previously covered by the statutory definition of artificial intelligence. It is not evidence that a bank has deployed a superhuman system. The practical question is more ordinary and more important: what can the system do, what happens if it is wrong or misused, and who can stop or correct it?
The OCC tells examiners that risk classification may depend on the consequences of a use case, the transfer of agency to the system and the degree of human oversight rather than the sophistication of the underlying model. Its example is useful: a relatively simple agent authorised to initiate transactions can create more material risk than an advanced model used internally to draft text. That moves attention away from model marketing and toward permissions, money movement, customer effects and recovery. It also means a bank cannot treat a less capable model as automatically low risk when the surrounding workflow gives it real authority.[1][2][3]
Six pillars turn a broad technology question into 16 supervisory components
The framework groups 16 components under purpose, risk, control, oversight, resilience and assurance. Purpose asks whether the bank has defined an intended use, permitted boundaries and a business justification. Risk covers classification, potential customer and financial consequences, legal obligations and how much agency moves from a person to software. Control includes model and data management, privacy, compliance, cyber security and third-party arrangements. Oversight covers accountable owners and meaningful human checkpoints. Resilience concerns failure modes, continuity and the ability to constrain or stop an agent. Assurance covers monitoring, testing, auditability, internal audit and reporting to management or the board.
Those categories deliberately build on existing banking disciplines rather than inventing a separate AI risk universe. The OCC says examiners should not treat the technology as a distinct risk category and banks should not be required to create a separate AI risk appetite. Existing obligations for safe and sound operation, consumer compliance, information security, model risk and vendor management remain the foundation. The framework is intended to help an examiner identify technology-specific considerations inside those established systems. For a small community bank, the work is meant to be proportional to the institution's size, complexity and risk profile as well as the autonomy and consequences of the use case.[1][2]
The distinction that matters
The OCC's six-pillar examination path
The framework moves from intended use to evidence that the bank can control, recover from and independently test the system.
Purpose and risk
Define the use, authority and possible consequences
Materiality-based classification
Control and oversight
Set permissions, human checkpoints and accountable owners
Bounded agency and escalation
Resilience
Prepare for failure, attack, outage and change
Recovery, limits and stop mechanisms
Assurance
Monitor, test, audit and report
Evidence for examiners and management
Human oversight must be designed around a decision, not added as a label
A statement that a human is “in the loop” is not enough. The framework asks whether checkpoints occur before an irreversible or consequential action, whether the reviewer has enough information and time to challenge the output, and whether escalation routes are clear. A person who routinely approves hundreds of opaque recommendations may provide less effective control than a smaller number of targeted interventions with visible evidence and authority to stop the process. Banks therefore need to connect human review to the specific harm it is intended to prevent, such as an improper credit decision, an unauthorised payment, a misleading customer message or a trading action outside limits.
The same logic applies to autonomy. An agent may be allowed to gather information or prepare a recommendation while remaining unable to execute a transaction. Higher-consequence systems may need tighter permissions, transaction limits, dual approval, circuit breakers or kill switches. Controls should also cover non-human identities and credentials, because an agent acting through a service account can accumulate privileges outside ordinary employee access reviews. The framework does not prescribe a single architecture. It gives examiners a way to ask whether authority, monitoring and recovery are proportionate to the consequences if the system behaves unexpectedly.[2]
Testing has to cover repeated operation, change and hostile conditions
A successful demonstration is not the same as reliable operation. The OCC directs attention to testing before use, monitoring after deployment and reassessment when data, models, prompts, tools or external services change. For agents, one pass through a happy-path scenario cannot show how the system behaves after many steps, when a tool fails, when instructions conflict or when an attacker manipulates the context. A bank should be able to reproduce important actions, retain evidence of the inputs and permissions involved, and determine whether the output came from a model, a rule, a human edit or a downstream tool.
Vendor relationships add another layer. Banks may not control a provider's model updates, data handling or subcontractors, yet they remain responsible for regulated activities. Examiners can ask whether contracts require adequate notice of material change, whether the bank can test the service, what happens during an outage and how the bank exits or replaces the provider. Concentration matters when many institutions depend on the same model or cloud service. The framework does not quantify a safe failure rate or mandate a particular test suite. It establishes the questions that evidence should answer.[2]
Customers should see remediation and accountability, not just efficiency claims
For customers, the immediate value of the framework is indirect. It does not create a new right, but it tells examiners to look at compliance, privacy, fair treatment and remediation inside AI-enabled banking processes. If an automated system blocks a payment, recommends a product, handles a complaint or contributes to a credit decision, the bank still needs an accountable route to investigate errors and correct customer harm. Monitoring should not stop at model accuracy. It should include overrides, complaints, disparate outcomes, security incidents, failed transactions and cases in which staff could not understand or reverse an action.
For bank workers, the document makes responsibility more explicit. Product teams need an inventory of use cases and permissions; risk staff need classification tied to real consequences; compliance teams need evidence that legal duties survive automation; security teams need controls for tools and non-human identities; internal audit needs records it can independently test. Front-line employees also need a clear way to report strange behaviour without being told that the model is outside their remit. None of this proves that a particular bank has implemented these controls. It defines what an examiner may reasonably ask to see.[2][3]
What the framework does not do
The OCC repeatedly limits the document's legal effect. It does not establish new supervisory expectations for banks, set enforceable standards or prescribe requirements. Banks are not expected to use it as their own assessment template. The intended audience is the examination workforce, and examiners are told to tailor their work to materiality and the bank's risk profile. The release contains no examination results, incident data, adoption denominator or list of institutions using agents. It therefore cannot support a claim that banks are already compliant, that agentic systems have caused widespread harm or that the framework will prevent failures.
The regulator also discloses that artificial intelligence assisted in preparing the framework and that OCC staff reviewed, edited and approved it. That is a useful transparency note, not a validation of the framework's effectiveness. The document will be updated as technology and supervisory understanding change. Its real influence will become clearer through examination findings, risk alerts, public enforcement actions and banks' documented changes in practice. Until then, it is a consequential supervisory signal rather than an observed outcome.[1][2][3]
Global context and what would change the assessment
The framework directly applies to OCC examiners and federally supervised institutions in the United States. International banks with US branches may encounter it, while other regulators will continue applying their own model-risk, outsourcing, operational-resilience and consumer-protection rules. The broader principle travels well: classify a system by the authority it receives and the harm it can produce, not by a vendor's capability label. But equivalent terminology should not be assumed across jurisdictions, and the US label “Super Intelligence” should not be read as an international technical standard.
Confidence in the framework would rise if the OCC publishes anonymised findings showing which controls failed or worked, how examiners applied proportionality and whether remediation reduced incidents or customer harm. Independent evidence comparing banks with different oversight designs would be stronger than compliance attestations alone. Material enforcement cases would clarify the boundary between this nonbinding resource and existing legal duties. Because today's publication is examiner guidance without an active incident, market disruption or urgent protective action, it belongs in Latest—not Breaking.[1][2]
Continue the story
Read next
Continue exploring banking: compare the finding and its evidence limits.
What will Hong Kong expect from banks using AI agents?
Hong Kong's banking supervisor has now launched Project SAGE, formed a cross-sector expert group and published its agent-governance discussion paper. The earlier ART controls remain a direction of travel, not final rules.
Read the next reportExplore the full source and evidence record
What this means for people
- Customers should have a route to challenge and correct AI-assisted decisions or transactions; a human label without authority and evidence is not meaningful oversight.
- Bank employees need clear ownership, escalation and stop mechanisms when an agent acts unexpectedly.
- Community-bank examinations are meant to be proportional to size, complexity, autonomy and consequences rather than copying the controls of the largest institutions.
Global context
The document directly guides US federal bank examiners, but its consequence-and-agency approach is relevant to any jurisdiction assessing AI in credit, payments, fraud, customer service, operations, risk or markets. Local law still determines enforceable duties.
What the evidence does not yet show
- Nonbinding examiner resource, not a new rule or enforceable standard
- No examination outcomes, incident denominator or bank-level findings
- No quantified performance threshold for any control
- US supervisory context; other jurisdictions use different legal frameworks and terminology
- The framework's effect on customer harm and bank resilience has not yet been measured
What to watch next
- OCC risk alerts, examination findings or enforcement actions applying the framework
- Evidence on customer remediation, discriminatory outcomes and unauthorised transactions
- Bank disclosures about agent permissions, monitoring and kill switches
- Updates covering new agent architectures and shared service-provider concentration
- Comparable guidance from other banking regulators
Living evidence record
Impact record IAI-0RRFT0M
Evidence stage
Announced
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Not yet
Record status
Monitoring
Last checked
4 November 2026
Source trail
3 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 3 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 4 November 2026
Research question, topics and regions
Topics: Banking · Financial regulation · AI agents · Human oversight · Operational resilience
Regions: United States · North America · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.