When AI Can Take Action: How Community Banks Should Evaluate AI Agent Risk

A bank may approve an AI assistant to summarize internal reports or help employees draft communications. The risk changes when that technology can retrieve customer information, update records, interact with other applications, or initiate business actions.

These capabilities are increasingly associated with AI agents, which can perform tasks through connected systems rather than simply generate responses.

For community and regional banks, evaluating these applications requires understanding their permissions, the decisions they influence, and the consequences of an incorrect or unauthorized action.

On November 4, 2026, the Office of the Comptroller of the Currency (OCC) issued a new supervisory framework addressing banks' use of advanced artificial intelligence. Its emphasis on potential consequences, autonomy, and human oversight provides useful context for evaluating AI-enabled workflows.

How should community banks evaluate AI agents?

Community banks should evaluate AI agents according to the actions they can perform, the systems and information they can access, and the potential impact of errors or misuse. Higher-consequence applications may warrant tighter permissions, stronger validation, additional monitoring, and more direct human oversight.

The level of control should reflect the individual use case and the bank's existing risk-management practices.

The OCC's November 2026 supervisory framework

On November 4, the OCC published Bulletin 2026-54, Super Intelligence: Supervisory Framework.

The framework provides OCC examiners with a resource for evaluating banks' use of advanced AI, including agentic systems capable of performing tasks or influencing decisions.

The OCC uses the term super intelligence to encompass technologies previously described as artificial intelligence, as well as emerging and advancing capabilities.

Its framework identifies several considerations for assessing these systems, including the potential consequences of failure, the degree of human oversight, and the extent of authority transferred to an AI system.

The OCC also makes an important distinction: the framework does not establish new regulatory requirements, enforceable standards, or supervisory expectations for banks.

It is intended for examiners, and banks are not expected to use the document to assess their own AI risk-management programs.

For community banks, the framework provides useful context for discussing how advanced AI use cases can fit within existing governance, technology, compliance, and operational-risk practices.

The OCC also instructs examiners to tailor their reviews according to an institution's size, complexity, risk profile, and the nature of its AI use.

An AI assistant and an AI agent may create different risks

Consider two hypothetical applications.

The first is an AI assistant authorized to summarize internal policy documents. Employees review the summaries before using the information.

The second is an AI-enabled workflow that can retrieve customer records, update information in a business application, and initiate follow-up actions.

Both applications deserve appropriate review, but the second introduces additional considerations because it can change information or initiate activity.

The distinction depends on what the system is authorized to do.

An advanced language model used only for internal drafting may present less operational risk than a relatively simple automated system authorized to initiate transactions.

The OCC's November framework specifically recognizes that a system's technical sophistication is not necessarily the best indicator of its potential consequences.

Banks should therefore examine an application's actual capabilities and permissions.

Review the actions an AI system can perform

An AI inventory helps identify which applications are present in the bank's environment.

For an application capable of taking action, additional detail may be necessary.

The bank should understand whether the system can retrieve information, modify records, send communications, approve activities, or interact with another application.

These capabilities can change the consequences of an error.

For example, an AI tool that drafts a suggested customer response creates one type of review requirement.

A system authorized to send that response automatically creates another.

When evaluating an AI-enabled workflow, consider:

• Which business process does the application support?
• What actions is it authorized to perform?
• Which actions require approval?
• What customer or bank information can it access?
• Which systems can it interact with?
• Can its permissions be restricted or revoked?

An application's documented purpose should match the permissions it actually has.

A tool approved for internal research does not necessarily need the ability to modify customer records or interact with production systems.

Match human oversight to the potential consequences

Human review is particularly important when an AI-enabled workflow can affect customers, financial activity, sensitive information, or critical operations.

The appropriate controls depend on the use case.

For a low-impact drafting activity, employee review before publication may be sufficient.

For an application authorized to take consequential actions, the bank may need additional approval steps, restrictions, testing, or escalation procedures.

The OCC's framework recognizes that higher-consequence use cases may warrant stronger oversight and limitations on autonomy.

Bank leadership should consider where human intervention is necessary and whether that intervention would be meaningful.

For example, requiring an employee to approve a transaction without providing enough information to evaluate it may offer limited protection.

A useful approval process should give the reviewer sufficient context to understand the proposed action and its potential impact.

Evaluate the information and systems an AI agent can access

AI agents may depend on connections to business applications, document repositories, databases, or other technology services.

Those integrations can improve efficiency, but they may also extend the application's access to information or capabilities beyond its original purpose.

A bank evaluating an AI agent should consider the permissions associated with both the application and its connected services.

Relevant questions include whether access is limited to approved information, how integration credentials are protected, and who can authorize additional connections.

The review should also address what happens when an employee changes roles, an application is retired, or a vendor relationship ends.

Banks can incorporate these considerations into existing identity management, cybersecurity, and third-party risk processes.

Include vendor-managed AI capabilities in the review

A bank may encounter AI agents through software it already licenses.

An existing technology provider might introduce an automated workflow, AI assistant, or new integration without requiring the bank to purchase a separate standalone AI product.

That creates a practical vendor-management question: has the functionality changed in a way that affects the relationship's existing risk assessment?

When a provider introduces AI-enabled capabilities, bank leadership should understand which functions are available, what permissions they require, and whether those functions are enabled by default or require approval.

The bank should also consider who is responsible for configuring access, monitoring relevant activity, and addressing security concerns.

These reviews can complement the bank's broader third-party risk-management program.

For additional guidance, see AvTek's article on AI governance for community banks, which discusses visibility into approved applications, policies, and governance documentation.

Document decisions and review material changes

A bank's AI governance records should help leadership understand why an application was approved and which controls apply to it.

For AI-enabled workflows capable of taking action, useful documentation may include the permitted use case, authorized functions, relevant systems, assigned owner, and required human oversight.

For example:

These are practical governance examples, not a regulatory checklist created by the OCC.

A bank may also need to reassess an application when its capabilities change. An AI assistant that originally generated draft text could later gain access to new business systems or automated actions.

Material changes to access or functionality deserve review before they become routine operations.

Questions bank leadership should ask before approving an AI agent

When evaluating an AI-enabled application, CIOs, IT Directors, Risk Officers, and other bank leaders can begin with a few questions:

• Can the application take actions, or does it only provide information for employees to review?
• What business consequences could result from an incorrect or unauthorized action?
• Are permissions limited to the application's approved purpose?
• Which activities require meaningful human oversight?
• Can the bank identify relevant activity and investigate exceptions?
• Who is responsible for reviewing the application when its capabilities or integrations change?

The answers can help a bank determine whether its existing governance practices adequately address the application or whether additional safeguards should be evaluated.

Supporting risk-based AI adoption

As banks introduce more AI-enabled workflows, they need a clear understanding of the applications in use, their permissions, and their potential cybersecurity and operational risks.

AvTek Solutions offers AI Governance Services to help organizations assess current AI use, establish acceptable-use requirements, review approved applications, and define governance responsibilities.

AvTek's vCISO Services also support cybersecurity risk assessments, policy development, remediation planning, and communication of technology risks to leadership.

These services can help community and regional banks evaluate the security and governance considerations associated with expanding AI adoption.

If your bank is considering AI applications that interact with sensitive information or take actions across business systems, understanding their permissions and potential consequences is a useful place to begin.

Talk with AvTek about evaluating AI governance and security risks at your bank.