Vendor AI and the OCC: what section 2.9 says about AI a bank buys

Published Nov 21, 2026 · Last updated Nov 21, 2026

Section 2.9 of the OCC's November 2026 AI framework for examiners covers AI a bank gets from vendors, and section 2.4 says accountability for the bank's risk generally remains with the bank even when vendors build or host the system. Section 2.9 adds four considerations on top of the existing interagency third-party guidance: provider agent visibility, AI incident definitions, contract liability for model errors, and notice and retesting when vendors change or add AI.

TRACKED DOCUMENTS: OCC Bulletin 2026-54 · OCC Bulletin 2023-17

Does the OCC framework cover AI a bank buys from a vendor?

Yes. Section 2.9 of the framework is written for AI that comes from outside the bank, and section 2.4 says the bank stays accountable for the risk even when a vendor builds or hosts the system.

The OCC calls AI “Super Intelligence (SI)” in this framework. This page writes “AI” and keeps “SI” only inside quotations. The framework was sent to examiners with OCC Bulletin 2026-54 on November 4, 2026. It is a resource for examiners. The bulletin says it “does not establish supervisory expectations for banks”.

Section 2.9 opens with a plain fact: “Banks frequently obtain SI capabilities from third parties rather than develop them internally.” A third party is any outside company the bank relies on, such as a software vendor, a cloud provider or a fintech partner. The section says this reliance can bring “model, operational, data, legal, concentration, and resilience risks that are not fully addressed by conventional service-level and information-security reviews.” Put simply, the usual vendor checks look at uptime and security. They do not ask whether the vendor's model gives wrong answers, changes without notice, or is the same model every other bank uses.

On responsibility, section 2.4 says: “Although development or hosting could be distributed across vendors, accountability for managing the bank's risk generally remains with the bank.” For example, if a vendor chatbot on the bank's website quotes a customer the wrong overdraft fee, the customer complaint lands at the bank, not at the vendor. Who pays for the loss between the two companies is a contract question, covered below.

What extra considerations does section 2.9 list for vendor AI?

Section 2.9 lists four: provider agent visibility, AI incidents, liability, and change and exit. They apply “where the provider's SI can affect the bank's decisions, customers or data.”

The framework says these considerations “may supplement existing interagency third-party risk management guidance.” So they sit on top of the normal vendor process. They do not replace it. Each one below gives the framework's words and then an example of our own.

  • Provider agent visibility. Banks “may define the scope of authority, identity, permitted tool use, data use and escalation limits for provider agents that interact with bank systems and data.” For example, a vendor's agent that reconciles payments could be given read access to one ledger and no ability to send money. The next section explains provider agents.
  • AI incidents. The bank “may choose to define a specific SI incident vocabulary and notification timelines to capture events such as hallucination, misfire, drift or manipulation.” A hallucination is a confident answer that is false. Drift is a slow change in how a model behaves as its inputs or versions change. In practice, a standard contract may only make the vendor report outages and data breaches. A chatbot that starts inventing loan rates is neither, so it may never be reported unless the contract names it.
  • Liability. “The bank should consider how contracts assign responsibility for model related errors or losses including customer harm for which the bank may seek clarity, indemnification, caps or loss responsibility.” Indemnification means one party agrees to cover the other's losses. A cap limits how much it pays. For example, if a vendor's fraud model wrongly freezes hundreds of accounts, the contract decides who pays for refunds and fees. Here section 2.9 uses “should”, where most items use “may”.
  • Change and exit. “Existing vendors may add SI features, functionality or change models connected to products already in use by the bank.” The bank “may require notice and transparency of new SI functionality and may reassess the relationship or retest the model before or when the changes are made.” For example, a core banking or CRM vendor may switch on an AI email writer in a routine update. Nobody at the bank bought an AI product, yet staff now use one.

The framework's summary table names the risk this component guards against: “Opaque provider chains, weak contractual rights, switching constraints, or concentration in common models and infrastructure”. Concentration means many banks, or many use cases in one bank, depending on the same model or provider. If that provider fails, they all fail together.

What is a provider agent?

A provider agent is an AI agent run by a vendor that works inside or alongside the bank's own systems and data. An AI agent is software that can take actions, such as calling tools or updating records, not only write text.

Neither section 2.9 nor the glossary in Appendix A defines “provider agent”. The glossary does define “SI agent” and “third-party SI risk”. Section 2.9 describes provider agents as agents “that interact with bank systems and data.” It says the bank “may also monitor and potentially limit what provider agents do without impacting the relationship with the provider.” Put simply, the bank can keep watch over the vendor's agent and put limits on it, and doing so need not damage the business relationship.

The five things section 2.9 says a bank may define for a provider agent, with our examples:

Item in section 2.9Plain meaningExample (ours)
Scope of authorityWhat the agent is allowed to decide or doIt can draft a dispute letter but not send it
IdentityA distinct login, so its actions can be told apart from a person'sA named service account, not a shared staff password
Permitted tool useWhich systems and functions it may callSearch the knowledge base, but no access to the payments system
Data useWhich data it may read, keep or pass onAccount balances yes, full card numbers no
Escalation limitsWhen it must stop and hand over to a personAny refund above an amount the bank sets goes to a human

Section 2.11 on cybersecurity adds a related point. Agent identities and credentials “may be governed in a manner comparable to privileged users”. A privileged user is an account with wide powers, such as a system administrator. The resilience and cybersecurity page covers that side.

Where else does vendor AI come up in the framework?

In at least eight other places. Vendor AI is not only a section 2.9 topic, so a bank that buys its AI may want to read these too.

  • Risk tiers (section 3.1). Provenance is one of seven factors. A system can be built in-house, “fine-tuned from a third-party base model, or consumed as a fully external vendor service”, and “each origin carries a different visibility and control profile.” See the four risk tiers.
  • Risk classification (section 2.2). “opaque third-party models could increase uncertainty and reduce the bank's ability to oversee the use case.” Shared providers can also create “concentration or correlated risk” across separate use cases.
  • Model risk (section 2.5). Adaptive models, retraining or “vendor updates could alter behavior between periodic reviews and may justify more frequent monitoring or re-validation.” Validation is an independent check that a model works as intended. See model risk management for generative AI.
  • Data governance (section 2.6). The bank may assess where data came from, including “data embedded in third-party foundation models.” A foundation model is a large general model, such as a large language model, that vendors adapt for many tasks.
  • Privacy (section 2.7). “Contracts and technical configurations may clarify whether bank data can be used to train or improve models made available to other customers.” See data, privacy and compliance.
  • Performance monitoring (section 2.12). Monitoring may identify “upstream version changes” and other changes “that alter behavior without a traditional model release.”
  • Internal audit (section 2.15). Audit may need access to “vendor contracts” among other evidence, “rather than only management summaries.”
  • Reporting (section 2.16). Reviewing a bank's reporting may help supervisors understand its “reliance on significant providers.”

What does the existing third-party guidance already cover?

The base layer is OCC Bulletin 2023-17 of June 6, 2023. It sends banks the interagency guidance on third-party relationships from the OCC, the Federal Reserve and the FDIC, and it reaches AI vendors like any other vendor.

The bulletin says the guidance “applies to all banks with third-party relationships.” It replaced the OCC's 2013 third-party guidance (Bulletin 2013-29) and the 2020 FAQs (Bulletin 2020-10). It “outlines the third-party risk management life cycle” and “clarifies that not all third-party relationships present the same level of risk or criticality.” The life cycle has five stages. The plain explanations in the right column are ours.

Stage (2023 guidance)Put simplyWhat it can mean for an AI vendor (our example)
PlanningDecide why you need the vendor and what could go wrongWrite down the use case and its likely risk tier before the demo
Due diligence and selectionCheck the vendor before signingAsk about the underlying model, security, resilience and subcontractors
Contract negotiationPut rights and duties in writingNotice of model changes, incident reporting, data use, liability
Ongoing monitoringKeep checking after go-liveSample outputs, track error and complaint rates, watch version changes
TerminationBe able to leave in an orderly wayGet the data back, have a fallback process, know how long a switch takes

The 2023 guidance also covers governance around these stages: oversight, documentation and independent review. In our reading, section 2.9 of the AI framework adds AI-specific questions on top of these stages. It does not replace them.

What would the proposed 2026 third-party guidance change?

It is a proposal, not final guidance. OCC Bulletin 2026-46 of September 11, 2026 proposes guidance, with the Federal Reserve, FDIC and NCUA, that would replace the 2023 guidance and focus vendor oversight on the relationships that carry real risk.

The bulletin says comments are due 60 days after Federal Register publication. The proposal was published there on September 15, 2026 (91 FR 58536). A notice of November 18, 2026 extended the comment deadline from November 16 to December 16, 2026. The proposal text is dated September 10, 2026. Main points, all stated as proposals:

  • Risk first. Risk identification and assessment become the foundation. The agencies say the 2023 guidance “frequently has been interpreted in an overly broad manner.”
  • Four components. Identifying and assessing risks, oversight in proportion to risk, decisions on residual risk, and governance. The life cycle stages (due diligence, contract negotiation, ongoing monitoring, termination) remain under oversight.
  • No standard contract terms. “There are no generally applicable expected contract terms for third-party relationships—even for higher-risk relationships.”
  • Residual risk can be accepted. “The agencies do not expect banking organizations to eliminate third-party risk.” A bank may lack bargaining power to get custom terms, and the relationship may still be worth keeping.
  • Not enforceable. The proposal “does not set forth enforceable standards or prescriptive requirements.” The agencies may still act on violations of law or unsafe or unsound practices.
  • Responsibility stays with the bank. Using third parties “does not diminish” a bank's responsibility to meet its requirements “to the same extent as if the activities were performed by the banking organization internally.”

The proposal text does not mention AI. Whether a final version will absorb section 2.9's points on provider agents, AI incidents, liability and notice of new AI features is not known. In practice, the two fit together: the proposal says how much effort a vendor deserves, and section 2.9 says what is different when the vendor sells AI. For example, a small bank buying a standard AI tool with no room to negotiate could, under the proposal's approach, record that it accepted some residual risk and lean more on monitoring.

How does each section 2.9 consideration turn into a contract or oversight question?

Each consideration can be turned into one question for the contract and one for day-to-day oversight. The questions are our suggestions, not OCC text.

Section 2.9 considerationContract question (ours)Oversight question (ours)
Provider agent visibilityDoes the contract fix the agent's identity, permissions, data use and hand-off points, and let the bank limit them?Can we see each action the agent takes in our systems, and can we switch it off?
AI incidentsDoes the incident clause name hallucination, misfire, drift and manipulation, with notice deadlines?Who at the bank receives these notices, and what happens next?
LiabilityWho pays for model errors and customer harm? Is there an indemnity, and how large is the cap?Do we track losses and complaints linked to the vendor's model?
Change and exitMust the vendor give notice before adding AI features or changing the model?Do we retest after a change, and do we have a fallback if we leave?
Opaque provider chains (summary table)Must the vendor name the underlying model and its own providers?Do several of our use cases depend on the same model or cloud?

What could a bank ask an AI vendor?

The list below is our suggestion, not an OCC checklist. Each question is tied to the point in the framework it comes from.

  1. Which underlying model does the product use, and who provides it? (Summary table, row 2.9, opaque provider chains; section 3.1, provenance.)
  2. Will your AI agents act in our systems, and with what identity, permissions and data access? (Section 2.9, provider agent visibility.)
  3. Where must your agent stop and hand over to our staff? (Section 2.9, escalation limits.)
  4. Will you tell us about hallucinations, misfires, drift or manipulation, and how fast? (Section 2.9, AI incidents.)
  5. Who is responsible if the model's error harms a customer? (Section 2.9, liability.)
  6. Will you give notice before you add AI features or switch models? (Section 2.9, change and exit.)
  7. Will our data train models you offer to other customers? (Section 2.7, vendor data use.)
  8. Where is our data stored and processed? (Section 2.7, cross-border processing.)
  9. What testing and monitoring evidence can you share, including for each new version? (Sections 2.5 and 2.12.)
  10. Can our auditors see the evidence behind your reports? (Section 2.15, access.)
  11. If we leave, how do we get our data back and keep the service running? (Section 2.9, change and exit; 2023 guidance, termination.)

For how examiners may approach these topics during a review, see preparing for an OCC exam of your AI. Smaller banks can read what it means for community banks.

What does this mean for a company selling AI to banks?

The framework is written for OCC examiners, not vendors. But section 2.9 shows a vendor which topics its bank customers may raise, so a vendor can prepare answers.

In practice, that could mean a few things. Publish which model you use and who hosts it. Describe your agents' permissions in terms a bank can test. Offer an incident clause that covers hallucination and drift, not only downtime. Give advance notice of new AI features. Say plainly whether customer data trains shared models. Expect negotiation on liability, because section 2.9 says the bank “should consider” how contracts assign responsibility for model errors.

The framework also notes that a bank “may also monitor and potentially limit what provider agents do without impacting the relationship with the provider.” Put simply, a bank restricting your agent is a normal control, not a sign of distrust. If the 2026 proposal is finalised as written, it would let banks scale their checks to the risk a vendor poses, so a low-risk tool could face lighter review.

For the whole framework in one place see the section-by-section guide to Bulletin 2026-54, the sixteen components, or the OCC tracker.

If a vendor's AI makes a mistake, is the bank still responsible?

Generally yes, as far as the framework says. Section 2.4 states that accountability for managing the bank's risk generally remains with the bank even when development or hosting is spread across vendors. How losses are shared between bank and vendor is a contract matter, which section 2.9 raises under liability.

Does the OCC require specific clauses in AI vendor contracts?

No. The framework is a resource for examiners and does not establish supervisory expectations for banks. Section 2.9 describes things a bank may define or should consider, such as incident terms and liability. The 2026 third-party proposal also says there are no generally applicable expected contract terms.

Does section 2.9 cover AI features that appear in software we already use?

Yes. Under change and exit, section 2.9 notes that existing vendors may add AI features or change models in products the bank already uses. It says the bank may require notice and may reassess the relationship or retest the model when that happens.

Is a provider agent the same as a vendor chatbot?

Not exactly. Section 2.9 talks about provider agents that interact with bank systems and data. A chatbot that only answers questions may not act in bank systems, while an agent can take actions such as updating records. The considerations apply where the provider's AI can affect the bank's decisions, customers or data.

Is the 2026 third-party proposal in force?

No. OCC Bulletin 2026-46 of September 11, 2026 is proposed guidance with a request for comment. It was published in the Federal Register on September 15, 2026, and the comment deadline was extended to December 16, 2026. Until any final version is issued, the 2023 interagency guidance in Bulletin 2023-17 is the current text.

Does the 2026 third-party proposal mention AI?

The proposed guidance text does not mention artificial intelligence, AI or machine learning. It is a general rewrite of third-party risk management. The AI-specific points sit in section 2.9 of the November 2026 framework, which says they may supplement the existing third-party guidance.

Can a small bank that cannot negotiate with a large AI vendor still use it?

The framework does not forbid it. The 2026 proposal says agencies do not expect banks to eliminate third-party risk and that a bank may lack bargaining power for custom terms. It describes accepting some residual risk where the relationship is still beneficial, though this is a proposal, not final guidance.

  1. Super Intelligence: Supervisory Framework (full text, PDF) — Office of the Comptroller of the Currency, Nov 4, 2026 (occ.gov) · our page on OCC Bulletin 2026-54
  2. OCC Bulletin 2026-54: Super Intelligence: Supervisory Framework — Office of the Comptroller of the Currency, Nov 4, 2026 (occ.gov) · our page on OCC Bulletin 2026-54
  3. OCC Bulletin 2023-17: Third-Party Relationships: Interagency Guidance on Risk Management — Office of the Comptroller of the Currency, Jun 6, 2023 (occ.gov) · our page on OCC Bulletin 2023-17
  4. OCC Bulletin 2026-46: Third-Party Risk Management: Proposed Guidance and Request for Comment — Office of the Comptroller of the Currency, Sep 11, 2026 (occ.gov)
  5. Proposed Third-Party Risk Management Guidance (interagency text, PDF) — OCC, Federal Reserve, FDIC and NCUA, Sep 10, 2026 (occ.gov)
  6. Proposed Third-Party Risk Management Guidance; Extension of Comment Period (Federal Register) — OCC, Federal Reserve, FDIC and NCUA, Nov 18, 2026 (federalregister.gov)

When this rule moves, the next morning's brief says so

when one of these regulators moves, the next morning's brief says so · six sourced stories · 7 am PT · free

plus every regulation, bank and agent page update, the morning after · leave any morning