Before relying on an AI agent vendor, the OCC expects a bank to plan the relationship, do due diligence scaled to risk, and negotiate a contract that protects it. Afterwards, it must monitor the vendor and be ready to exit. The bank stays responsible for what the agent does, as if its own staff had done it.
What does the OCC require, in plain words?
The OCC, with the Federal Reserve and FDIC, has issued interagency guidance on managing risks from third-party relationships. Its core idea is that using a third party does not reduce the bank's responsibility to operate safely and soundly and in compliance with law. Risk management should follow the life of the relationship:
- Planning: understand the risks before entering the relationship and how it fits the bank's strategy.
- Due diligence and selection: assess the vendor's ability to perform, its controls, its financial condition and its own reliance on others.
- Contract negotiation: secure terms on performance, data, audit, reporting, incidents, subcontracting and termination.
- Ongoing monitoring: watch performance and risk throughout.
- Termination: exit in an orderly way.
Across the lifecycle, the guidance expects board and management oversight, independent review, and documentation. The depth of each stage scales with the risk and criticality of the activity.
Who does it apply to?
Banks supervised by the OCC, and in parallel by the Federal Reserve and FDIC under the same guidance. It covers any business arrangement with a third party, which includes agent platforms, model providers reached through them, and connector and data suppliers. Our broader guide to third-party risk for agent platforms under the NIST AI RMF covers the voluntary framework's view of the same chain.
What does this mean when the vendor supplies an agent?
Agent vendors differ from most technology vendors in one way that matters to the OCC: their product acts inside the bank. That raises the stakes at each stage.
Planning must decide criticality early. An agent that drafts internal summaries is low risk. One that answers customers about their accounts, or changes records, may support a critical activity. Classify each intended use before choosing the vendor, because the classification sets how deep diligence must go.
Due diligence must reach behind the vendor. The agent runs on a model from a model provider the bank may never contract with. Diligence should find out who that provider is, how model changes are communicated, and what the vendor does when the model behaves differently. It should also establish who sets the agent's limits and where the record of its actions is kept.
Contracts must secure evidence and control. Standard technology contracts cover uptime. An agent contract also needs notice of model changes, the bank's right to set and change agent limits, a complete exportable record of actions and decisions, audit access, a ban on using the bank's data to train models without consent, and incident notice.
Monitoring must test behaviour, not just service levels. The bank should re-test agents on its own cases when the vendor or model changes, and track overrides, refusals and complaints. Vendor reports alone are not enough.
Exit must be practical. If the vendor fails, the bank needs to stop agents quickly, return work to staff and retain the records.
Take an illustrative case. A mid-sized bank selects a vendor whose agents handle card dispute intake. Planning rates the activity as significant because it affects customer rights under consumer protection law. Diligence reveals the vendor's model provider can change models without notice to the vendor's customers. The bank negotiates notice through the vendor, the right to cut the agent to drafting at any time, and full export of dispute records. When the first change notice arrives, the bank re-tests before restoring the agent's authority.
How do you meet OCC expectations for agent vendors?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Planning | Criticality rating per intended agent use | Planning memo with risk rating |
| Due diligence | Vendor and model provider assessed: controls, testing, change practice, data use | Diligence file |
| Subcontractor risk | Model provider and other subcontractors identified | Supply chain map |
| Contract terms | Change notice, limit control, record export, audit rights, data use limits, incident notice | Executed contract |
| Ongoing monitoring | Re-testing on the bank's own cases after changes; live behaviour tracked | Test results; monitoring reports |
| Oversight and accountability | Named relationship owner; reporting to management and board for critical uses | Reporting packs |
| Independent review | Risk or audit review of the vendor relationship | Review report |
| Termination | Exit plan: agents stopped, work returned to staff, records retained | Exit plan and test |
What will an examiner ask?
- How did you rate the criticality of each agent use, and why?
- Who is the model provider behind your agent vendor, and how do you learn of changes?
- Can you produce the agent's action records without the vendor's help?
- Who can change what the agent is allowed to do: the bank or the vendor?
- What did you do after the last change the vendor made?
- How would you exit if the vendor failed tomorrow?
What does the guidance not cover?
It does not set specific contract terms or approve vendors. It does not answer whether an agent is a model; model risk guidance does. It does not cover non-US regimes: banks operating in the EU face DORA, which may treat agent vendors as ICT third-party providers, as explained in are agent vendors ICT third-party providers under DORA, and the EU AI Act, discussed in the EU AI Act and third-party agents you buy. For the full map, see our guide to AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Is a model provider a third party if the bank has no contract with it?
It is part of the third-party relationship's supply chain. The guidance expects banks to understand subcontractors that matter to the activity, so identify the model provider and secure the protections you need through the vendor contract.
Does vendor certification satisfy due diligence?
It is useful evidence about the vendor's controls, but it does not show how an agent behaves on the bank's cases. Diligence should combine certification reports with the bank's own testing.
How often should agent vendors be reviewed?
In line with criticality, and at every material change. For agents, a model change is material, so reviews are event-driven as well as periodic.