From Sidhant Tamrkar | Product & Market Analysis
The Bank Third-Party Risk Proposal, Read Clause by Clause for AI Suppliers
On this page
The banking agencies' new vendor risk proposal fills 9 pages of the Federal Register. It does not contain the words "artificial intelligence" once. It also states that there are no generally expected contract terms, even for higher-risk relationships. For an AI supplier selling into banks, the bank vendor risk AI question is no longer what the checklist says. It is how a bank sorts you into a risk tier, because that tier now decides almost everything else.
Key takeaways
- The proposal would replace the 2023 interagency guidance entirely. The OCC, Federal Reserve, FDIC and NCUA published it on 15 September 2026, and comments close on 16 November 2026.
- Contract expectations fall from a 17-topic list to none. The 2023 guidance walked through 17 contract considerations, while the proposal says no contract terms are generally expected, even for higher-risk relationships.
- The text never names artificial intelligence or models. AI suppliers are governed by the same magnitude-of-harm and likelihood test as a cloud host or a compliance tool vendor.
- Lighter oversight is available only to vendors banks rate lower risk. A model in a lending, fraud or customer-facing workflow is likely to fail that test and still face full diligence, monitoring and exit planning.
Short answer for AI vendors: the proposal would let banks tailor oversight to each relationship's assessed risk. Banks could also accept residual risk within their stated appetite and drop expected contract terms. It does not exempt AI suppliers. A model that could cause a legal violation, material financial harm or a significant disruption still sits in the higher-risk tier, with full diligence attached.
What the third-party risk management proposal replaces
On 11 September 2026 the four federal banking regulators jointly proposed new third-party risk management guidance. It was published in the Federal Register on 15 September at 91 FR 58536. Stinson LLP notes it is the first proposed interagency guidance on the subject to include the NCUA. Credit unions now sit inside the same text as banks.
The proposal would rescind the June 2023 interagency guidance and two supplemental documents. Those are the July 2024 joint statement on bank and fintech deposit arrangements and the May 2024 community bank guide. Footnote 5 also asks whether other guidance or interpretive letters should go at the same time.
The agencies' stated reason is that the 2023 text was read too broadly. Skadden's 17 September client note describes checklist-style examination, heavy use of "should," and a focus on "critical activities" that implied "an impossible goal of risk elimination". The proposal swaps that for what the agencies call a principles-based approach.
Two companion documents arrived the same day. The Federal Reserve proposed a separate guide for traditional community banking organizations, which Crowell & Moring describes as covering Fed-supervised banks under $30 billion. The agencies also issued a joint statement on community banks and their core service providers, covering transparency, contract features and technology.
| Point | 2023 interagency guidance | 2026 proposal |
|---|---|---|
| Organising concept | Critical activities, with illustrative characteristics | Magnitude and likelihood of harm, relationship by relationship. |
| Inventory | All third-party relationships tracked | Banks "may decide not to maintain extensive inventories" for limited-risk relationships. |
| Due diligence | 14 topic headings | 6 named topics plus risk management practices, "as applicable". |
| Contracts | 17 contract considerations, from audit rights to regulatory supervision | No generally applicable expected terms, even for higher-risk relationships. |
| Subcontractors | Diligence on reliance on subcontractors | Subcontractor use alone "does not typically create an independent third-party relationship". |
| Residual risk | Implied elimination, per commentators | Banks may accept risk within their risk appetite. |
| Where 2023 still wins | A vendor knew exactly which clauses to pre-draft. Under the proposal every bank can ask for something different, so negotiation cost may rise for smaller suppliers. | |
Counts of 2023 headings are from the 2023 Federal Register text (88 FR 37920). Commentator characterisations are attributed to Skadden and Crowell & Moring. The final row is this post's judgement, not a finding.
Whether your AI product is a third-party relationship at all
The proposal's definition is deliberately broad. A third-party relationship is "a business arrangement between a banking organization and an entity or individual" that provides products, services or other activities supporting the bank. A model API, an agent platform, a transcription tool and a fraud scoring vendor all fit inside it.
Technology appears explicitly. The text says third-party relationships let banks access "technological advances," and it gives "cloud processing, cloud storage, and compliance tools" as examples. Models and AI do not appear, so an AI vendor is read through the same lens as any software supplier.
The written agreement question in footnote 9
Footnote 9 asks whether the guidance should apply only to third parties under a written agreement. For AI suppliers this is the most consequential question in the document. A large share of model access starts on click-through online terms, often bought on a corporate card by a single team.
If the agencies narrowed scope to written agreements, usage that never reached procurement could fall outside formal oversight. I think that would be the wrong outcome for AI specifically. The highest-risk AI use inside a bank is often the unreviewed kind, as the evidence on what shadow AI breaches cost already shows.
Your model provider is probably a subcontractor, not a separate relationship
The proposal states that "the use of subcontractors alone does not typically create an independent third-party relationship". If you build an agent product on a frontier model, the bank's relationship is with you. The model lab becomes your subcontractor, and your contract is where the bank's oversight of it lives.
That raises the stakes on your own upstream terms. Whatever your model provider will not commit to on retention, training use or notice of change, you cannot pass on to a bank. A side-by-side of what the major model providers commit to in their terms is the starting point for that gap analysis.
The higher-risk test AI suppliers will be sorted by
The proposal opens its risk section bluntly: "Not all third-party relationships present the same level of risk". Banks assess each relationship on two axes, the magnitude of harm and the likelihood that the harm occurs. Higher risk attaches where failure carries a material likelihood of a non-trivial violation of law, material financial harm, or significant operational disruption.
The text names one category outright. Core processing relationships are "likely to be assessed by most banking organizations as higher risk." Everything else, including every AI product, is left to the bank's own assessment.
Where an AI product lands on that test
My read is that placement follows the workflow, not the technology. A drafting assistant used internally with human review is a plausible lower-risk relationship. A model that sets credit limits, flags suspicious activity or answers customers in writing touches fair lending, anti-money-laundering or consumer protection law directly.
Those are the obligations where a model failure becomes a non-trivial legal violation. Selling the same model into both uses means you may be a lower-risk vendor to one team and a higher-risk vendor to another inside the same bank. Price and package for the higher tier, because that is the one that sets your sales cycle.
Due diligence for AI vendors: six named topics and the gap between them
The proposal lists what a bank may consider before selecting a third party. The named topics are financial condition, business experience and key personnel, staffing capabilities and training, legal and regulatory compliance, insurance coverage, and information security and cybersecurity, "as applicable". It also points at the effectiveness of the third party's own risk management, policies and internal controls.
For lower-risk relationships, banks may rely on "less-detailed due diligence information or on public or alternative sources". Skadden adds that incomplete diligence would not "automatically preclude engagement," and that shared assessments and certifications count. A certification, in Skadden's words, "would not itself transfer the bank's accountability".
What none of the six topics covers for a model
None of the named topics asks how a model performs, how it changes, or how its outputs are tested. Those questions used to land in model risk management. The April 2026 revision of that guidance, SR 26-2 and OCC Bulletin 2026-13, excludes generative and agentic AI from scope, according to CRA International's analysis.
So a large language model vendor sits in a gap. It is outside the revised model guidance and unnamed in the vendor guidance. I expect banks to fill that gap themselves by stretching "risk management practices" and "legal and regulatory compliance" into evaluation questions. What a validation file for a language model should hold is set out in the LLM validation file built on the old model risk letter.
Contract terms: from 17 listed topics to none expected
This is the clause that changed most. The proposal states: "There are no generally applicable expected contract terms for third-party relationships," and it applies that to higher-risk relationships too. For lower-risk relationships, a bank "may determine that standard form contracts are sufficient".
The 2023 guidance had walked through 17 contract considerations. They included the right to audit and require remediation, ownership and license, subcontracting, foreign-based third parties and regulatory supervision. Examiners and procurement teams often treated that list as mandatory. The proposal removes the list, not the subjects.
| Proposal text | Where it appears | What it means for an AI supplier |
|---|---|---|
| Confidentiality and information security | Contract negotiation | Prompt and output retention, training use and deletion will be asked about first. |
| A service level agreement "can define performance measures" | Contract negotiation | Uptime is easy to commit to. Accuracy and drift thresholds are what higher-risk buyers want. |
| Use and oversight of subcontractors | Cross-cutting topics | Name your model provider and hosting region, and say how you are told of model changes. |
| Insurance, indemnification and limitations on liability | Cross-cutting topics | Caps set at fees paid will be tested on any use the bank rates higher risk. |
| Operational resilience, backed-up data and "alternative back-up providers" | Cross-cutting topics | A documented fallback model or provider becomes a selling point. |
| Termination, data retention and destruction, joint intellectual property | Termination | Fine-tuned weights and evaluation sets built on bank data need an owner in writing. |
The audit right that is no longer listed
The proposal refers to audits "conducted by the third party (or an independent entity hired by the third party)". It does not set out a bank audit right or a clause granting regulators access to the vendor, both of which the 2023 list covered. For an AI vendor that looks like relief, and in the short run it is.
I would not strip audit and regulator access clauses from your paper because of it. Large banks will keep asking, and a vendor that offers them unprompted shortens review. Your SOC 2 report and its AI-relevant controls carry more weight now, which is the subject of what auditors accept as AI evidence in a SOC 2.
Liability caps are now a negotiated risk acceptance
The proposal treats insurance, indemnification and limits on liability as a cross-cutting oversight topic. It also lets banks accept residual risk within their risk appetite. Put together, a low liability cap is no longer a compliance blocker. It is a residual risk that a bank's governance has to own.
That shifts the conversation from legal to risk teams. Expect higher-risk buyers to ask for a cap sized to plausible harm, not to annual fees. The patterns vendors use in practice are compared in the review of AI agent liability caps and remedies.
Ongoing monitoring, subcontractors and the model you do not own
Ongoing monitoring under the proposal aims to "confirm the quality of a third party's controls". It should also catch problems such as "material or repeat audit findings" and "data loss". The examples are performance reports, periodic visits, testing, complaint reviews and, for certain large third parties, agency examination reports. Frequency scales with risk.
The text says subcontractor use "may heighten risks," which is the closest it comes to the fourth-party problem. For AI suppliers the fourth party is usually the model. When a provider deprecates or silently updates a model, a bank's monitoring of you is only as good as your monitoring of them.
Exit planning is where AI vendors are least ready
Termination planning covers cost, transition time, data retention and destruction, and jointly developed intellectual property. Operational resilience points banks at "alternative back-up providers that can provide an easy transition of services". For a model vendor, easy transition means prompts, evaluation sets and fine-tuned assets that move with the customer.
Crowell & Moring also flags that OCC Bulletin 2002-16 on foreign-based service providers would be rescinded with no replacement text. That removes a specific checklist for vendors processing bank data abroad. It does not remove the data residency questions, which banks will now ask under general information security, as covered in the comparison of AI data residency rules in the EU and India.
Where this reading is weakest
Three objections deserve a straight answer, and one of them may simply win.
Guidance is not law, and this is only a proposal
The release calls the guidance "non-binding," and the proposal can still change after 16 November. Crowell & Moring expects finalisation is "unlikely to occur before 2027". Federal Reserve Governor Michael Barr dissented, warning the proposals would "reduce safe and sound operations". A future agency leadership could reverse direction, and vendors who rewrite their paper for a lighter regime could be early.
Banks may not use the flexibility at all
Large banks built their vendor programmes on the 2023 structure and have every reason to keep them. Supervisors will still review those programmes. The likely outcome at the largest institutions is that nothing changes for AI vendors, because AI uses land in the higher-risk tier anyway. Smaller banks are where the real difference shows up.
There is also a source caveat. The finding that the proposal never names AI comes from a full-text read of the published notice, not from an agency statement. The agencies could argue, fairly, that technology-neutral guidance is the point, and that naming AI would date the text within a year.
Frequently asked questions
What is the 2026 third party risk management proposal?
It is proposed interagency guidance from the OCC, Federal Reserve, FDIC and NCUA, published in the Federal Register on 15 September 2026 at 91 FR 58536. It would rescind and replace the 2023 interagency third-party guidance with a principles-based approach that tailors oversight to each relationship's assessed risk. Comments are due by 16 November 2026, and the guidance is non-binding.
Does the bank third-party risk proposal cover AI vendors?
Yes, though it never names them. The guidance covers any business arrangement that supports a bank's products, services or activities, which includes model APIs, agent platforms and AI software. The text mentions technology, fintechs and cloud services, but not artificial intelligence or models. AI suppliers are assessed on the same magnitude and likelihood of harm test as any other vendor.
Does the proposal require specific contract terms with banks?
No. The proposal says there are no generally applicable expected contract terms, even for higher-risk relationships, and that standard form contracts may be enough for lower-risk ones. It still discusses confidentiality, information security, subcontractors, service levels, insurance, indemnification, limits on liability, resilience and termination. Banks will keep negotiating those subjects, tailored to risk.
What due diligence will banks ask an AI vendor for?
The named topics are financial condition, business experience and key personnel, staffing and training, legal and regulatory compliance, insurance coverage, and information security and cybersecurity, plus the vendor's own risk management and controls. None covers model performance directly, so banks using AI in higher-risk workflows will likely add their own testing and change management questions.
When are comments on the third-party risk guidance due?
Comments are due by 16 November 2026. Each agency runs its own docket: OCC-2026-0793 for the OCC, OP-1881 for the Federal Reserve, RIN 3064-ZA58 for the FDIC and NCUA-2026-1684 for the NCUA. The Federal Reserve's companion guide for traditional community banking organizations carries the same deadline.
Is a model provider a third party or a subcontractor to the bank?
If a bank contracts with an AI application vendor that runs on a model provider, the provider is generally a subcontractor of that vendor. The proposal says subcontractor use alone does not typically create an independent third-party relationship. The bank oversees it through the vendor's contract and controls. The exception is a bank that contracts with the model provider directly.
Where to start before 16 November
If you sell AI into banks, map each product to the workflows it touches. Mark which ones could plausibly cause a legal violation, material financial harm or a significant disruption. Write that map down. It is the assessment your buyer will run on you, and arriving with it first changes the meeting.
Then decide whether to comment. The footnote 9 question on written agreements and the footnote 15 question on listing high-risk characteristics both shape how AI gets tiered. A short letter from a supplier explaining how model access is actually bought is the kind of evidence the agencies rarely receive.
References
- OCC, Federal Reserve, FDIC and NCUA, Proposed Third-Party Risk Management Guidance, 91 FR 58536 to 58544, 15 September 2026. Used for all quoted proposal text, footnotes and docket numbers.
- OCC, News Release 2026-77: Agencies seek comment on proposed third-party risk management guidance, 11 September 2026. Used for the non-binding status and the core service provider statement.
- OCC, Federal Reserve and FDIC, Interagency Guidance on Third-Party Relationships: Risk Management, 88 FR 37920, 9 June 2023. Used for the 2023 contract and due diligence topic counts and critical activities language.
- Skadden, Arps, Slate, Meagher & Flom, US Federal Banking Agencies Propose Revised Third-Party Risk Management Guidance, 17 September 2026. Used for the agencies' rationale, certification language and the Barr dissent.
- Crowell & Moring, From Process to Principles: Federal Regulators Propose a New Third-Party Risk Management Framework, 1 October 2026. Used for the community bank guide threshold, OCC Bulletin 2002-16 and timing.
- Stinson LLP, Banking Agencies Propose Revisions to Third-Party Risk Management Guidance, 23 September 2026. Used for the NCUA inclusion.
- CRA International, Model Risk Management Guidance SR 26-2 in the Era of AI, July 2026. Used for the generative and agentic AI scope exclusion.
Last verified 10 November 2026. The weakest source in this post is the SR 26-2 scope point, which rests on CRA International's reading rather than the agencies' own text. The proposal is open for comment and its wording may change in final form.
Related reading