Check out our upcoming events | Learn More
Articles

Third-party vendor risk management in an AI world (part 1)

Conventional third-party risk assumes the vendor stays put and the risk lives at the vendor. AI breaks both.

  • July 29, 2026
  • 10 Minute Read
Back to All Insights

Third-party risk management, as most organizations practice it today, rests on two quiet assumptions. The first is that a vendor is a relatively stable thing; assess it once and the picture you form holds until the next review cycle. The second is that risk lives at the vendor, that the company you contract with is the entity whose controls, data handling, and security posture you need to understand.

Vendors built on AI break both.

They break the first assumption because AI systems do not sit still. The model behind a vendor’s product can change continuously; retrained, swapped for a cheaper provider, or updated by an upstream lab with no version number, no contract amendment, and often no notice. The tool you validated in the spring may behave materially differently by the fall, and nothing in a conventional point-in-time assessment is designed to catch it.

They break the second because the risk frequently does not live at the vendor at all. It sits one or two layers down, at a foundation model provider you never assessed and that your vendor has little or no leverage over. Your data, and your exposure, flow through a supply chain the vendor may not fully control and that you cannot see into by questioning the vendor alone.

Nearly every complication in the pages that follow is a variation on these two breaks. And the failure is a quiet one. A program built for static, single-layer vendors will not break loudly when it meets AI. 

This piece sets out how to adapt a third-party risk program to that reality.

Mapping the Model Supply Chain, Not Just the Vendor

The first task is to stop thinking about an AI vendor as a single entity and start thinking about the chain behind it. Begin with provenance; is the capability built in-house, fine-tuned on an open-weight base such as Llama or Mistral, or simply an API call to a frontier lab? Each answer carries a different profile for data flow, residency, and liability, and the vendor rarely volunteers which one applies unless you ask directly. That question matters because a vendor’s “AI feature” frequently routes your data to a model provider that is itself an unassessed sub-processor. A fourth party your intake process was never designed to surface. Treating the feature as self-contained hides the entity where much of the real exposure sits.

Two further wrinkles make this chain harder to pin down. Vendors change inference providers for cost or performance reasons, swapping one model family for a cheaper one, and they do so without a contract amendment or any notice to you; your data flow and risk posture shift underneath you with no event in your process to catch it. 

The second is geographic.  Where the model actually runs may differ from where the application is hosted. A vendor’s residency commitments on its platform do not automatically extend to the location of inference, so the assurance you think you have about data staying in a given jurisdiction can quietly fail to cover the part that matters most.

Data-Use Questions Unique to AI

The single largest AI-specific data risk is training use; does the vendor or the upstream model provider use your prompts, documents, or outputs to train or improve their models? Default terms often permit exactly that unless you are on an enterprise tier or have explicitly opted out. But even a clean answer here is not the end of the inquiry, because “we don’t train on your data” rarely means the data disappears. Providers routinely retain inputs for abuse monitoring or debugging, often for thirty days or more, which is a separate data-exposure surface that a training-use assurance does nothing to close.

The exposure also extends to data you might not think of as your data. Embeddings, vector stores, and cached outputs are all derived from what you send, yet they frequently fall outside a vendor’s “we don’t store your data” claim, so it is worth asking specifically where those derived artifacts live and how long they persist. 

Finally, in multi-tenant inference there is the question of leakage, whether sensitive inputs can be memorized by a model or surface in another customer’s outputs. These are the questions a conventional data-processing assessment tends not to ask, and they are precisely where AI vendors differ from the SaaS tools that came before them.

Treating the Model as a Moving Target

Even a vendor you have assessed thoroughly is not the fixed quantity your process assumes. AI systems are non-deterministic; the same input does not reliably produce the same output, which quietly undermines repeatable testing and the whole premise of point-in-time assurance. The practical response is to build acceptance criteria around behavior envelopes rather than fixed outputs, accepting a range of acceptable behavior instead of a single expected answer.

Compounding this is drift. There is no version numbering for model behavior; the vendor, or the lab upstream of it, updates the model and your validated workflow now behaves differently, with no version bump you can point to. That is why it is worth defining, in advance, what counts as a material change and requiring the vendor to notify you when one occurs. 

It is also why the cadence of assurance has to change. An annual SOC 2 decays quickly against a vendor shipping changes weekly, so the program should be built around reassessment triggers (a change of model provider, a change in training use, a new region, the addition of agentic capabilities) rather than a fixed calendar date that bears no relationship to when the risk actually moved.

Agentic and Action Risk

When a product stops merely processing data and starts taking actions on your behalf, the nature of the risk changes. With agents and tool use, the blast radius is no longer confined to data handling; it extends to actions executed inside your environment. That makes the scope of permissions granted to the tool a first-order concern, what it can touch, whether it has been over-permissioned, and whether you retain the ability to constrain it. 

Agentic tools also introduce a failure mode that has no real analogue in traditional software; prompt injection. A tool that ingests untrusted content such as emails, web pages, or documents can be manipulated into taking actions it was never meant to take, so it is reasonable to require evidence that the vendor tests for this specifically. 

And because the tool is now acting rather than just responding, auditability becomes an assessment item in its own right: if you cannot reconstruct what an agent did and why, you cannot govern it.  You should confirm that logging of an agent’s decisions and actions exists before you rely on it. 

Contractual Levers, and the Limits of Them

Most of the problems above are only manageable if you can see them coming, which makes notification rights the foundational contractual term; the right to be told about model version changes, sub-processor or model-provider changes, changes in training use, and changes in residency. Without those rights, the “ever-changing” nature of AI vendors is simply invisible to you. Alongside them, an opt-out from training use should be guaranteed in the contract rather than left to a settings toggle that can be reset or reinterpreted.

It is important to be honest about how far these levers reach. Your vendor almost certainly cannot impose your requirements on a provider like OpenAI, Anthropic, or Google; it is a small customer of a very large lab operating on non-negotiable terms. This flow-down ceiling caps how much assurance the vendor can genuinely pass through to you, and it is worth naming as a structural feature of the market rather than treating it as a failing on the vendor’s part. 

The same realism applies to liability. There is genuine exposure around intellectual property and copyright when training data flows into outputs, and around errors when a model’s output is relied upon in a decision. Meaningful accuracy SLAs for AI output remain rare. Knowing where the contract cannot protect you is as valuable as knowing where it can.

Assurance Artifacts Worth Demanding, and How to Scope Them

The familiar attestations still have a role, but they need to be read more carefully in an AI context. A SOC 2 report is useful only if its scope actually covers the AI features rather than just the underlying platform, and that distinction is easy to miss. ISO/IEC 42001, the emerging standard for AI management systems, is worth asking for even where few vendors hold it yet, and alignment with the NIST AI RMF serves as a useful signal of program maturity. 

Beyond the certifications, the more revealing artifacts are often the technical ones: model or system cards, data-flow diagrams that show where inference happens and what gets logged, and evidence of red-teaming or prompt-injection testing. These tell you how the vendor actually operates, which is frequently more informative than a certificate that may or may not extend to the component you care about.

The Concentration Risk Most Programs Miss

Assessing vendors one at a time hides a risk that only becomes visible at the portfolio level. Many of your vendors may quietly depend on the same two or three foundation labs, which means a single model provider’s outage, policy change, security incident, or rate-limiting decision can strike your entire vendor portfolio at once. No individual vendor assessment will ever reveal this, because the correlation lives across your inventory rather than within any single relationship. 

The remedy is to map the underlying model dependency across the whole portfolio and to know not just that a given vendor uses a given lab, but how much of your overall footprint rests on that same handful of providers so that systemic exposure is something you can see and plan for rather than discover during an incident.

Designing the Program and Its Governance

The program itself needs to reflect how different AI vendors are from the rest of your third parties. The cleanest way to do that is to branch the intake: a vendor that uses AI, or processes your data through a model, should route to an AI-specific questionnaire and risk tier rather than through the standard flow. Risk should be rated by use case as much as by vendor, because the same vendor can be trivial in one deployment and high-stakes in another depending on the data it touches and the decisions it informs. 

A large part of the “ever-changing landscape” problem is not new vendors at all but existing ones, a CRM or ticketing tool that switches on an AI feature and begins routing your data to a model with no new procurement event to trigger review. Inventory and monitoring therefore have to catch capability changes in tools you have already onboarded, not just watch the front door. 

And underpinning all of it, someone has to own AI risk internally with enough authority to act. The program should assess lock-in from the outset: whether, if you leave, you can actually extract your data and the derived artifacts built from it.

The Regulatory Overlay

Finally, each AI use case should be classified against the regulatory regimes that apply to you. The obligations that matter depend on your industry, the jurisdictions you operate in and serve, the data you process, and the kinds of decisions your AI capabilities touch. A hiring or credit-scoring use case sits in a different regulatory posture than an internal productivity tool; a firm handling health or financial data inherits sector rules that a retailer does not; and the moment prompts cross a border, data-protection and transfer obligations come into play that have nothing to do with AI-specific law at all. 

The right instinct is not to memorize a single framework but to ask, for each use case, which regimes could plausibly reach it and who inside the organization owns that determination. What makes this harder than a normal compliance mapping is that the ground is still moving. Major AI regulation is still being written, phased in, and in some cases amended and deferred, even as organizations try to build against it. The EU AI Act is the most prominent example, but far from the only one, with a growing patchwork of national and sub-national rules emerging in parallel.  Deadlines shift, definitions get refined, and supporting technical standards often lag the law itself. 

The practical consequence is that a regulatory position taken today can quietly go stale, so this belongs on a watch list with a named owner rather than being treated as a one-time classification exercise. The habit to build is to confirm the current state of any regime before you rely on it, so the organization notices when the rules change rather than learning it from an enforcement action. 

None of this calls for tearing down the third-party risk program you already have. It calls for extending it to a class of vendor that refuses to sit still and refuses to sit at a single layer. The through-line of everything above is that AI turns vendor risk from a snapshot into a moving picture; the model changes, the sub-processors change, the capabilities change, and much of what matters lives one or two layers deeper than the company whose name is on the contract. 

A program built for static, single-layer vendors will not fail loudly when it meets that reality; it will simply keep issuing clean assessments while the risk quietly moves out from under them. The organizations that manage this well will be the ones that stop treating an AI assessment as a gate to pass once and start treating it as a relationship to monitor, with clear ownership, the right to know when things change, and honest recognition of where their leverage ends. Get those instincts in place and the specific techniques, frameworks, and regulations will keep falling into position around them, even as all three keep changing.

Those instincts are where it starts. The second part of this series takes them the rest of the way, into the sequence for standing up an AI vendor-risk program the business will actually use, one that survives well past its first quarter.

AI vendors challenge traditional third-party risk programs because models, providers, data flows, and capabilities can change without notice. Learn how risk leaders can adapt their assessments, contracts, monitoring, and governance to manage a fast-moving AI supply chain.