Skip to content
ReportThe 2026 State of TPRMDownload

Managing AI vendor risk: What most TPRM programs are missing

ArticleJune 2, 202615 min read

Most vendor risk questionnaires were written before any of your vendors started running your data through AI. These questionnaires focus on data security basics like encryption, access control, and incident response, all of which still matter. However, none can tell you whether a vendor is training its AI on your customer records, sending your data to an AI company you’ve never heard of, or letting an AI agent take actions on your behalf.

That gap creates a kind of risk exposure that sits right in the blind spot of most assessment processes. This guide covers what AI vendor risk actually is, why standard vendor reviews miss it, how to assess and manage vendors that use AI, and which standards to follow along the way.

Key takeaways

  • AI vendor risk is the risk you inherit when a vendor uses AI on your data or in your systems. It applies whether the vendor sells AI as its main product or has added AI features to a tool you already use.
  • Standard vendor questionnaires weren’t built to catch it. They rarely ask how a vendor’s AI uses your data, and they assume the vendor’s system stays the same between reviews.
  • The AI companies behind your vendors are your fourth parties. Many AI features rely on the same few AI models and tools, so one incident can affect many of your vendors at once.
  • AI agents raise the stakes. When an AI can take actions, the question shifts from what it can see to what it’s allowed to do.
  • You don’t need a separate program for vendors that use AI. Add AI questions to your existing program, ask for proof instead of promises, and keep checking vendors after approval.
  • A mature program can answer a few basic questions quickly. It knows which vendors use AI, which AI companies they depend on, and which have agreed in writing not to train on your data.

What is AI vendor risk?

AI vendor risk is the risk your company takes on when one of your third-party vendors uses AI to handle your data, make decisions that affect you or your customers, or work inside your systems. That includes AI vendors whose main product is AI as well as ordinary vendors that have added AI features to tools you already use. It’s a subset of third-party risk, but it introduces problems that most third-party risk management (TPRM) programs weren’t designed to catch.

Two things make it different from ordinary vendor risk:

  • You can’t see how an AI system works. With a traditional vendor, you can check the controls they have in place, like who has access to your data and how it’s stored. With AI, the system’s behavior comes from the data it learned from and the internal settings it built during training, and neither can be inspected. Even the vendor may not be able to explain why their AI produced a specific answer.
  • A vendor’s AI can change after you’ve approved it, often without warning. A vendor can switch to a different AI model, retrain the one they have, or change the instructions it follows. The system you approved in March can behave differently by June, with no contract change and no notification.

The new risks that come with vendor AI

When a vendor uses AI, whether it’s the vendor’s main product or a feature added to an existing tool, it opens up six kinds of risk that a standard vendor review doesn’t cover:

1. Data use and training

If a vendor uses your data to train or fine-tune its AI, the information you share can influence the answers that AI gives other customers. Once your data is built into a model, it’s very difficult to remove, which creates a lasting data privacy risk. Get the vendor’s commitment not to train on your data in writing, then verify that they follow it.

2. Model provenance and subprocessors

You need to know which AI models a vendor uses and which company provides them. Most vendors don’t build their own AI. They use a model from a separate AI company, which means that company also handles your data. That makes it a subprocessor (a company your vendor relies on to process your data), with its own data-handling terms. Your vendor inherits those terms, and by extension, so does your company.

3. Automated decisions

If a vendor’s AI makes a decision about a person, such as a credit limit, a fraud flag, or whether a job application moves forward, your company remains accountable for that decision. That includes any bias in the outcome and your ability to explain how it was reached. Regulators don’t accept “the vendor’s AI made the decision” as a defense. Under the General Data Protection Regulation (GDPR), for example, people can ask for a human to review certain decisions made entirely by automated systems.

4. Accuracy and fabrication

AI tools can produce answers that sound confident but are incorrect, and some generate information that isn’t true at all. If a vendor uses AI to summarize, classify, or write content for you, inaccurate details can make their way into your work. The more your team relies on that output without reviewing it, the greater the risk.

5. Model-layer security

AI systems face types of attacks that standard vendor reviews don’t assess. In prompt injection, an attacker hides instructions in content the AI reads, such as a webpage or document, causing it to take actions it shouldn’t. In training data poisoning, corrupted data is introduced into what the AI learns from, affecting how it behaves later. In model extraction, an attacker reconstructs how an AI works by sending it a high volume of questions and analyzing the answers. AI systems can also expose sensitive information in their responses, which can lead to data breaches. Even a vendor with strong cyber risk controls isn’t automatically protected against any of these attacks.

6. Shadow AI

Shadow AI is the use of AI tools a company never approved, such as consumer chatbots, often with sensitive data. The AI vendors behind these tools were never reviewed and have no contract with your company, but they’re still third parties, and your data is exposed as soon as it’s submitted. Coverbase’s State of Continuous Monitoring report found that 30% of the apps running inside large companies had never been reviewed for risk, and AI tools were the largest group among them.

Why traditional vendor assessments miss AI risk

A standard vendor questionnaire is built around a fixed picture of what a vendor does with your data, including where it’s stored, who can access it, and how it’s protected. However, AI complicates these assessments in three ways:

1. It asks the wrong questions

Most questionnaires don’t ask whether a vendor uses your data to train or fine-tune its AI, which AI companies sit behind its product, or whether its system makes automated decisions about your customers. If the questionnaire doesn’t ask, the vendor doesn’t answer, and the risk never gets recorded.

2. It assumes the vendor’s system doesn’t change

Traditional assessments are based on the idea that you evaluate a vendor once and its system stays largely the same until the next review. AI models are updated constantly, and as AI adoption grows, vendors often release new AI features faster than an annual review cycle can keep up with. Coverbase’s report also found that under questionnaire-based review, it takes an average of 6 to 9 months for a customer to learn that a vendor has changed its subprocessors.

3. It stops at the vendor’s edge

Many AI features are built on top of an AI model from a separate AI company. That means your data may pass to another party that you never assessed and have no contract with.

All three gaps come from the same source. Traditional assessments rely on what vendors tell you at a single point in time, and that’s why more programs are shifting to evidence-based TPRM, which checks vendors directly and continuously instead of waiting for the next questionnaire.

How to prepare your TPRM program for vendors that use AI

You don’t need a separate program for vendors that use AI. If your company already has a third party supplier management program, you can extend it to cover AI. The two biggest changes are matching your review to the level of risk each vendor’s AI carries and asking every vendor for evidence rather than taking their word for it.

Sort vendors that use AI into higher and lower risk

Not every vendor’s AI carries the same risk. Base your risk scoring on what the AI actually does, then match the depth of each review to the score.

Lower risk vendors

A lower risk vendor uses AI for tasks like drafting internal documents or summarizing meeting notes. For these vendors, a short disclosure question at intake is usually enough.

Higher risk vendors

A higher risk vendor uses AI to make or influence decisions about your customers, such as approving loans or flagging fraud. For these vendors, set a documented trigger to reassess them whenever their AI model or its behavior changes, rather than waiting for an annual review date.

Ask every vendor for proof of its AI use

No matter the risk level, be sure to ask your vendors to back up their claims about their AI use with verifiable documentation. Here are four types of documentation you can request:

  • A data processing addendum that names AI subprocessors. This is the part of your contract that lists every company handling your data, and it should include the AI companies behind the vendor’s product.
  • An ISO 42001 certification or a NIST AI RMF self-assessment. These show the vendor manages AI risk against a recognized standard. Both are covered in more detail below.
  • Model cards or other AI documentation. These describe what an AI model was built to do, its training data sources, and its known limits.
  • Red-teaming results. These come from tests where security experts try to make the AI fail or misbehave on purpose, and they show how the vendor’s AI holds up under attack.

How to manage vendor AI after approval

A vendor’s AI can change after approval and vendor onboarding, which means the commitments made at intake only hold up if someone keeps track of them. Managing vendor AI over time comes down to three practices:

1. Tying contract renewals to AI commitments

Renewal is your best opportunity to update terms. Before you renew, confirm the vendor still meets the AI commitments in its current contract, such as not training on your data and disclosing the AI companies it relies on. Then, add any terms the original contract missed, like advance notice before the vendor changes AI models or adds AI features that use your data.

2. Reassessing when a vendor’s AI changes

Set clear triggers for a new review instead of relying on the calendar. Common triggers include the vendor switching AI models or AI companies, launching an AI feature that uses your data, giving its AI access to new systems, or reporting an AI-related incident. Any of these can change the risk you approved, even when the contract stays the same.

3. Tracking AI obligations for the life of the relationship

Commitments in vendor contracts such as data-use limits, notice requirements, and audit rights are easy to agree to, but they’re also easy to lose track of. Keep them in one place, such as an obligations tracker, assign an owner to each, and check them on a regular schedule. When a vendor misses one, treat it as a finding, the same way you would a failed security control. This ongoing monitoring is where many programs fall behind, especially as more vendors add AI. For smaller teams, managed TPRM services can be a practical way to keep up.

What agentic AI adds to third-party AI risk

Agentic AI refers to AI systems that take actions instead of only answering questions. They can browse the web, send messages, move data, complete transactions, and use other software tools on their own.

That changes the risk in a fundamental way. When a vendor’s product includes an AI agent, or when your company gives an AI vendor’s agent access to your systems, the agent can do anything its permissions allow, including things nobody anticipated when those permissions were granted. A mistake is no longer just a wrong answer. It can be a sent email, a deleted file, or a completed payment.

The questions that matter once an AI can act

With a traditional vendor, the key questions are about data, such as what the vendor can see and how it protects it. With an AI agent, the key questions are about authority, meaning what the agent is allowed to do.

  • What actions can the agent take? Get a specific list, such as reading files, sending messages, or approving transactions. The more an agent can do, the more damage a mistake or attack can cause.
  • What credentials does it hold? Find out which accounts, API keys, or logins the agent uses and what they give it access to. An agent using a broad admin account carries far more risk than one limited to a single tool.
  • What stops it from following hidden instructions? Attackers can plant instructions in a webpage or document the agent reads, tricking it into taking actions it shouldn’t. Ask what safeguards the vendor has in place against this.
  • Is a person involved in important actions? For actions with real consequences, like payments or changes to customer records, there should be a step where a human reviews and approves before the agent proceeds.

A useful rule is to treat an AI agent’s permissions the way you’d treat a new employee’s access. Give it only the access it needs to do its job, and keep a record of everything it does, because what you’re granting an agent is much closer to employee access than to a software license.

The fourth-party model problem

A fourth party is a company your vendor relies on to deliver its service, and AI adds an important new one in the AI company behind a vendor’s AI. Whether you work with an AI vendor or a tool that has added AI features, you’ve often taken on a dependency on an AI company you didn’t choose and can’t see in your contract.

That creates concentration risk. Many AI features across the software market run on a small number of AI models and the shared tools that connect to them. You can spread your business across dozens of vendors and still find that most of them depend on the same few AI companies or tools. When one of those has an outage, changes its policies, or gets breached, the impact can reach many of your vendors at once and put business continuity at risk.

The LiteLLM attack in March 2026 showed how risk moves through the AI supply chain. LiteLLM is a popular open-source AI gateway that many companies use to connect their software to AI models. When attackers slipped credential-stealing code into two of its releases, every vendor running those versions was exposed, along with customers who may never have heard of LiteLLM. With 15 to 20 million installs a week, the reach was enormous. Coverbase’s State of Continuous Monitoring report breaks down how the attack spread and what would have caught it sooner.

Knowing which AI companies and tools your third-party ecosystem depends on is now a core part of managing vendor risk. A strong TPRM program with fourth-party monitoring maps those dependencies, so when an incident like this happens, you can see right away which vendors are affected.

Figure 2.1. Nth-party dependency graph. Source: Coverbase, State of Continuous Monitoring report.
Nth-party dependency graph connecting an organization, its vendors and shared upstream services, including LiteLLM.

Figure 2.1. Nth-party dependency graph. Source: Coverbase, State of Continuous Monitoring report.

How to assess a vendor that uses AI

Add these questions to your vendor risk assessment process. For each one, ask the vendor for evidence behind the answer, the same way you would for any other control.

  • Does your product use AI, and for what specific functions? Start here before anything else. Even for an AI vendor, you need to know exactly which parts of the product use AI and how before you can assess the risk it carries.
  • Do you use our data to train, fine-tune, or improve any model? If the answer is yes, ask whether you can opt out and whether that commitment is in the contract. Confirm it also applies to any AI companies the vendor works with, not just the vendor’s own systems.
  • Which AI companies and subprocessors are involved, and where is our data processed? Any AI company behind the vendor’s product is handling your data too. Make sure each one is named in the data processing addendum.
  • Does your AI make or influence automated decisions about people? If it does, ask how the vendor tests for bias and how a decision gets explained to the person it affects. Regulators treat automated decisions about people as high risk.
  • What protects your AI from prompt injection, data leakage, and misuse? Ask for red-teaming results or independent security testing of the AI itself. A written policy alone isn’t enough.
  • How will you notify us when your AI model or its behavior changes? If the vendor won’t commit to notifying you, treat that as a risk on its own. Without that commitment, a vendor can retrain or replace its AI without telling you.
  • If your product includes AI agents, what can they do, what access do they have, and where does a person review their actions? Use the authority questions from the agentic AI section above to go deeper. An agent’s permissions deserve the same scrutiny as a new employee’s access.

0

What is AI compliance?

AI compliance is the practice of making sure the AI your company uses meets the laws, regulations, and industry standards that apply to it. This includes AI that reaches you through vendors, since your company can still be held responsible when a vendor’s AI breaks the rules. In practice, most companies build their approach to AI regulatory compliance around the established frameworks below.

AI governance frameworks: ISO 42001, NIST AI RMF, and the EU AI Act

Standards for managing AI risk are still new, but three frameworks have emerged as the main reference points. They serve different purposes, from vendor certification to legal requirements, so it helps to know which one applies when. Here’s how they compare:

Scroll to compare →

FrameworkWhat it isWhy it matters for vendor risk
ISO/IEC 42001An international standard, published in late 2023, that companies can be certified against.It’s the first certifiable standard for managing AI, similar to what ISO 27001 is for information security. A certified vendor has put a formal AI management system in place. Ask vendors for it the way you’d ask for a SOC 2 report.
NIST AI Risk Management Framework (AI RMF)A voluntary US framework, published in 2023, with a generative AI profile added in 2024.It gives companies a shared way to identify, measure, and manage AI risk. It isn’t certifiable, but it’s widely used in the US as a baseline for internal AI practices and vendor conversations.
EU AI ActA European Union law regulating AI, the first of its kind, taking effect in phases over several years.It bans certain uses of AI outright, sets strict requirements for high-risk uses like credit, hiring, and essential services, and applies lighter rules to lower-risk AI. It also assigns different obligations to companies that build AI and companies that use it, so an AI vendor and a vendor that has added AI may carry different responsibilities. If your company or your vendors operate in the EU, it affects which AI uses are allowed and what documentation you’ll need.

Sources: ISO/IEC 42001, NIST AI RMF, and EU AI Act.

What good management of AI vendor risk looks like

A program that has caught up to AI risk can answer a few questions quickly. For many programs, these are the gaps worth closing first:

Vendor AI inventory

A mature program knows which of its vendors are AI vendors, which have added AI to their products, and what each one uses AI for. That knowledge lives in an active inventory, not in a guess or a policy document, and it gets updated whenever a vendor changes its product.

AI dependency mapping

A mature program knows which AI companies and tools it depends on through its vendors. With data lineage tracking, it can also trace where its data goes and how it’s processed. That way, when an AI company or tool has an incident, the team can see right away which vendors are affected.

Contractual data-use commitments

As part of good data governance, a mature program has clear commitments from vendors that handle sensitive data on whether they can train AI on that data. Those commitments are written into the data processing addendum, apply to the AI companies the vendor works with, and have been verified rather than accepted on a sales call.

Trigger-based reassessment

A mature program doesn’t wait for an annual review to recheck third-party vendors whose AI makes important decisions. When a vendor’s AI model or its behavior changes, the program reassesses it and updates its vendor risk score.

An approved path for shadow AI

A mature program gives employees approved AI tools, from AI vendors it has already reviewed, along with a clear policy for using them. Banning AI tools on their own tends to push employees toward using them in secret, while an approved alternative actually reduces the risk.

Using AI in your own TPRM program

Everything above is about assessing vendors that use AI. Using AI to run your own third-party risk program is a separate topic, and the two often get mixed up. The most familiar version speeds up the paperwork: pulling control evidence from a vendor’s SOC 2 report, drafting first-pass assessments, and summarizing findings. That helps with volume, but it only moves the same questionnaire-based process along faster, and most TPRM platforms now offer some version of it.

The bigger opportunity is work your team can’t do today at all, because it would take more people, more access, or more data than any risk team has. With AI-driven application inspection, an AI agent gets read-only access to a vendor’s application and keeps reviewing its security settings, access controls, and integrations. Findings come from how the vendor is actually set up rather than from what it wrote on a questionnaire. When something changes, like a new integration or a change to how users log in, you find out without waiting for the next review cycle.

The caveat is that AI used to assess risk needs the same scrutiny you’d give any vendor’s AI. A TPRM platform that uses AI is itself an AI vendor you’re relying on. Every finding it produces should show its reasoning and its source, and a person should review it before it drives a decision.

To learn how other risk teams are handling the shift, read the State of Third-Party Risk Management in the Age of AI report. To find out which of your vendors use AI and what they depend on, book a demo with Coverbase.

Frequently asked questions

01What is the difference between AI risk and AI vendor risk?

The difference between AI risk and AI vendor risk is where the risk comes from. AI risk covers all the ways an AI system can cause harm, including bias, inaccurate answers, and misuse. AI vendor risk is the part of that risk your company takes on through third parties, when vendors use AI on your data, in their products, or inside your systems. That applies whether a vendor sells AI as its main product or has added AI features to an existing tool.

02Can a standard security questionnaire cover AI vendors?

A standard security questionnaire can’t cover AI vendors, or vendors that have added AI, on its own. Most were written for systems that don’t change much between reviews, and they rarely ask about training data, the AI companies behind a product, automated decisions, or what AI agents are allowed to do. You don’t need a separate process, though. You can extend your existing questionnaire with AI-specific questions.

03Is it safe if a vendor says they don’t train on our data?

A vendor saying they don’t train on your data is the right commitment to get, but it isn’t enough on its own. Make sure it’s written into the contract rather than promised on a sales call, confirm it covers any AI companies the vendor works with, and verify it the way you would any other control. Those AI companies may have different default policies than the vendor itself.

04What is shadow AI?

Shadow AI is employees using AI tools their company hasn’t approved, such as consumer chatbots, with company or customer data. Each of those tools comes from an AI vendor no one has reviewed, and the data is exposed as soon as it’s submitted. The most effective fix is usually an approved alternative and a clear policy, since banning AI tools tends to push employees toward using them in secret.

05Which AI framework should we map our vendor questions to?

The AI framework to map your vendor questions to all depends on what you need. If you want a certification you can ask vendors for, ISO 42001 is the one to use. If you want a framework to organize your own program, the NIST AI RMF is widely used in the US. If your company operates in the EU, the EU AI Act sets requirements you’re legally required to meet. Many programs use more than one, since they serve different purposes and fit together well as part of an AI compliance approach.

06How often should vendors that use AI be reassessed?

Vendors that use AI should be reassessed whenever their AI changes, not only on a fixed schedule. A vendor that switches AI models or AI companies, adds an AI feature that uses your data, or gives its AI access to new systems may carry different risk than the one you approved. Vendors whose AI makes important decisions about your customers need these change-based reviews most. Lower risk vendors can usually follow your standard review cycle, with a quick check on whether their use of AI has changed.

Share this post

Clarence Chio, headshot

Clarence Chio

CEO, Coverbase