Blog

Why Traditional Vendor Risk Assessments Miss AI Adoption

How to strengthen third-party risk programs as AI use evolves.

On May 22, 2026, Microsoft updated its Data Protection Addendum (DPA) and set up a separate notification track for AI. If a new subprocessor supports AI functionality, customers now get 30 days’ notice before it starts to handle their data, instead of the six months they used to get.

Not all vendors are alike, and we cannot expect them to send timely communications outlining how AI is being baked into their workflows, products, and services.

Additionally, those same vendors may not be quick to share what they are doing with your data. The review cycle you built for onboarding doesn’t trigger when a vendor you already have starts using AI, which means the first real question is how to find out it happened.

Where AI shows up in vendors you already onboarded

There are four places, and only the first is obvious.

In the product, announced. The vendor ships an AI feature and tells you about it, either in an email, a patch note, or on their marketing page.

In the product, unannounced. ENISA’s July 2026 procurement guidance makes the point directly: AI risk extends past products a vendor markets as AI-enabled. AI gets embedded in unrelated products or used during development without ever surfacing in a security review, and there’s frequently no contractual obligation to tell you how AI is used in the product development lifecycle. ENISA’s multilayer framework for AI cybersecurity goes a step further here, mapping threats like model poisoning, prompt injection, and data extraction that don’t show up in a standard vendor questionnaire because they’re specific to how the model itself can be attacked, not how the product is marketed.

In the subprocessor chain. Your Software-as-a-Service (SaaS) vendor routes data through a model provider, which runs on an inference host on some cloud infrastructure, so you’re into fourth- and fifth-party territory. The list you were given at signing goes stale the moment they switch providers, and that can happen between assessments, before a reassessment catches it.

In your service providers’ operations. Your law firm, benefits administrator, and outsourced support desk have no integration to inspect, so you have no visibility into how AI is used as part of the business operations that touch your data to meet their service obligations to you. Thomson Reuters surveyed more than 1,500 professionals across legal, tax and accounting, corporate and government roles in 2026 and found organization-wide AI usage at 40%, nearly double the year before, with a majority of individuals reporting they use publicly available tools like ChatGPT. As a result, your vendors are using AI in ways you can’t see unless you’ve performed a vendor assessment that identifies how they’re using it as part of their service operations. The worst part is that some of these vendors may have employees using shadow AI to deliver services using your data because of poor AI governance.

Why annual vendor risk assessments miss AI adoption

Vendor reviews are built around events like onboarding, renewal, and an annual refresh, and a vendor adopting AI may not show up on that calendar because it happens on their release schedule instead.

Between those events, two things change: the model behind a feature may be swapped, which changes the output, the data handling, and sometimes the processing region; and the policies requiring notice of that change lag behind it.

A clean security review, a signed DPA, and a completed AI questionnaire still leave you unable to answer which model backs the feature you bought. That isn’t evasion, because the person answering your questionnaire works in trust and compliance, while the answer lives with a product team that considers it an implementation detail.

AI is introducing a wide variety of new risks to the business, and staying diligent with vendor assessments and reassessments, including ad hoc reviews, can significantly reduce risk.

Per Verizon’s Data Breach Investigations Report, breaches involving a third party now account for 48% of all breaches in 2026, up 60% year over year, while employee use of shadow AI tripled to 45%, spiking data leakage.

It’s therefore essential that vendor risk organizations stay diligent about performing vendor risk assessments throughout the vendor lifecycle to identify risks introduced by AI usage. Organizations should also prioritize an AI-specific risk assessment now if the business or vendor risk team hasn’t performed one in 2026, given how ubiquitous the technology has become. Going forward, questionnaires should include AI as a standing domain in both the inherent risk and due diligence assessments.

The pace of AI advancement, combined with how ubiquitous it has become, means businesses can’t wait for the next scheduled reassessment to find out how their vendors are using it.

Third-party risk frameworks for AI, and where each one fits

These are standards and frameworks that shape program structure, not documents you hand a vendor to fill out. Use them to decide what “good” looks like before you decide how to ask for it.

NIST SP 800-161r1, for program structure. Updated November 2024. Its supply chain control family works well as a standing set of supplier requirements you paste into contracts rather than negotiating from scratch each time.

NIST AI RMF, for AI-specific risk categorization. Built around four functions, Govern, Map, Measure, and Manage, that sit at a different altitude than 800-161r1. It’s the framework most likely to already be sitting somewhere in a vendor’s own governance program, which makes it a useful shared language when you’re asking them to categorize risk the same way you do.

ISO/IEC 27036, for the relationship lifecycle. It covers planning through to termination, and the operation stage is the one that matters here, since it’s also the easiest stage to treat as a placeholder between renewals. Paid standard, worth it mainly if you’re already ISO 27001.

ENISA’s multilayer framework, for AI-specific threat modeling. Free, and useful once you need to go past procurement-level questions into how model-level attacks such as poisoning, prompt injection, and extraction should shape your technical due diligence rather than just your contractual language.

ISO/IEC 42001, for evidence you accept instead of asking. Writing down what a certification lets you skip is the fastest thing I know for shrinking cycle time. A vendor holding ISO/IEC 42001 may be able to skip your AI governance section entirely.

Questionnaire instruments for asking vendors about AI

These are the actual instruments you send to a vendor, or the certifications that let you skip sending one.

SIG and SIG Lite, for the questionnaire you’re probably already sending. Shared Assessments added an AI domain covering model risk, training data provenance, and human oversight starting with recent releases. The update path here may just be turning on questions inside a tool you’ve already licensed, rather than adopting something new.

The CSA AI-CAIQ, for AI-native depth. It pairs with the AI Controls Matrix, which runs to 247 control objectives across 18 domains with roughly 320 questions behind it. What you want is the handful that change your answer: whether your data is used for training, which subprocessors and model providers sit in the path, and what happens to your data when you leave.

SOC 2, for evidence you accept instead of asking. Like ISO 42001 above, a current SOC 2 report with the right trust service criteria can let you skip re-verifying controls a vendor has already had audited, though it won’t cover AI-specific practices unless the report scope was written to include them.

The laws underneath all of this, and why none of them map to a single jurisdiction

There is no single AI law that governs your vendors, wherever you or they operate. It is tempting to assume one jurisdiction’s compliance covers you everywhere, and it does not.

The EU has gone furthest with the EU AI Act, where whether a vendor is a provider or a deployer changes what they owe you. The US, by contrast, has no federal AI law, though Colorado’s AI Act is the most comprehensive state-level version so far.

Canada’s proposed Artificial Intelligence and Data Act (AIDA) died along with Bill C-27 in 2025 and has not been replaced. The federal government has since shifted toward targeted policies and a national AI strategy instead.

Elsewhere, the pace is accelerating. Brazil’s PL 2338 is moving through its legal system with a risk-tiered structure modeled partly on the EU’s. China’s generative AI measures have been binding since 2023, and South Korea’s AI Basic Act took effect on January 22, 2026.

These laws do not map cleanly onto a single questionnaire question, so ask the vendor directly which of these regimes applies to them and how they are classified under each. Their answer, or their inability to answer, tells you whether they have mapped their own regulatory exposure, and whether their compliance claims (a SOC 2 report, an AI Act classification, a certification) actually cover how they use AI with your data specifically, rather than describing a compliance program that predates that use or was scoped around something else.

Here are five recommendations for updating the third-party risk program you already have:

1. Run an AI re-baseline across vendors you’ve already onboarded. This isn’t a full reassessment. It’s a short, standardized set of questions sent to every vendor above your existing threshold (typically your critical or high-data-sensitivity tier). Ask which parts of the service now involve AI, which model providers or subprocessors sit in that path, and whether your data is used for training. Expect some vendors not to have a ready answer or one that may be unsatisfactory.

2. Add a second tiering axis. Contract value and regulated data still matter, but exposure now belongs alongside them: specifically, whether your data ends up inside an AI system at this vendor, and whether anyone there can tell you where. A small consultancy running your case files through a public model can outrank a large vendor that simply emails you a monthly report.

3. Assign clear ownership of contract terms, not just the vendor relationship. DPAs and subprocessor notices need a reader when they arrive. Microsoft’s 30-day subprocessor notification track, for example, turns this from an annual chore into an ongoing responsibility.

4. Add a notification clause to your next renewal. Require notice before the model behind a feature changes materially, with the right to pause the service until the vendor provides sufficient information about how it uses AI, the controls it has in place, and how it handles your data when used in conjunction with AI models.

5. Make sure your termination clause covers derived data, not just original data. Most contracts only require the vendor to return or delete the data you originally gave them. But if that vendor ran your data through an AI system, it may now exist in other forms, including embeddings, fine-tuning datasets, logged prompts, or generated outputs that were built from your data but aren’t the data itself. ENISA calls this out as a common gap: these derivatives are rarely named in the contract, which means a vendor can fully comply with deletion requirements and still retain a model trained on your information indefinitely. When assessing a vendor or renewing a contract, check whether “data” is defined broadly enough to include these derivatives, and ask directly whether anything persists in a trained model after termination.

If you’re revisiting your program at this level, it’s also worth checking whether your Vendor Risk Management Maturity Model (VRMMM) or another maturity model has kept pace. A program scored as “mature” in 2023 may not have had AI in its control library at all, meaning that score may be as stale as your vendor list.

Key Takeaways

  • Your first problem is detection, not assessment. A short, AI-risk-framework-aligned follow-up questionnaire to existing vendors matters more than a full reassessment, because nothing in a standard onboarding-shaped review cycle triggers automatically when a vendor you already work with ships a new AI feature.
  • Check four places to determine whether AI has entered a vendor’s operations and become a supply chain risk: announced product features, patch or version notes, the subprocessor chain (sometimes disclosed via a link on the vendor’s marketing page or on their trust center), and proactive outreach to service providers with no system-to-system integration who may still access your data through channels like email.
  • AI risk management frameworks, AI-domain questionnaires, and AI laws serve three different functions: NIST, ISO, and ENISA shape your program’s structure; SIG and the CSA AI-CAIQ help you assess a specific vendor’s risk, while SOC 2 reports offer independent audit evidence. AI laws, such as the EU AI Act, impose mandatory legal obligations regardless of internal frameworks or vendor assessments.
  • Exposure deserves a place alongside contract value as a tiering axis and the re-baseline should start with vendors you onboarded before any of this AI risk existed.

Vendor risk used to be mostly a question of who you did business with. Now it’s just as much a question of what your vendors are building and or analyzing from your data, which models and subprocessors sit behind that work, and how much of it happens outside your line of sight. The work now needs to be closing that gap.

Explore our work in action

Our Enterprise Advisory and Management Consulting capabilities span cybersecurity, AI, and adjacent domains with comprehensive services that extend beyond conventional boundaries.

Ready to advance your business goals?

However you bring us in, the work is delivered by practitioners. Engage our Advisory & Management Consulting team to shape the strategy and roadmap, then hand the keys to Managed Services.