Last chapter was about finding the AI already inside your walls. This one is about the AI you never built, never trained, and don't control — but that your organization is fully exposed to anyway, because it's sitting inside a vendor contract you signed for something else entirely.
The Assumption That No Longer Holds
Most third-party risk programs were built for a world of static software. AI vendors break that assumption completely.
The model underneath a tool you assessed in January can be silently swapped, retrained, or fine-tuned by March — and your questionnaire has no way of knowing. The frameworks, checklists, and review cycles your organization relies on were designed for software that stays still. AI doesn't stay still.
The Vendor Risk You Inherited Without a New Contract
Third-party AI risk doesn't arrive through a single, obvious door. It enters through three distinct channels — each with its own level of visibility and control.
Direct AI Vendors
Tools procured specifically because they're AI — at least flagged for review during procurement. These are the visible ones.
AI-Upgraded Incumbents
Vendors you've used for years that quietly added AI features, entering through a door your process doesn't watch. No new contract. No new review.
Fourth-Party AI
The AI your vendor's vendor uses — with no contract, no visibility, and no leverage on your end. The most invisible risk of all.
Why "They Passed Our Security Review" Isn't Enough
Standard vendor questionnaires were not designed with AI in mind. Two critical questions are almost never asked — and their absence leaves significant exposure unaddressed.
Question 1: What happens to our data?
After the AI processes your data — is it used to train the vendor's models? Retained for "product improvement"? Or fully isolated to your instance? Most questionnaires never ask. Most contracts never specify.
Question 2: Who is accountable?
When the model gets something wrong — a biased output, a hallucinated fact, a harmful recommendation — who bears responsibility? The vendor? Your organization? The answer is rarely written down anywhere.
The Three Layers of Third-Party AI Exposure
Understanding where your exposure actually lives requires mapping the full supply chain of AI capability flowing into your organization.
The deeper you go in the supply chain, the less visibility you have — and the more your risk depends entirely on your vendor's own governance practices, which you may have no right to audit.
Building the Map: What to Actually Do
Governing AI you didn't build requires deliberate, structured action across your vendor lifecycle. Here is the practical playbook.
Re-open your vendor register with an AI-specific lens
Ask in writing: has any AI capability been added since your last review? Don't rely on vendors to volunteer this information — require it explicitly.
Add AI-specific clauses to renewal cycles
Data-use restrictions, model-change notification requirements, and audit rights should be standard inclusions at every contract renewal going forward.
Segment your vendor population by AI risk tier
Not just by spend or criticality — by the nature and depth of AI involvement. A low-spend vendor processing sensitive decisions via AI may carry more risk than a high-spend infrastructure provider.
Ask about the fourth-party layer explicitly
Who are your vendor's AI subprocessors? This question alone will surface risks that no standard questionnaire currently captures.
Set a review cadence tied to AI's actual pace of change
Not your default annual cycle. AI models change faster than annual reviews can track. Quarterly touchpoints for high-risk AI vendors should be the baseline.
The AI Vendor Risk Lifecycle
Governing third-party AI is not a one-time assessment — it is a continuous cycle that must be embedded into your existing vendor management processes, running at the pace AI actually changes rather than the pace your legacy program was designed for.
Key Contract Clauses You're Probably Missing
When renewing or negotiating vendor contracts where AI is involved, these are the provisions that matter most — and that most standard templates currently omit entirely.
Data-Use Restrictions
Explicit prohibition on using your organization's data to train, fine-tune, or improve the vendor's models — unless you have affirmatively consented in writing.
Model-Change Notification
Contractual obligation for the vendor to notify you before making material changes to the underlying AI model — with a defined notice period and your right to re-assess.
Audit Rights
The right to audit or request third-party audits of the vendor's AI systems, data handling practices, and subprocessor relationships — not just their security controls.
Accountability Allocation
Clear language specifying who bears responsibility when the AI produces harmful, biased, or incorrect outputs — and what remediation obligations the vendor carries.
Subprocessor Disclosure
Requirement to disclose all AI subprocessors (fourth parties) and to notify you of any changes to that list — mirroring the GDPR subprocessor model applied to AI.
Toolkit for This Chapter
This pillar doesn't have a dedicated artefact in the toolkit yet — vendor and third-party AI risk is the next template planned for the register.
The vendor and third-party AI risk register template is actively in development. When published, it will provide a structured framework for capturing AI-specific vendor attributes, risk tier classifications, contract clause status, and fourth-party subprocessor mapping — all in a single, auditable register.
What Comes Next
The next chapter moves from vendors back inward: data governance and consent in the age of AI.
This Chapter Covered
The three entry points of third-party AI risk, why standard security reviews fall short, and the five concrete actions to build your vendor AI map.
The Core Insight
You are exposed to AI you never built, never assessed, and don't control. Governing it requires new questions, new contract clauses, and a faster review cadence.
Coming Up
Data governance and consent in the age of AI — turning the lens back inward to examine how your own data practices hold up under AI's demands.
Borrowed Risk: Governing AI You Didn't Build