AI Supply Chain Risk: The Gaps Your Contracts Miss

Most organizations have a vendor risk management program. Most of those programs were designed for traditional software vendors — SaaS applications, cloud infrastructure providers, managed service providers. They were not designed for AI tools, and the gaps are significant.

AI vendors introduce risk categories that standard questionnaires do not address: model governance, training data provenance, output reliability, bias and fairness, and the contractual protections that govern what the vendor can do with your data to train their models. If your vendor risk program has not been updated to address these categories, you have blind spots.

The Risks Standard Questionnaires Miss

Model governance: How does the vendor manage changes to the underlying model? If the model is updated, how are you notified? How do you evaluate whether the update affects your use case? Standard questionnaires ask about change management for software — they rarely ask about change management for AI models, which can change behavior in ways that are difficult to detect.

Training data provenance: What data was the model trained on? Does it include data that could create legal exposure — copyrighted material, personal data, data obtained without appropriate consent? Standard questionnaires do not ask these questions, and most vendors do not volunteer the answers.

Output reliability and hallucination risk: What is the vendor's documented approach to managing hallucination and output reliability? For AI tools used in consequential decisions — customer communications, legal documents, financial analysis — this is a material risk that standard questionnaires do not address.

The Contractual Gaps

The most significant contractual gap in most AI vendor agreements is the training data clause. Many AI vendors include provisions that allow them to use customer data to train or improve their models. For organizations handling sensitive data — customer PII, financial records, health information, proprietary business data — this is a material risk that must be addressed contractually.

Specifically, your AI vendor agreements should address: whether the vendor can use your data to train their models (and if so, under what conditions); what data is retained after your contract ends and for how long; what happens to model weights or fine-tuned models trained on your data if you terminate the relationship; and what audit rights you have to verify compliance with these provisions.

What to Add to Your AI Vendor Assessments

Update your Tier 1 vendor questionnaire to include AI-specific questions: Does the vendor use customer data to train or improve their models? What is the vendor's approach to model versioning and change notification? How does the vendor test for bias and fairness in model outputs? What is the vendor's incident response process for AI-specific failures — hallucinations, model drift, adversarial inputs?

For vendors deploying AI in high-risk use cases — hiring decisions, credit decisions, healthcare triage — also ask about conformity assessments, human oversight mechanisms, and EU AI Act compliance status if you have European operations.

The Governance Implication

AI supply chain risk is not just a vendor management problem — it is a governance problem. Your organization needs a policy that governs which AI tools can be deployed, under what conditions, and with what level of review. Without that policy, individual teams will adopt AI tools without security review, and your AI supply chain risk will grow faster than your ability to manage it.

The organizations that manage AI supply chain risk effectively are the ones that treat it as a first-class risk category — not an afterthought to their existing vendor risk program.

For a broader look at how to build the governance program that sits behind your vendor risk process, see our blog post on [building an AI governance program with ISO/IEC 42001 and the NIST AI RMF](/blog/building-an-ai-governance-program-iso-42001-nist-ai-rmf). And for a deeper dive into the full range of AI supply chain risks — including shadow AI, sub-processor dependencies, and prompt injection — see our blog post on [AI supply chain risk: what your vendors are not telling you](/blog/ai-supply-chain-risk-what-your-vendors-arent-telling-you).