Governance · AI governance and assurance · 04
AI vendor and model risk review
An AI vendor and model risk review assesses the third-party AI systems, models and AI-enabled services an organization relies on: what data they receive and retain, how they are trained and updated, what the provider commits to contractually, how outputs are controlled, and what happens when the provider changes or fails. It is needed before adopting an AI vendor, when existing vendors have added AI features to services already in use, when customers ask how client data is protected from use in model training, or when a regulator or contract requires documented third-party AI oversight. The review produces a vendor-by-vendor assessment with data handling, contractual and operational findings, the conditions that should apply to each, and a register entry that keeps the oversight current.
Scope
What the review covers.
- Data handling: inputs, retention, training use, residency, deletion
- Model provenance, updates, versioning and change notification
- Contractual terms: data use, confidentiality, liability, audit rights, exit
- Output controls, human review and logging
- Operational dependency and continuity if the vendor changes or fails
- Conditions, monitoring and register entry
Deliverables
What you receive.
| Deliverable | Form |
|---|---|
| Vendor and model assessment | Findings and rating per vendor |
| Contract conditions | Terms the agreement should include |
| Operational controls | Review, logging and dependency measures |
| Register entries | For the AI use register and vendor risk program |
Prerequisites
What we need from you.
- The AI vendors and features in scope, from the AI use inventory
- Contracts, terms of service and the vendor’s documentation
- A contact at each vendor for clarifications, where available
Related
Where this sits.
By David Selakovic · Published 2026-10-16 · Revised 2026-10-16 · Reviewed every 180 days
Questions
Asked before engaging.
Do large AI providers answer questions?
Their published terms and documentation answer most. Where they do not, the gap is recorded and treated as a risk decision.
What about AI features in tools we already use?
They are the most common finding: terms changed to permit training on customer data, or features enabled by default. The review catches both.
How does this relate to vendor security assessment?
Same discipline, with AI-specific questions on data use, model behavior and output control. The two are often combined.