Third-Party AI Risk Policy¶
Ready
Purpose. How you assess, contract for, and monitor AI you did not build — including AI features that appear inside software you already licence. For most organisations this is the majority of their AI exposure.
When to use it. At procurement, at renewal, and whenever a vendor announces new AI functionality in an existing product.
How to use it. The hard part is not the assessment; it is finding out that a vendor switched AI on. Build the trigger into contract renewal and vendor release-note reviews, and accept that you will discover some AI features after the fact.
Closes assessment gaps
This template addresses Q07, Q22, Q23, Q40 in the readiness assessment.
The template¶
1. Scope¶
Applies to any AI capability we consume rather than build: standalone AI products, AI features inside existing SaaS, APIs and foundation models, and outsourced services using AI on our behalf.
The invisible majority
Most organisations' largest AI exposure is AI switched on inside software they already bought. Your CRM, HR system, and support desk have all shipped AI features. Assume you have more than you know.
2. Vendor risk tiering¶
| Tier | Criteria | Assessment depth |
|---|---|---|
| Critical | High-risk use case, or personal / Restricted data, or customer-facing decisions | Full assessment + contractual controls + annual review |
| Significant | Confidential data, or a material business process | Standard assessment + contract terms |
| Limited | Internal data, non-critical | Lightweight screen |
3. Assessment questions¶
Ask before purchase, and record the answers.
The system
- What does it do, and what decisions does it inform or make?
- Are we the provider or the deployer under the AI Act? (Rebranding or fine-tuning can make you the provider.)
- What is its risk classification?
Data
- What data do we send, and at what classification?
- Is our data used to train their models? Can that be disabled contractually, not merely by a setting?
- Where is data processed and stored, and who else can access it?
- What is retained after termination?
Model & assurance
- What model underlies it, and are we told when it changes?
- What documentation exists — model cards, evaluations, bias testing?
- What accuracy and robustness evidence can they provide?
- Do they hold relevant certifications (ISO/IEC 42001, SOC 2)?
Compliance
- Can they evidence AI Act compliance for their role?
- Will they support our obligations — documentation, incident notification, audit?
- What is their incident notification commitment, and in what timeframe?
4. Contract clauses to require¶
| Clause | Why |
|---|---|
| AI Act obligation flow-down | You remain accountable for what you deploy. |
| No training on our data (or explicit consent) | Prevents your confidential data improving a shared model. |
| Notification of material model change | A silent model swap can change system behaviour overnight. |
| Documentation rights | You cannot evidence what you cannot obtain. |
| Audit or third-party assurance rights | Trust needs verification. |
| Incident notification within 24 hours for critical vendors; 72 hours otherwiseUnless: Match the tightest window that applies to you — if your own regulator gives you 72 hours, a vendor telling you at 72 hours is already too late. | Your own reporting clock may already be running. |
| Data location and sub-processor control | Transfer and access obligations. |
| Exit: data return and deletion | Avoids hostage data at renewal. |
5. Ongoing monitoring¶
| Activity | Frequency |
|---|---|
| Reassess Critical vendors | Annually for critical vendors, and at every contract renewalUnless: Immediately whenever a vendor announces new AI functionality in a product you already use. |
| Review vendor release notes for new AI features | Quarterly, including a review of vendor release notesUnless: Monthly if you have a large SaaS estate or an active procurement pipeline. |
| Confirm certifications still current | Annually |
| Review incidents involving the vendor | Each occurrence |
6. Recording¶
Every third-party AI system is entered in the AI System Inventory, marked as third-party, with vendor, tier, assessment date, and review date.
7. Review¶
| Version | Date | Owner | Approved by |
|---|---|---|---|
| 0.1 | [date] | [role] | [committee] |
Adaptation notes¶
- Small organisations: You will not renegotiate the contracts of large vendors. Focus on questions 4–7 (data handling), choose products with acceptable defaults, and document the residual risk you are accepting.
- Regulated sectors: Align with your existing outsourcing and operational resilience regime — AI vendors supporting important business services likely fall inside it already.
- Public sector: Fold these questions into your standard procurement framework so the assessment happens inside a process staff already follow.
Related¶
- AI System Inventory — Ready
- AI Governance Framework — Ready
- Control Library & Assurance Map — Ready
- Acceptable AI Use Policy — Ready
Not legal advice
These templates are a head start, not a substitute for professional judgement. Adapt them to your jurisdiction, sector, and risk appetite, and have qualified counsel review anything material before you rely on it.