AI supplier management for ISO 42001

Many organisations do not build all of their AI themselves. They buy models from OpenAI , Google or Microsoft , or they implement software with AI components. ISO 42001 says that you are also responsible for the AI you buy. On this page we explain how to approach that.

Why does ISO 42001 require AI supplier management?

The risks of purchased AI are just as real as those of AI you build yourself.

Inheriting bias. A purchased model can contain bias. If you use it, the problem is yours.

Black box. You often do not know exactly how a supplier built its AI. What is in the training data?

Responsibility. Towards your customers and regulators, you are responsible, not your supplier.

Supply chain risk. If your supplier has an incident, you have an incident.

The EU AI Act makes this even more urgent. You must be able to show that your entire AI supply chain is in order.

Which suppliers count?

First map out which suppliers provide AI or are connected to AI:

Direct AI suppliers:

Suppliers of AI models or services (OpenAI , Google Cloud AI , Azure AI )
SaaS products with AI functionality
AI consultants and system integrators

Indirect suppliers:

Data suppliers (if their data is used for AI)
Cloud providers where AI runs
Development tools for AI

Embedded AI:

Software you already use that contains AI
Hardware with AI components

Many organisations discover during the inventory that they have more AI suppliers than they thought. That handy chatbot in your CRM? That is an AI supplier.

The supplier management process

Managing AI suppliers follows six steps. Below you can see what each step involves.

Step	What do you do?	Result
1. Inventory	Register all AI suppliers	Supplier register
2. Classification	Rank suppliers by criticality and dependency	Priority per supplier
3. Risk assessment	Assess operational, AI-specific and compliance risks	Risk profile per supplier
4. Due diligence	In-depth investigation of critical suppliers	Questionnaires, documentation, references
5. Contractual safeguards	Record agreements on transparency, performance and liability	Contract with AI clauses
6. Monitoring	Continuously track performance, incidents and compliance	Periodic review reports

We elaborate on each step below.

Tip: Most organisations underestimate the number of AI suppliers. Start broad at step 1. That handy chatbot in your CRM or the smart search function in your HR tool is also an AI supplier for which you are responsible.

Step 1: Inventory

Create a register of all AI suppliers:

Name of the supplier
What they supply
Which AI system it concerns
Contractual relationship
Contact person

Step 2: Classification

Not all suppliers are equally critical. Classify based on:

Criticality of the AI:

High: AI that makes decisions about people or supports critical processes
Medium: AI that performs important but not critical functions
Low: AI for supporting, non-critical functions

Dependency:

High: Difficult to replace, deeply integrated
Low: Easy to switch

Focus your control efforts on the highly critical, highly dependent suppliers.

Step 3: Risk assessment

For each supplier, assess the risks:

Operational risks:

Availability and uptime
Performance and scalability
Continuity (financial stability, exit strategy)

AI-specific risks:

Transparency about how the AI works
Bias and fairness
Data practices (privacy, security)
Responsibility for incidents

Compliance risks:

Does the supplier comply with relevant legislation?
Does the supplier hold its own certifications (ISO 42001, ISO 27001)?
How does the supplier deal with the EU AI Act?

Step 4: Due diligence

For critical suppliers, you carry out more in-depth research:

Questionnaires. Send a detailed questionnaire on AI practices, security and compliance.

Request documentation. Model cards, datasheets, audit reports, certifications.

Check references. Speak to other customers.

Contract review. What do the terms say about liability, data ownership and compliance?

Step 5: Contractual safeguards

Record agreements in the contract:

Transparency:

Supplier informs you of significant changes to the AI
Supplier shares relevant documentation

Compliance:

Supplier complies with relevant legislation (EU AI Act, AVG)
Supplier cooperates with audits where necessary

Performance:

SLAs for availability and response time
Measurable quality criteria

Incidents:

Duty to report AI-related incidents
Cooperation in investigations

Liability:

Who is liable for AI errors?
Indemnity for third-party claims

Exit:

How do you get your data back?
How do you migrate to another solution?

Large tech suppliers (OpenAI, Google, Microsoft) have standard terms that are difficult to change. Your room for negotiation is limited. Understand what you accept and document the residual risks.

Step 6: Monitoring

Supplier management is not a one-off task. Monitor continuously:

Performance: Do they meet the SLAs?

Incidents: Have there been any problems?

Changes: Has the AI changed?

Market: Is there news, reviews or incidents involving the supplier?

Compliance: Does the supplier continue to meet the requirements?

Plan periodic reviews, at least annually for critical suppliers.

Annex A controls

ISO 42001 Annex A contains specific controls for supplier management:

A.10.2 Supplier policy. You have a policy for dealing with AI suppliers.

A.10.3 Supplier risk assessment. You assess, select and monitor AI suppliers for risks.

A.10.4 Customer relationships. If you supply AI yourself, you identify customer requirements and manage the relationship.

A.10.5 Providing components. If you supply AI components to others, you provide adequate information.

The challenge with large tech suppliers

The reality: many organisations use AI from large tech companies. OpenAI, Google, Microsoft, Amazon. These parties have millions of customers and standard terms.

What you can do:

Understand the terms you accept
Use enterprise versions with better terms where possible
Document which risks you accept
Monitor for incidents and changes
Plan for alternatives (avoid vendor lock-in)

What you usually cannot do:

Negotiate individual contracts (usually)
Force full transparency about how their AI works
Obtain guarantees on specific performance

Accept that residual risks exist. Document them. Explain to the auditor how you handle them.

Practical example: using OpenAI

Your organisation uses the OpenAI API for a customer-facing chatbot.

Inventory:

Supplier: OpenAI
Product: GPT-4 API
Application: Customer service chatbot
Criticality: High (customer contact)

Risk assessment:

Operational: Dependent on OpenAI uptime; the API may change
AI-specific: No control over model updates; possible bias; hallucinations
Compliance: OpenAI is US-based; data leaves the EU unless an EU endpoint is used

Mitigation measures:

Fallback to a human agent when problems occur
Content filtering on output
Monitoring of quality and incidents
Use the EU endpoint where available
Inform users that they are talking to AI
Regular evaluation of alternatives

Documentation:

OpenAI’s usage policies and data handling
Internal risk acceptance by management
Monitoring dashboards and alerts

Supplier management and the EU AI Act

The EU AI Act places obligations on the supply chain. If you use high-risk AI, you must be able to show that the entire chain is compliant.

What the AI Act requires:

Traceability along the supply chain
Documentation from the AI supplier
Responsibility when problems occur

Practical implication:

Ask suppliers for:

Technical documentation of the AI
Evidence of compliance (or a plan to achieve it)
Cooperation with conformity assessment

This will become more important in the coming years. Suppliers that do not cooperate will become less attractive.

ISO 42001 supplier management checklist

Have all AI suppliers been inventoried?
Have suppliers been classified by criticality?
Has a risk assessment been carried out for each critical supplier?
Has due diligence been carried out?
Are there contractual agreements on AI?
Is monitoring in place?
Have periodic reviews been scheduled?
Have residual risks been documented and accepted?