Zum Hauptinhalt springen
ISO/IEC 42001 · Annex A · AI control

A.10 Third-party and customer relationships

Updated on 6 min Reviewed by: Cenedril Editorial
ISO/IEC 42001EU AI Act

A logistics company plugs a major provider’s language model into its customer communication. A service provider handles the integration, and a third partner supplies the data for fine-tuning. When the system gives a customer wrong customs information, all three point at each other. A.10 asks that, for every AI system, it is settled in advance which party is responsible for which part, how suppliers are steered and which customer expectations the organization’s own approach takes into account.

Hardly any AI system is built by a single party. Models, data, computing power, integration and operation are spread across several players. A.10 makes sure the organization keeps its own accountability and that risks are shared appropriately. The control looks in both directions: towards suppliers whose deliveries flow into the organization’s systems, and towards customers to whom the organization itself supplies AI.

What does the standard require?

Annex A.10 of ISO/IEC 42001 contains three controls. The implementation guidance sits in Annex B.10.

  • Allocate responsibility (A.10.2). The organization makes sure that responsibilities in the life cycle of its AI systems are distributed between itself, partners, suppliers, customers and other third parties. It documents every party involved with its role, such as who supplies data, who provides models or algorithms and who develops or uses the system, and determines their responsibilities. If it supplies a system to a third party itself, it develops that system responsibly and can provide the necessary documentation. Where personal data is involved, it clarifies whether it acts as controller, joint controller, processor or several of these; ISO/IEC 27701 offers suitable controls for this.
  • Steer suppliers (A.10.3). A process ensures that services, products and materials obtained from suppliers fit the organization’s own responsible approach. This ranges from a single data set, algorithm or model through software libraries to a complete AI system, used on its own or built into a product. Selection, requirements and monitoring depend on the type of supplier, what it delivers and the risk. The organization documents how bought-in systems and components are integrated into its own systems, requires corrective action if a delivery does not work as intended or has unwanted effects on people, and makes sure the supplier delivers appropriate and adequate documentation.
  • Consider customers (A.10.4). If the organization supplies products or services involving AI, its approach takes customer expectations and needs into account. These show up as product requirements, contract clauses or terms of use and can differ from one customer relationship to the next. The organization clarifies which responsibility lies with the provider and which with the customer. If it identifies risks arising from the customer’s use, it can treat them by giving the customer suitable information, for instance the limits of the domain in which the system is valid.

In practice

Tier suppliers by risk. A model that prepares decisions about people inside a high-risk system needs closer management than a translation tool. For each tier, decide which documents must be available before the contract, which commitments are agreed in writing and how often the assessment is repeated.

Secure documents and cooperation by contract. Ask for a system description, details on training data and performance limits, instructions for use and security information. Agree that the supplier reports incidents and significant changes and cooperates in providing information to authorities. A rule for third-party intellectual property, trade secrets and source code received through the supply chain has also proven useful: access only for named people, classified storage, return or deletion once the purpose has ended.

Demand corrections and follow them up. If a bought-in system does not work as promised or disadvantages groups, the organization requires a remedy from the supplier and tracks the case as a corrective action. If no remedy follows, use is restricted or ended. The standard explicitly allows working with the supplier on the solution.

Give customers the limits. Whoever supplies AI hands over, along with the system, the limits of the validated domain of use, the measured accuracy and the split of responsibilities. This helps customers treat their own risks and protects the organization against uses the system was never meant for.

Typical audit evidence

For A.10, auditors typically expect:

KPI

Share of AI systems in the AI register for which every party involved is documented with its role and responsibility

The target is 100%. Typical gaps are subcontractors of the main supplier and data suppliers who delivered only once.

Supporting KPIs:

  • Share of AI suppliers assessed within the defined review cycle
  • Number of open correction requests to suppliers past their deadline
  • Share of systems supplied to customers with a handed-over description of the limits of use

The AI Act (Regulation (EU) 2024/1689) distributes duties along the value chain and expects the parties involved to cooperate. Anyone who supplies an AI system, tool, service, component or process for a high-risk system concludes a written agreement with the provider under Article 25(4) covering the necessary information, capabilities and technical access. Components made available under a free and open-source licence are exempt, but this exemption does not cover general-purpose AI models. Article 25(1) sets out when a distributor, importer or deployer becomes a provider itself, and Article 25(2) then obliges the original provider to cooperate. Providers of general-purpose AI models make available to downstream providers, under Article 53(1)(b), the information those providers need for their own duties.

Importers and distributors check under Articles 23 and 24, before placing on the market or making available, whether the provider has met its obligations. For deployers, Article 26 remains decisive: they use the system in accordance with the provider’s instructions for use and inform the provider of risks and incidents (→ EU AI Act).

A.10 extends internal responsibility beyond the boundaries of the organization:

Sources

Frequently asked questions

Who is responsible when a bought-in model produces wrong results?

That depends on the allocation A.10.2 asks for. Towards affected people and authorities, the organization remains accountable for its own use, even if the fault lies with the supplier. It requires corrective action from the supplier and restricts use if no remedy follows.

Which documents should I ask an AI supplier for?

Annex B.10.3 asks for appropriate and adequate documentation and points to technical documentation and user information for that. A system description, details on training data and performance limits, instructions for use, security information and commitments to report incidents and changes have proven useful. For high-risk systems, add whatever you need to meet your deployer obligations under Article 26 of the AI Act.

Does A.10.4 apply if we only use AI internally?

A.10.4 applies when the organization supplies products or services involving AI to customers. If you only use AI internally, you can exclude A.10.4 in the statement of applicability with a justification. A.10.2 and A.10.3 usually remain relevant, because hardly any system works without suppliers.

No Cookies!

This wiki collects nothing, bakes nothing and leaves nothing behind. There's nothing to consent to. Privacy doesn't get better than this.