A provider of a system that prepares performance appraisals in HR departments builds on a general-purpose AI model and buys annotated training data from a service provider. For the technical documentation it needs information on the capabilities and limits of the model and on where the annotations come from. The model provider points to a generic model card, and after three reminders the data service sends a spreadsheet with nothing on annotation guidelines or quality checks. The conformity assessment comes to a halt. The provider carries full responsibility for a system whose building blocks it only partly knows.
EN 18286 requires providers, in clause 6.1.3, to consider when dealing with risks to the quality management system (QMS) that other parties have to supply information and assistance across the life cycle that the regulatory requirements depend on. The risk behind it: these deliveries fail to arrive or arrive too late.
What is behind it?
Hardly any provider builds an AI system entirely on its own. General-purpose AI models, pretrained components, purchased data sets, annotation services, external test labs and data centres all contribute to the result. The obligations under the AI Act still stay with the provider of the high-risk system. It therefore needs from its suppliers everything it cannot produce itself for documentation, assessment and operation.
What information is typically missing
- Origin and quality of data. Without information on origin, preparation and known biases, the requirements on data and data governance are hard to demonstrate.
- Capabilities and limits of a model. For the instructions for use and the risk assessment, the provider needs to know what a purchased model has been tested for and where it fails.
- Changes. A model or component is updated, and the news of it arrives late or never.
- Support when it matters. In a serious incident, the provider needs logs, analyses and contact persons at the supplier, and within the deadlines of Article 73.
Why it goes wrong so often
Supplier contracts often date from before the AI Act and cover performance and price. Information duties, deadlines for documents and cooperation in incidents are missing. Legitimate supplier interests add to this, such as protecting trade secrets, and so do diverging timelines: the supplier has no reason to plan around its customer’s conformity assessment.
Impact
Missing information tears holes in the technical documentation. The conformity assessment is delayed, the market launch slips, and in the worst case conformity cannot be demonstrated at all. In operation, the lack of support weighs even more: without the supplier’s cooperation, the provider cannot find the cause of an incident and cannot take an effective corrective action. Pointing to the supplier offers no protection towards the authority, because the provider obligations stay with the provider.
Practical examples
Generic model card. The example at the start is typical: a public model card describes the model for many purposes and for none in particular. For the risk assessment of a performance appraisal system, information on performance across groups of people and on known error patterns is missing.
Data vendor without proof of origin. A provider of a system for assessing loan applications buys training data from a data broker. As to origin, the broker points to “public sources”. The chain of custody that the provider wants to record for each data set stays empty.
Incident without cooperation. A provider of software that pre-screens applications for social benefits at public authorities uses a purchased text recognition component. After a serious incident it asks for the component’s logs. The manufacturer replies after two weeks with a reference to its support terms. The provider reports on time with incomplete facts and has to run the root cause analysis without the decisive data.
Effective measures
The AI Provider Policy from the AI Compliance Starter Kit addresses this risk in these places:
- Section 23 (Resources and security of supply): For every third-party component built in, a written agreement under Article 25(4) is concluded and kept per system with its contract status and follow-up date. Suppliers are selected and assessed against documented criteria, including their ability to meet regulatory requirements. An unsuitable supplier leads to a corrective action.
- Section 12 (Development and quality assurance): Model version and source of purchased models are tracked separately from the application.
- Section 15 (Data management): The origin of every data set is recorded, for externally sourced data together with the supplying party.
- Section 18 (Reporting of serious incidents): The report is filed on time even with incomplete facts, and missing information is supplied later.
- Section 22 (Control of documented information): Documents from suppliers are kept with their reference alongside the system they concern.
These ISO/IEC 42001 controls contribute as well:
- A.10 Third-party and customer relationships: Responsibilities are allocated between provider and suppliers, and suppliers provide suitable documentation.
- A.7 Data for AI systems: Origin and quality of data are recorded for purchased data sets as well.
- A.4 Resources for AI systems: Documenting resources shows which of them come from third parties.
Link to the AI Act
Article 25(4) of Regulation (EU) 2024/1689 requires the provider of a high-risk system and the third party that supplies AI systems, tools, services, components or processes for it to set out in a written agreement which information, capabilities, technical access and other assistance the provider needs for its obligations. For components released under a free and open-source licence the duty falls away, except for general-purpose AI models. Providers of such models make the elements of Annex XII available under Article 53(1)(b). Article 17(1)(j) also requires the QMS to govern communication with other operators. EN 18286 adds requirements on controlling the supply chain in clause 9.3 (→ EN 18286: Quality Management System for AI Providers).
Related QMS risks
- QMS-04 AI technologies in use overstretch the processes: Third-party models are hard to test when information from their provider is missing.
- QMS-07 Outage of software holding QMS documentation: The tool in which the QMS is kept usually comes from a third party as well.
- T04 Poor data quality: The threat at system level when purchased data goes in unchecked.
Sources
- EN 18286:2026, clause 6.1.3 c and clause 9.3 (supply chain)
- Regulation (EU) 2024/1689 (AI Act), Article 17(1)(j), Article 25(4), Article 53 and Annex XII
- AI Provider Policy, Sections 12, 15, 18, 22 and 23
- EN 18286: Quality Management System for AI Providers in the Cenedril wiki