A provider of an AI system that assesses risks in applications for health insurers keeps its entire quality management system (QMS) in a cloud document management service: policies, approvals, test reports, technical documentation. One Monday the service is unreachable after an attack on its operator, and on Thursday it still is. On Tuesday, a reasoned request from the market surveillance authority arrived. The evidence exists, yet nobody can get to it.
EN 18286 names exactly this case in clause 6.1.1 as an example of a risk to the QMS: documentation sits in a software solution that a third party controls and that can fail. As an action, the standard suggests a regular backup. The AI Provider Policy takes up the example and adds a described fallback.
What is behind it?
QMS documentation is memory and evidence at once. It shows that the procedures were applied, and it has to do so for many years. Today it almost always lives in external tools: document management, wiki, ticket system, code repository, platforms for model experiments, compliance software. When one of them fails, more than a working tool is missing. Access to evidence that the provider must present on request is gone.
How documentation becomes unreachable
- Disruption or attack. The service is down for hours, days or weeks.
- End of the product. The vendor shuts down the service, is acquired or goes bankrupt.
- Lost access. A subscription lapses because a payment failed, or the only admin account belongs to someone who has left the company.
- Unusable export. The data can only be exported in a proprietary format, and versions and approval history get lost in the process.
- Untested backup. A backup exists, yet its restore was never practised and fails when it matters.
Why the retention period makes it worse
The AI Act requires key documents to be kept available for ten years (Article 18). EN 18286 requires documented information to be controlled so that it stays available when needed, protected and legible over time, and to remain accessible to notified bodies and authorities (clause 4.5). Over such a period, tools almost certainly change at least once.
Impact
In the short term, processes stall: approvals cannot be granted, corrections cannot be documented, incidents cannot be handled with the necessary records. A request from the authority under Article 21 cannot be answered on time, and an audit by the notified body has to be postponed. In the long term, a permanent loss weighs most. What no longer exists cannot be demonstrated after the fact, and the provider breaches its retention duty even if it always applied the procedures correctly.
Practical examples
Attack on the operator. The example at the start shows the core of the problem: the authority is waiting, and the documents exist and are still out of reach. With weekly exports of released documents to storage of its own, the provider could have answered the request.
Discontinued platform. A provider of an AI system for exam proctoring keeps its test reports on a platform for model experiments. The vendor is acquired and shuts down the service at short notice. The export only covers current runs. The test evidence for the first two versions of the system is lost afterwards.
Locked account. At a small provider of candidate pre-selection software, the subscription for the ticket system runs on the credit card of a former managing director. The payment fails and the account is locked. It holds every release record from the past three years.
Effective measures
The AI Provider Policy from the AI Compliance Starter Kit addresses this risk in these places:
- Section 16 (Risk management): The outage of external software holding QMS documentation is the example given there for a risk of the QMS. It is met with a regular backup and a described fallback and kept in the AI Risk Register under the category “Quality management system”.
- Section 19 (Communication with authorities and operators): The evidence of conformity is compiled per system in a conformity dossier and provided on reasoned request.
- Section 21 (Records and retention): Storage location, access rights and an annual completeness check are defined per system.
- Section 22 (Control of documented information): Earlier versions are kept and stay legible, also outside the document system, for example as PDF. Deletion only happens on an explicit decision after the retention period ends.
- Section 23 (Resources and security of supply): For third-party tools, the source, contractual notice periods and the consequences of an outage or discontinuation are recorded.
These ISO/IEC 42001 controls contribute as well:
- A.4 Resources for AI systems: The documented tooling resources include the software in which design, development and evidence are kept.
- A.10 Third-party and customer relationships: The vendor of a tool is a supplier whose performance and risk are managed.
Link to the AI Act
Article 17(1) of Regulation (EU) 2024/1689 requires systems and procedures for keeping all relevant documentation and information (point (k)) and resource management including measures related to security of supply (point (l)). Article 18 obliges providers to keep, among other things, the technical documentation, the QMS documentation, changes approved by notified bodies and their decisions, and the EU declaration of conformity at the disposal of the competent authorities for ten years. Under Article 21 they provide, upon a reasoned request, all information and documentation that demonstrates conformity. An outage of the QMS software hits all three duties at once. For the reach of the presumption of conformity when applying the standard, see EN 18286: Quality Management System for AI Providers.
Related QMS risks
- QMS-05 Missing information or support from third parties: The vendor of the QMS software is also a third party on whose cooperation the provider depends.
- QMS-01 The QMS misses its intended outcomes: Without access to the documentation, the QMS can no longer demonstrate anything.
- QMS-06 Missing resources or competences: If only one person manages access to the tool, their absence turns into an outage of the documentation.
Sources
- EN 18286:2026, clause 4.5 (documented information) and clause 6.1.1 with its note
- Regulation (EU) 2024/1689 (AI Act), Article 17(1)(k) and (l), Articles 18 and 21
- AI Provider Policy, Sections 16, 19, 21, 22 and 23
- EN 18286: Quality Management System for AI Providers in the Cenedril wiki