Erster Auftrag und der Grund für eine GmbH: Mein Start in die Selbstständigkeit | Umair Zaffar
Umair Zaffar November 2, 2026 0 Comments

EU AI Act Article 17: Quality Management System, Implemented in Practice

EU AI Act Article 17 requires providers of high-risk systems to run a quality management system with thirteen mandatory elements, and at SIFAMO we build it not as a binder but as a set of versioned procedures whose compliance the registry, the gateway and the pipeline check by machine at every release.

Article 17 is the article that ties the individual obligations of Section 2 together. Each of them needs a procedure, and Article 17 says those procedures have to form one system that the provider writes down and lives by. It addresses the provider, which includes the insurer whose platform team builds models for the business units. I am an engineer, not a lawyer. What interests me is how the QMS becomes something that is checked at every deployment rather than read once a year. The basics on deadlines and roles are in my EU AI Act guide.

For context: the EU AI Act has been in force since 1 August 2024. The prohibitions and AI literacy obligations have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. The obligations for high-risk systems under Annex III apply, after the postponement by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, and for Annex I products from 2 August 2028. Article 17 belongs to these high-risk obligations. The Omnibus moved deadlines; it removed no obligation.

What Article 17 requires

Article 17 sits in Chapter III, Section 3, among the obligations of providers; Art. 16(c) points to it. The text does not say how good a system has to be. It says how the organisation behind it has to work:

  • Art. 17(1), the core duty: The provider puts in place a quality management system that ensures compliance with the Regulation. It is documented “in a systematic and orderly manner in the form of written policies, procedures and instructions” and covers at least thirteen aspects, points (a) to (m).
  • Points (a) to (d), development: a strategy for regulatory compliance, including conformity assessment and procedures for managing modifications (a); techniques for design, design control and design verification (b) and for development, quality control and quality assurance (c); examination, test and validation procedures before, during and after development, including how often they run (d).
  • Points (e) and (f), specifications and data: the technical specifications and standards applied and, where harmonised standards do not cover everything or are not fully applied, the means by which the requirements are met anyway (e); data management systems from acquisition to retention (f).
  • Points (g) to (i), the three references: the risk management system under Art. 9 (g); post-market monitoring under Art. 72 (h); reporting of serious incidents under Art. 73 (i).
  • Points (j) to (m), organisation: communication with authorities, notified bodies, other operators and customers (j); record-keeping of all relevant documentation (k); resource management including security of supply (l); an accountability framework with the responsibilities of management and staff (m).
  • Art. 17(2): Implementation is proportionate to the size of the organisation; the wording after the Omnibus names SMEs, start-ups and small mid-caps explicitly. The rigour and the level of protection the Regulation requires remain in any case.
  • Art. 17(3): Providers that already run a QMS or an equivalent function under sectoral Union law may integrate the aspects of paragraph 1 into it.
  • Art. 17(4): For financial institutions subject to internal governance rules under Union financial services law, the QMS obligation is deemed fulfilled by complying with those rules. Points (g), (h) and (i) are excluded. Harmonised standards under Art. 40 have to be taken into account.

The technical example

An insurer has an internal AI platform team. It builds and operates models for the business units: a risk assessment and pricing model in health insurance, an assistant for pre-screening job applications in HR, and a model that sorts disability insurance claims by urgency and depth of review. The first two are high-risk under Annex III point 5(c) and point 4(a). The third only sorts, but it shapes the review; we treat it as high-risk, because that line moves quietly in projects.

Legally the insurer is the provider: it develops the systems and puts them into service under its own name. The business units are internal deployers. So the insurer needs a QMS. Not three. Article 17 addresses the organisation, not the individual system. From that follows the most important design decision: the procedures exist once, and the artefacts per model, the risk register under Art. 9, the technical documentation under Art. 11 and the registration record, are instances of those procedures.

The data flow is the same for all three: the business application calls the policy gateway with a use-case ID, never the model. The gateway resolves the registry record, checks the data zone, replaces names, policy numbers and diagnoses with pseudonyms, calls the model and writes the evidence log. Where does Article 17 bite? Everywhere three teams currently run three variants of the same process. The pricing team tests differently from the HR team, changes get approved sometimes by ticket, sometimes by chat, and nobody can say which version of the test procedure applied at the last release. That is exactly what an auditor means when writing “no QMS”.

How we implement it

At SIFAMO we build the QMS as a repository of procedures, not as a manual. Every procedure has an ID, a version, an owner, the point of Art. 17(1) it fulfils, a machine check and a place where the evidence is produced. The governance layer runs the checks. An excerpt, simplified:

{
  "qms_version": "qms-2026-11",
  "owner": "Head of AI Platform; approved by: management board, risk function",
  "scope": ["DE-V-2026-0007 health pricing", "DE-V-2026-0012 applicant pre-screening",
            "DE-V-2026-0019 disability claims prioritisation"],
  "procedures": [
    {"id": "QMS-A-02", "art17": "1a", "name": "Change management",
     "rule": "Every change to model, prompt, index or threshold creates a new registry version; no route without passed gates",
     "check": "registry.version == gateway.deployed_version", "evidence": "registry history"},
    {"id": "QMS-D-01", "art17": "1d", "name": "Quality gates",
     "rule": "Business-owned evaluation set, regression per case, adversarial tests, bias run, rollback rehearsal; before every release and at expiry",
     "check": "pipeline.gates_passed(all)", "evidence": "gate report per release"},
    {"id": "QMS-F-01", "art17": "1f", "name": "Data zones and lineage",
     "rule": "Red never reaches a model, gray only pseudonymised and in scope; provenance per source and per chunk",
     "check": "gateway.zone_check && lineage.complete", "evidence": "lineage record, evidence log"},
    {"id": "QMS-I-01", "art17": "1i", "name": "Incident reporting under Art. 73",
     "rule": "Triage the same working day, draft report from the evidence log, deadline by severity",
     "check": "incidents.open_without_owner == 0", "evidence": "incident file"},
    {"id": "QMS-M-01", "art17": "1m", "name": "Accountability",
     "rule": "Three signatures per release: business owner, risk, platform",
     "check": "registry.signoffs == 3", "evidence": "registry record"}
  ]
}

One system, many models

A model that is not in the manifest’s scope gets no valid registry record and therefore no route. The QMS version travels into every registry record and every entry of the evidence log.

Compliance strategy and change management (point a)

For Annex III systems like those in our example, the conformity assessment is the internal control procedure under Annex VI. Our procedure for it is a checklist the registry record enforces: documentation under Annex IV complete, declaration of conformity under Art. 47 signed, entry in the EU database under Art. 49 present. For change management we closed the gap in the law ourselves: the procedure defines what counts as a substantial modification triggering a new assessment under Art. 43(4). New purpose, new data zone, new model family: substantial. A prompt fix within the tested bounds: not substantial, but still a new version and all gates.

Design, development, testing (points b to d)

Design for us means that, before any model, the business unit answers four questions. Who owns the system, what may it see and do, how do we know afterwards what it did, who switches it off. Development runs in the shared pipeline with code review, versioned prompts and reproducible training. Testing is the quality gates from the manifest above. Point (d) explicitly demands the frequency, so it is in the procedure: before every release, at the expiry date of the registration, and monthly as a regression run against drift.

Specifications and data (points e and f)

Harmonised standards are still missing, so the second half of point (e) applies: we document the internal specifications we apply, for logging, metrics, data zones and oversight tiers, and name the gap honestly. Data management under point (f) is the three zones plus lineage: red is raw personal data and never reaches a model, gray is pseudonymised with access control, green is non-critical. Each source has a lineage record with origin, legal basis, pseudonymisation proof and deletion deadline; for retrieval, access rights apply at query time and provenance is stored per chunk.

The three references and the financial sector (points g to i, paragraph 4)

Points (g) to (i) are exactly the three that paragraph 4 does not waive. Risk management is the versioned risk register per model. Post-market monitoring is a plan under Art. 72 fed from the evidence log: reviewer actions, drift alerts, out-of-scope attempts, held against the risk register monthly, with the kill switch in the hands of the business owner. Incident reporting is a procedure with triage the same working day, a named person who decides whether an event is classified as a serious incident, a draft report generated from the evidence log, and the deadlines of Art. 73.

On paragraph 4: an insurer is subject to the governance requirements of Solvency II, which BaFin spells out in Germany in its VAIT circular; for banks it is MaRisk and BAIT. So the presumption applies, but we do not rely on it. Instead we maintain a mapping table: which existing process, such as change management, information risk management or outsourcing control, covers which point, and where a gap remains. The auditor reads the mapping, not the presumption, and you have to build (g), (h) and (i) regardless.

Authorities, records, resources, accountability (points j to m)

For communication with authorities the procedure names one person with the authority to hand over documentation and logs, as Art. 21 requires. Record-keeping means registry history and evidence log are immutable and retained for ten years under Art. 18. Security of supply means, for us, a second model provider evaluated in advance, a rollback version ready to go, and deputies for key roles. The accountability framework is the three-way sign-off per release: the business owner carries the residual risk and holds the kill switch, the risk function checks as second line, the platform team runs the QMS.

Proportionality (paragraph 2)

A small provider needs the same procedures, not the same volume: five procedures instead of thirty, the registry as a git repository, the evidence as a pipeline report. What must not shrink, in my view, is gates before every release, versioned changes, and an incident procedure with names in it.

What the auditor sees: the versioned QMS manifest, per procedure the history of check results, the mapping to VAIT or MaRisk, per model the instances of the procedures, and in the evidence log the QMS version on every decision. More on this approach under AI governance.

Typical mistakes

  • The QMS is a binder from an ISO 9001 template. Generic, tied to no system, outdated after the first release.
  • One QMS per project. Three teams, three test procedures, three approval paths. Article 17 requires one per provider.
  • Paragraph 4 as a free pass. “We have VAIT” without a mapping, and (g), (h), (i) are missing entirely.
  • Changes without a definition. Nobody has defined what a substantial modification is, so chance decides whether a new conformity assessment is due.

EU AI Act Article 17 checklist

  • There is exactly one QMS for all of the provider’s high-risk systems, with version, owner and approval by management.
  • Every point of Art. 17(1) is mapped to at least one procedure with ID, version, machine check and evidence location.
  • The procedure defines what a substantial modification is and forces a new version with all gates for every change.
  • Test and validation procedures state their frequency: before every release, at expiry and at a fixed rhythm.
  • Risk register, post-market monitoring plan and incident procedure exist independently of paragraph 4.
  • For financial institutions, a mapping of VAIT, MaRisk or BAIT to points (a) to (m) exists, gaps included.

Open questions

  • Which standard? The harmonised standard on quality management for AI has not been published. Until then we document the gap under point (e).
  • What counts as Union law in paragraph 4? Solvency II and the banking directives are Union law; VAIT, BAIT and MaRisk are national supervisory practice. Whether the presumption also covers the national elaboration is undecided.
  • How far does security of supply reach? What a QMS under point (l) has to know and hold in reserve about a hosted foundation model is open.

Frequently asked questions

Does this apply to us as a deployer too?

No, Article 17 addresses the provider; as a deployer under Art. 26 you do not need a QMS of your own. But as soon as you develop a system yourself, put it into service under your own name or substantially modify it, you become the provider under Art. 25, and internal platform teams are the standard case for that.

Is our ISO 9001 certification enough?

As a skeleton yes, as content no. Art. 17(3) allows integration into existing systems. But the thirteen aspects have to be demonstrable within it specifically for the AI systems, above all the risk register, post-market monitoring, incident reporting and the frequency of tests.

Do we need a notified body?

Not for the systems in our example, Annex III points 4 and 5. There the internal control procedure of Annex VI applies: you verify yourself that your QMS complies with Article 17 and you vouch for it with the declaration of conformity.

This post is part of the series on the EU AI Act article by article, all articles at a glance.

Related: EU AI Act · All articles · AI Governance · AI Consultant

Leave Comment