What are EU AI Act notified bodies and how do they certify high-risk AI systems?

The Direct Answer: What Notified Bodies Are Under the EU AI Act

Notified bodies under the EU AI Act are independent, third-party conformity assessment organizations that have been formally designated by an EU member state and notified to the European Commission to assess whether high-risk AI systems comply with the regulation's requirements before they can be placed on the EU market. They are not EU institutions, not private consultants, and not government regulators in the enforcement sense. They are accredited assessment organizations — many of them the same entities that already handle CE marking under the Medical Device Regulation (EU) 2017/745 (MDR), the In Vitro Diagnostic Regulation (IVDR), machinery directives, and radio equipment rules. Organizations such as DQS Medizinprodukte GmbH, TÜV Rheinland, TÜV SÜD, BSI Group, DEKRA, and Bureau Veritas are examples of the type of entity that either is or is expected to become notified for AI Act conformity assessment work.

Also worth reading: How Can Responsible AI Deal Networks Transform Regional Health Systems? · How Do Permission-Aware AI Retrieval Systems Protect Private Deal Data in 2026? · Can an AI private deal flow network help founders and operators find better opportunities?

The reason they matter is simple: under Article 43 of the AI Act, most providers of high-risk AI systems cannot self-certify. They must undergo a third-party conformity assessment covering their risk management system, data governance practices, technical documentation, logging, transparency obligations, human oversight measures, accuracy and robustness testing, and cybersecurity controls. Only after a notified body issues a certificate can the provider affix CE marking and legally sell into the EU's roughly 450 million consumer market. For founders and operators building AI products with any European exposure, understanding which notified body to engage — and when — has become as commercially consequential as choosing an investor or a lead engineer.

Why Notified Bodies Exist: The Regulatory Logic Behind Third-Party Assessment

The EU chose a delegated-assessment model rather than building a new pan-European AI regulator from scratch, largely because the infrastructure already existed. The MDR and IVDR regimes run on a network of notified bodies supervised by national competent authorities (in Germany, ZLG; in France, ANSM; in Ireland, HPRA) and coordinated through the Commission's NANDO database, where every notified body's designation scope is publicly listed. The AI Act plugs into this machinery rather than duplicating it. A notified body designation requires proof of independence, impartiality, technical competence, quality management (typically ISO/IEC 17065 accreditation), sufficient staffing of qualified assessors, and financial stability.

This design choice has real consequences, both good and bad. On the positive side, it means companies that already navigate MDR or IVDR conformity assessment face familiar procedures, familiar auditors, and partially overlapping documentation. The intersection of the AI Act with medical device regulation is the single most active area of regulatory commentary right now: legal experts writing in RAPS' Euro Roundup have warned about changes to MDR notified body fee structures, while analyses in Medical Device Network and the Petrie-Flom Center at Harvard have debated whether the overlap between the AI Act and MDR/IVDR produces clarity or chaos for medical AI developers. On the negative side, notified body capacity is finite, appointment backlogs historically stretched to 12–18 months during the MDR transition, and there is no guarantee the same bottleneck will not repeat for AI Act designations through 2026–2027.

Which AI Systems Actually Require a Notified Body

Not every AI product needs one. The AI Act's risk-tiered structure means only Annex III high-risk systems and AI embedded in products covered by harmonized Union legislation (Annex I) require third-party conformity assessment. Annex III covers eight categories: biometric identification and categorization systems; critical infrastructure management; education and vocational training decisions; employment and worker management; access to essential services (credit scoring, insurance pricing, emergency triage); law enforcement; migration, asylum and border control; and administration of justice and democratic processes. Annex I captures AI as a safety component of regulated products — notably medical devices under MDR/IVDR, machinery, lifts, toys, aviation, and vehicles.

There is an important carve-out: if your Annex III system does not process personal data (biometric or otherwise), you may perform the conformity assessment yourself under Article 43(1). In practice almost no commercial AI system qualifies, because training and validation datasets nearly always contain personal data. Separately, general-purpose AI models (GPAI) like large language models do not go through notified bodies at all — they fall under the Commission's GPAI Code of Practice regime, with systemic-risk models above 10^25 FLOPs facing additional evaluation duties enforced directly by the AI Office. So the notified body question applies specifically to deployable high-risk AI systems, not foundation models themselves.

The Conformity Assessment Process, Step by Step

The typical engagement runs through seven stages. First, classification: you document why your system is or is not high-risk, using the Article 6 filters and the Commission's implementing guidance. Second, selection and contracting of a notified body whose NANDO-listed scope actually covers your product category — a mismatch here wastes months. Third, application submission with your full technical documentation file: intended purpose, risk management file per Article 9, data governance documentation per Article 10, model architecture, validation results, post-market monitoring plan per Article 72, and instructions for use. Fourth, assessment, which combines a documentation review with audits of your quality management system (ISO 13485-style for medical AI, ISO 9001-based otherwise) and, increasingly, sampling-based testing of the model itself. Fifth, resolution of nonconformities within deadlines set by the body. Sixth, issuance of an EU technical documentation assessment certificate, typically valid for up to five years with annual surveillance audits. Seventh, CE marking, EU declaration of conformity, and registration in EUDAMED or the equivalent AI database.

Realistic timelines matter enormously for planning. Industry experience from the MDR rollout suggests 6–12 months for a well-prepared applicant and 18 months or more for a first-time applicant with immature documentation. Add the fact that the AI Act's high-risk obligations became binding on 2 August 2026 following the December 2024 delay agreement that pushed the original 2 August 2025 date back twelve months, and the practical message is blunt: companies starting their notified body engagement now are already late relative to their compliance clock.

Comparing Your Certification Pathways

A third pathway deserves mention: for AI that is a component of a medical device, the conformity assessment under MDR may be conducted together with the AI Act assessment by the same notified body, since the AI Act explicitly allows joint assessments where legislation permits. This combined route is what the Petrie-Flom Center debate calls both the greatest efficiency gain and the greatest source of confusion — because MDR notified bodies must extend their designation scope to cover AI-specific requirements, and fee structures are being renegotiated, as the RAPS reporting on notified body fee changes highlights.

Common Mistakes Founders and Operators Make

The first and most expensive mistake is misclassification in either direction. Over-classifying a low-risk tool as high-risk burns six figures and a year on unnecessary certification; under-classifying invites market surveillance action, fines up to €15 million or 3% of global turnover for most high-risk violations (7% / €35 million for prohibited practices), and forced withdrawal from the EU market. The second mistake is treating the notified body relationship as a transactional audit rather than a multi-year partnership — switching bodies mid-lifecycle triggers re-certification. The third is arriving with incomplete Article 10 data governance files; assessors consistently report that data provenance, bias testing, and representativeness documentation are the weakest areas in first submissions. Fourth, companies conflate the notified body with the regulator: the body certifies conformity, but national market surveillance authorities and the Commission's AI Office enforce the regulation, and a certificate does not immunize you from enforcement if your deployed system drifts from its assessed configuration. Fifth, startups ignore the post-market monitoring plan requirement, then fail surveillance audits when they cannot show systematic field-data collection. Finally, some operators assume the August 2026 delay means enforcement is soft; the chatbot-transparency and prohibited-practice provisions were already live earlier, and the omnibus simplification discussions in Brussels have not removed the core high-risk assessment duty.

Costs, Timelines, and When to Act

Budgeting realistically: notified body application and assessment fees for a moderately complex high-risk AI system typically range from €40,000 to €120,000, with medical-device-adjacent AI running higher — often €150,000 to €300,000 all-in when combined MDR/AI Act assessment, QMS audits, and re-testing are included. Annual surveillance adds roughly 20–30% of initial fees. Consultant support for documentation preparation commonly runs €30,000–€100,000. Against this, the penalty asymmetry is stark: a single enforcement action can exceed the entire certification budget by an order of magnitude, and loss of EU market access is usually fatal for a growth-stage company.

On timing, the binding dates are fixed. Prohibited AI practices took effect 2 February 2025; GPAI obligations applied from 2 August 2025; chatbot disclosure and transparency rules are live; and the high-risk regime, delayed to 2 August 2026, is now binding law as of this month. Products placed on the market after that date need certification before sale, and legacy products get a transition window tied to their existing certificates. Given 6–18 month assessment queues, the last responsible moment to open a notified body dialogue was arguably early 2026. If you have not started, begin classification analysis immediately, shortlist two or three bodies with matching NANDO scopes, and request pre-application meetings — most offer them free, and queue positions effectively start at contract signature.

How This Fits Into a Private Deal-Flow Strategy

For founders and operators evaluating AI investments or acquisitions, notified body status is becoming a diligence signal in its own right. A target with an executed notified body contract, a mature technical file, and a realistic certification timeline carries materially lower regulatory risk than one claiming 'AI Act ready' without evidence. Conversely, an announced certification can inflate valuations prematurely — certificates take years, and interim 'assessment in progress' claims should be discounted. Operators building deal-flow networks around European AI assets increasingly screen for three artifacts: the Article 6 classification memo, the notified body engagement letter, and the post-market monitoring plan. Those three documents separate serious EU-market players from aspirational ones, and they are cheaper to verify than any pitch deck claim about model performance.