What are the AI Act conformity assessment steps for high-risk systems in 2026?

What the EU AI Act Actually Requires for Conformity Assessment

Conformity assessment under the EU AI Act (Regulation 2024/1689) is the formal procedure a provider runs to demonstrate that a high-risk AI system satisfies the obligations in Articles 8 through 17 before placing it on the EU market or putting it into service. As of 14 August 2026, the August 2 2026 enforcement milestone is binding law: the general-purpose AI (GPAI) transparency duties, the prohibited-practice bans, and the governance penalties are now fully applicable, while the high-risk obligations in Annex III are partially delayed under the Omnibus amendment that entered into force earlier in 2026. For any provider whose system is still classified as high-risk, the conformity assessment is not optional paperwork; it is the legal gateway to CE marking and to lawful deployment in all 27 member states.

Also worth reading: How Can Organizations Build Responsible AI Personality Assessment Systems? · How Does AI Profile Validation Actually Work for Psychological Assessment Systems in 2026? · How Do Forensic Risk Assessment Tools Work, and Which Ones Are Most Reliable?

The Act borrows the New Legislative Framework vocabulary from EU product-safety law. A provider must identify the applicable assessment route, compile or generate the technical documentation listed in Annex IV, establish a quality management system, run an internal or external control procedure, register the system in the EU database, affix the CE mark, and sign an EU declaration of conformity. Each step has a defined output that a national market-surveillance authority can request during an inspection.

The Two Assessment Routes and When Each Applies

Article 43 splits high-risk systems into two procedural tracks. The first track is internal control, set out in Annex VI, and applies to most high-risk systems under Annex III and to biometric systems that are not real-time remote biometric identification in publicly accessible spaces. The second track is third-party assessment by a notified body, set out in Annex VII, and is mandatory for biometric identification, critical-infrastructure safety components, AI used as a safety component in medical devices that require third-party conformity assessment under the Medical Device Regulation, and certain AI systems covered by sectoral legislation such as toys, machinery, and in-vitro diagnostics.

The practical difference is who signs the certificate. Under internal control, the provider itself issues the EU declaration of conformity after documenting compliance. Under the Annex VII route, a notified body designated by a member state and notified to the Commission reviews the technical documentation, the quality management system, and the post-market monitoring plan, then issues a certificate of conformity that is valid for a maximum of five years. Providers cannot self-select the lighter route when their product falls into the Annex VII list; doing so is a market-surveillance violation that can trigger fines of up to 15 million euros or 3% of global annual turnover, whichever is higher.

Step-by-Step Procedure for a Provider

The first step is classification. The provider must determine whether the system is prohibited (Title II), high-risk (Annex III or a safety component of a regulated product), limited-risk (transparency obligations only), or minimal-risk. Misclassification is the single most common cause of enforcement action, because every downstream obligation depends on it. Draft Commission guidelines published in 2025 and updated in 2026 clarify that a system remains high-risk if its intended purpose falls within an Annex III use case, even if the provider argues the system is "assistive" or "human-in-the-loop."

The second step is risk management under Article 9. The provider must establish a documented risk-management process that runs across the entire lifecycle, not just at launch. For a continuously learning system, this means defining acceptable performance drift, setting re-training triggers, and documenting how model updates are re-validated before they reach production. The Holland & Knight analysis of the August 2026 deadline notes that U.S. providers in particular have struggled with this step because U.S. product-safety regimes do not require a written, lifecycle-wide risk file.

The third step is data governance under Article 10. Training, validation, and testing datasets must be documented for relevance, representativeness, and statistical bias. For psychological profiling systems, this is the step where the dataset description must show coverage of protected demographic categories and must justify any proxy variables used to infer sensitive attributes. The Commission's draft high-risk guidelines specify that a single paragraph of "we used a balanced dataset" is not sufficient; the technical file must contain dataset cards, bias-metric results, and mitigation measures.

The fourth step is technical documentation under Annex IV. The file must include a general description of the system, its intended purpose, the level of human oversight, the hardware and software environment, the training methodology, the performance metrics, the cybersecurity measures, and a clear statement of which Annex III category applies. For providers using third-party models, the file must also describe how the upstream model is integrated and what additional fine-tuning or guardrails were applied.

The fifth step is record-keeping under Article 12. The system must generate automatic logs that allow traceability of the system's functioning for a period appropriate to the intended purpose. For psychological profiling, this typically means storing input prompts, model version, output, and the human reviewer identifier for at least six months, and longer where sectoral law requires.

The sixth step is transparency and instructions for use under Article 13. Deployers and, where relevant, end users must receive clear instructions describing the system's capabilities, limitations, and foreseeable misuse. The Commission has been explicit that marketing copy does not satisfy this obligation; the instructions must be a separate document delivered with the system.

The seventh step is human oversight under Article 14. The provider must design the system so that natural persons can intervene, override, or shut it down. For high-risk profiling, this usually means a documented escalation path and a UI that surfaces confidence scores and dissenting evidence.

The eighth step is accuracy, robustness, and cybersecurity under Article 15. The provider must declare the accuracy levels in the instructions for use, define resilience measures against adversarial attacks, and document cybersecurity controls. The Commission's GPAI Code of Practice, finalized in 2025, gives a reference set of metrics that providers commonly cite.

The ninth step is the quality management system under Article 17. This is the operational backbone: documented procedures for design and development, data quality, post-market monitoring, incident reporting, and change control. ISO 9001 or ISO 13485 (for medical-device-adjacent systems) are commonly used as the audit framework, but the AI Act does not mandate a specific standard; it mandates that the QMS exists and is effective.

The tenth step is the conformity assessment procedure itself, either Annex VI internal control or Annex VII third-party assessment. After successful completion, the provider draws up the EU declaration of conformity, affixes the CE mark, and registers the system in the EU database before making it available.

Comparison of the Two Routes

The cost bands above are drawn from public fee schedules of major notified bodies and from the Process Excellence Network analysis of the 2026 Omnibus changes; they are indicative, not statutory.

What Changed Under the 2026 Omnibus

The Omnibus amendment, which entered into force in early 2026, narrowed the high-risk perimeter by raising the threshold for what counts as an Annex III system and pushed back the application date for several high-risk obligations. The Tech Times reporting on the August 2 2026 enforcement date confirms that GPAI transparency, prohibited-practice bans, and penalty provisions are now binding, while the high-risk obligations for many Annex III systems are deferred. Providers whose systems were on the borderline of Annex III should re-run the classification test, because a system that was high-risk in 2025 may now be limited-risk, and vice versa for systems newly caught by the revised thresholds.

The Omnibus also introduced a six-day transparency deadline for certain GPAI providers and a December 2026 deadline for nudifier-style applications. These are not conformity-assessment steps, but they sit in the same compliance calendar and are enforced by the same national authorities.

Common Mistakes That Trigger Enforcement

The most frequent failure is treating the technical documentation as a one-time deliverable rather than a living document. The Act requires that the documentation be kept up to date, and a market-surveillance authority can request the current version at any time. The second most frequent failure is inadequate post-market monitoring; Article 72 requires a documented plan, not a generic promise to "watch for issues." The third is treating the EU declaration of conformity as a marketing artifact; it is a legal document that the provider signs under penalty of perjury in the member-state sense, and any material inaccuracy is a separate ground for fines.

A fourth mistake, specific to psychological profiling, is assuming that the system is "assistive" and therefore outside Annex III. The Commission's draft guidelines and the Cambridge Forum on AI: Law and Governance analysis of manipulative AI both treat profiling that materially influences decisions about access to services, employment, or education as high-risk regardless of whether a human makes the final call. A fifth mistake is failing to register in the EU database before deployment; registration is a precondition for lawful placement on the market, and an unregistered system cannot be regularized retroactively.

When to Act and What It Costs

For a provider that has not yet started, the realistic timeline from classification to CE marking is six to nine months under Annex VI and nine to eighteen months under Annex VII, assuming no major redesign is needed. For a provider that already has an ISO 9001 or ISO 13485 QMS, the marginal effort is closer to three to six months because the QMS scaffolding is already in place. Costs vary widely by system complexity, but a useful rule of thumb is that internal-control assessments for a software-only high-risk system fall in the EUR 50,000 to EUR 250,000 range, while notified-body assessments for a biometric or medical-device-adjacent system fall in the EUR 150,000 to EUR 600,000 range, with annual surveillance costs of 15 to 25 percent of the initial assessment fee.

Providers should treat the August 2 2026 deadline as the date by which the GPAI and prohibited-practice obligations must be live, and should plan for high-risk obligations to apply either on the deferred dates set by the Omnibus or immediately for systems that remain clearly within Annex III. Acting now, rather than waiting for the next Commission clarification, is the lower-risk path because the penalty regime is already active and the Italian implementation package shows that national authorities are willing to use it.

Practical Recommendations for a Psychological Profiling Provider

For a provider of an AI psychological profile product, the conformity assessment should start with a written classification memo that maps the intended purpose to the Annex III categories, with particular attention to access to services and employment. The technical file should include a bias audit across protected characteristics, a documented human-oversight workflow, and a transparency notice that meets Article 13. If the system is integrated into a regulated product such as a medical device, the Annex VII route is likely mandatory and the provider should engage a notified body early, because notified-body capacity is constrained and lead times of six to twelve months are common in 2026. Finally, the provider should align its post-market monitoring with the incident-reporting timelines in Article 73, which require notification of serious incidents to the market-surveillance authority within a defined window after becoming aware of them.