Annex VI Internal Control Conformity Assessment: An Implementation Guide

What the Annex VI route covers

Under the EU AI Act, conformity assessment is the process of demonstrating whether the requirements in Chapter III, Section 2 have been fulfilled for a high-risk AI system. Annex VI is the procedure for doing that through internal control. It is built on points 2, 3 and 4 of the Annex and does not provide for the involvement of a notified body.

Eligibility depends on the Annex III category. For high-risk systems referred to in points 2 to 8 of Annex III, the provider follows the Annex VI procedure. For systems listed in point 1 of Annex III, a provider that has applied harmonised standards (Article 40) or, where applicable, common specifications (Article 41) can opt for either Annex VI or Annex VII. Annex VII consists of an assessment of the quality management system and technical documentation with a notified body. Article 43 also sets conditions under which the Annex VII procedure must be followed.

Because no external body reviews the work, the provider's own procedure is the control. An auditor or authority will judge the quality of that procedure and the evidence it left behind, so it needs a named owner, a fixed sequence and defined artifacts.

Preconditions to settle before step one

Most inconsistency in internal assessments starts with unrecorded scoping decisions. Settle these before any verification work begins, and store each as a dated record.

The route decision matters most. Map the system to its Annex III point, record whether Annex VI is mandatory or optional, and for point 1 systems record which harmonised standards or common specifications were applied. Read the Article 43 conditions that require the Annex VII route and document why none applies. Also record which version of the regulation text the assessment used, since the rules are being amended.

• Annex III point and the resulting route (mandatory or optional Annex VI)
• Harmonised standards or common specifications applied, for point 1 systems
• Article 43 check showing the Annex VII route is not required
• Whether the Cyber Resilience Act also applies, which adds a separate cybersecurity conformity procedure
• Regulation text version used for the assessment
• Named assessment owner and independent reviewer

Showing that runtime behavior matches documented controls

Steps 3 and 4 share a weakness. Documentation describes what a system is supposed to do, and a reviewer can read it without learning what the system actually does. A control described in the technical documentation, such as a restriction on what data may leave the system or a requirement that certain actions get human approval, is a claim until runtime behavior shows it holds.

This is where enforcement in the execution path helps. Trussed enforces policy at runtime, not reviewed after the fact: for individual interactions and tool calls it can allow, redact, block or escalate for human approval. Each enforced decision produces audit-ready evidence mapped to the EU AI Act. That gives the assessor a record of what the system was permitted to do and what it did, which can be set beside the documented controls in step 4.

The scope is specific. Enforcement evidence supports the assessment record for interactions and tool calls. It does not conclude conformity. The provider's own examination, findings and sign-off do that.

Keeping the evidence current after the declaration

An assessment file starts aging on the day it is signed. Three practices slow that down.

First, triage every change to the system against the documented assessment. Any change that could affect a Chapter III, Section 2 requirement, including changes to the intended purpose, the design, the controls or the monitoring plan, should trigger a review of the affected steps and an update to the requirement-to-evidence matrix. Record the decision either way, including when no reassessment was judged necessary.

Second, feed post-market monitoring back into the file. Monitoring output is evidence for step 4 in future cycles, and continuous runtime evidence is easier to keep current than evidence assembled by hand before a review.

Third, run a regulatory-change watch. As of November 2026, the Digital Omnibus on AI, proposed by the Commission on 19 November 2025, has reached a provisional agreement between the Council and Parliament, announced on 7 May 2026. It is reported to set 2 December 2027 for stand-alone high-risk systems and 2 August 2028 for high-risk systems embedded in products. Schedule assessment milestones against the current text, and record which version each assessment used. Where the Cyber Resilience Act also applies, assign an owner for that procedure and define how its outcome is reconciled with the Annex VI file.