Annex III high-risk AI checklist: 8 areas, 12 obligations, one deadline
Annex III is where most companies discover they are regulated. It is a list of eight areas, and if your system falls into one of them, a fixed set of obligations applies. This is the checklist, in the order an auditor will walk through it.
High-risk classification comes from Article 6. Paragraph 1 covers AI systems that are safety components of products under the Union harmonisation legislation in Annex I — machinery, medical devices, toys, lifts. Paragraph 2 covers the eight areas in Annex III. Most software companies meet Article 6(2), not 6(1).
The eight Annex III areas
| # | Area | Typical systems | Reference |
|---|---|---|---|
| 1 | Biometrics | Remote identification, verification, categorisation of people | Annex III 1(a)–(c) |
| 2 | Critical infrastructure | Safety components in water, gas, electricity, heating, traffic | Annex III 2 |
| 3 | Education and vocational training | Admission, grading, proctoring, learning-path decisions | Annex III 3 |
| 4 | Employment and worker management | Recruitment, screening, ranking, promotion, termination, task allocation | Annex III 4 |
| 5 | Essential services | Credit scoring, insurance risk, eligibility for public benefits, emergency dispatch | Annex III 5 |
| 6 | Law enforcement | Risk assessment of individuals, evidence reliability, profiling, predictive policing | Annex III 6 |
| 7 | Migration and border control | Asylum assessment support, visa risk, document authenticity | Annex III 7 |
| 8 | Administration of justice | Judicial research, legal interpretation support, alternative dispute resolution | Annex III 8 |
Area 4 is the one that catches most SaaS companies. A tool that ranks applicants, scores interviews, drafts rejection reasoning, allocates shifts or recommends who to promote is in scope even when a human signs the final decision — the test is whether the output influences the decision, not whether it is automatic.
The Article 6(3) filter — and why it rarely saves you
Article 6(3) says an Annex III system is not high-risk if it does not pose a significant risk to health, safety or fundamental rights, because it performs a narrow procedural task, improves the result of a previously completed human activity, detects decision patterns without replacing human assessment, or performs a preparatory task.
That is a narrow exception, and the burden of proof is on you: you must document the assessment before placing the system on the market, and register it. Two further limits matter. If the system performs profiling of natural persons, Article 6(3) is unavailable and the system is always high-risk. And a provider who claims the exception still owes the Article 49 registration and information duties.
"It only recommends, a human decides." If the human sees only the recommended shortlist and approves it, the system materially influences the decision. The exception is for systems where the human does the assessment first and the AI only formats or checks it.
The obligation checklist
For a high-risk system the same twelve items appear in every audit. The left column is the legal hook; the right column is what you actually hand over.
| Obligation | Article | What you produce |
|---|---|---|
| Risk management system | Art. 9 | Risk register, owner, evaluation method, residual risk sign-off, review cadence |
| Data and data governance | Art. 10 | Source register, quality checks, representativeness analysis, bias testing, datasheet |
| Technical documentation | Art. 11 + Annex IV | System description, design choices, training data, validation results, limitations |
| Record-keeping | Art. 12 | Automatic event logs with actor, model version, data categories and approvals |
| Transparency and instructions | Art. 13 | Instructions for use: intended purpose, accuracy, known limitations, oversight measures |
| Human oversight | Art. 14 | Named approvers, intervention procedure, stop mechanism, records of decisions |
| Accuracy, robustness, cybersecurity | Art. 15 | Test results, metrics per group, adversarial testing, security controls |
| Quality management system | Art. 17 | Documented processes, responsibilities, change control, incident handling |
| Conformity assessment | Art. 43 | Internal control (Annex VI) or notified body procedure, with the technical file |
| EU declaration and CE marking | Art. 47–48 | Signed declaration of conformity, CE mark on the system or documentation |
| Registration | Art. 49 | Entry in the EU database before placing on the market |
| Post-market monitoring and incidents | Art. 72–73 | Monitoring plan, periodic reports, serious incident reports within the legal deadlines |
Two of these are usually missing in small companies: Article 14 records (approvals exist in email, not in a system) and Article 72 monitoring (nobody owns the system after launch). Both are cheap to fix before an audit and expensive to explain afterwards.
Timeline after the AI Digital Omnibus
| Date | What applies |
|---|---|
| 2 February 2025 | Article 5 prohibited practices; AI literacy duty |
| 2 August 2025 | General-purpose AI obligations; governance and penalties |
| 2 August 2026 | Article 50 transparency duties (chatbots, synthetic content, deepfakes) |
| 2 December 2027 | Chapter III obligations for Annex III high-risk systems (moved by Regulation 2026/1744) |
| 2 August 2028 | High-risk AI embedded in Annex I products |
Regulation (EU) 2026/1744 entered into force on 27 July 2026 and moved the two high-risk deadlines, citing the delay in harmonised standards from CEN and CENELEC. It did not move Article 5, Article 50 or the GPAI rules. For a company selling into enterprise procurement, the practical deadline is earlier than the legal one: security and AI questionnaires already ask for the technical file.
A 30-day plan that produces a file, not a slide
- 1Days 1–3: classify. Write down the Annex III area, the Article 6(3) analysis if you claim the exception, and the role you play (provider, deployer or both).
- 2Days 4–10: inventory the controls. For each of the twelve items, name the artefact and the owner. Anything you cannot point to is open.
- 3Days 11–17: instrument the logs. Turn on event recording with actor, model version and data categories; mask personal data; set the retention rule.
- 4Days 18–24: record oversight. Move approvals out of email into a system that stores approver, timestamp and note, and link each approval to an event.
- 5Days 25–30: seal and review. Generate the first evidence pack for a defined period, verify the hash, and fix the three gaps the pack makes obvious.
Twelve questions map your system to Article 5, Annex III and Article 50, and produce the list of controls that are still missing. No account needed.
Start the free risk checkFour classification mistakes to avoid
- Calling an HR tool "productivity software". If the output affects who gets hired, promoted or dismissed, Annex III 4 applies.
- Assuming a human in the loop removes high-risk status. Article 14 requires oversight for high-risk systems; it does not turn them into limited-risk ones.
- Treating a chatbot as out of scope. It is not high-risk, but Article 50 disclosure has applied since 2 August 2026.
- Relying on Article 6(3) without documentation. The exception has to be assessed and recorded before the system is placed on the market, and it is unavailable for profiling.
What to do if you are only a deployer
Deployers have a shorter list: use the system according to the instructions (Article 26(1)), assign human oversight (Article 26(2)), ensure input data is relevant (Article 26(4)), monitor operation (Article 26(5)), keep logs for at least six months (Article 26(6)), inform workers and affected persons where required (Article 26(7) and (11)), and cooperate with the provider and authorities. You do not produce the technical file, but you do have to be able to show your side of it.
Frequently asked
Is recruitment software always high-risk?
AI systems used for recruitment or selection, especially to place targeted job advertisements, analyse and filter applications, or evaluate candidates, are listed in Annex III point 4 and are high-risk. The Article 6(3) exception exists but is narrow and cannot be used for profiling.
When does the high-risk deadline actually apply?
For Annex III systems, 2 December 2027, after Regulation (EU) 2026/1744 moved the date from 2 August 2026. For AI embedded in Annex I products, 2 August 2028. Prohibitions have applied since 2 February 2025 and Article 50 transparency since 2 August 2026.
Do we need a notified body?
Usually not for Annex III systems: the default is internal control under Annex VI, based on the quality management system and technical documentation. Notified body involvement is required in specific cases, for example where harmonised standards were not applied, or where the system is a biometric identification system in some circumstances.
What is the penalty for getting this wrong?
Article 99 sets tiered fines: up to EUR 35 million or 7% of worldwide turnover for prohibited practices, up to EUR 15 million or 3% for most other obligations including the high-risk requirements, and up to EUR 7.5 million or 1% for supplying incorrect information. For SMEs the lower of the amount and the percentage applies.
Sources
Keep reading
An AI Act evidence pack is the fixed record an operator shows when asked how a high-risk AI system is governed: classification, controls, logs, human oversight and a way to prove nothing was edited.
A practical reading of Article 12 of the EU AI Act: automatic event recording, the minimum fields, retention under Articles 19 and 26(6), and the traps that make logs unusable as evidence.