In October 2026, a panel client sends a new section of its AI governance questionnaire: Has your organisation assessed whether any AI systems used in service delivery constitute High-Risk AI systems under EU AI Act Annex III?
Two types of organisations receive this question. The first opens its classification register, copies four assessment rationales into the response, and closes section 7(c) in under thirty minutes. The second sends a reply that begins: "We do not believe our AI tools are high-risk."
An organisation that has not performed an Annex III assessment is not in a "Not High-Risk" position. It is in an unassessed position. The EU AI Act does not classify silence. Article 6(2) requires a documented assessment per system. If a system is subsequently found to fall within Annex III, the absence of any prior assessment is an aggravating factor in any regulatory inquiry — not a neutral one.
The peer firm's response — "we don't believe our tools are high-risk" — is not a classification. It is a statement of opinion with no legal basis. The next question from any competent client, regulator, or auditor will be: On what basis?
Annex III classification is the entry gate to the EU AI Act's primary obligation regime. Every downstream obligation — risk management, data governance, technical documentation, human oversight, incident notification, the Fundamental Rights Impact Assessment — flows from a single determination: is this AI system High-Risk under Annex III?
Getting that determination wrong in either direction carries real cost. Wrongly classify a system as High-Risk when it is not, and you impose Arts. 9–15 and Art. 26 obligations on a system that is not subject to them — substantial compliance cost for no legal reason. Wrongly classify a High-Risk system as Not High-Risk, and you leave obligations unmet on a system that is actively regulated — enforcement risk, client trust risk, and the reputational damage of a wrong answer becoming public.
Applying Arts. 9–15 and Art. 26 to systems that are Not High-Risk. Generates disproportionate compliance cost. A common error where Area 8 (administration of justice) is misapplied to solicitor-side legal drafting tools.
Leaving a genuinely High-Risk system unregistered and unmanaged. Triggers enforcement risk at a deadline of 2 December 2027 (standalone AI systems) or 2 August 2028 (embedded in products). The more dangerous error.
The EU AI Act does not require you to discover every AI system you use is High-Risk. It requires you to have looked — properly, system by system — and to be able to show your working. The document you are reading is a structured guide to doing that looking correctly, in a way that produces a defensible, signable record.
Under ITIL 4, the Annex III assessment maps to two practices. As a Risk Management input, it determines whether a specific AI system creates a regulatory risk with a defined compliance deadline. As a Service Configuration Management record, it documents a property of each Configuration Item (AI system): its EU AI Act classification status. The classification register is the CMDB entry for compliance standing, reviewed annually and on material change — exactly the lifecycle cadence ITIL 4 requires.
"In addition to the high-risk AI systems referred to in paragraph 1, AI systems referred to in Annex III shall be considered to be high-risk."
Article 6 contains three operative sub-provisions. For most regulated organisations — law firms, financial institutions, healthcare providers, manufacturers — only one of them is routinely in play.
A High-Risk classification under Art. 6(2) triggers a cascade of obligations. These apply to deployers — organisations using High-Risk AI systems in their operations — as well as to providers. The deployer obligations are the ones most regulated organisations will face.
| Article | Obligation | Deployer Position | Deadline (Standalone AI) |
|---|---|---|---|
| Art. 9 | Risk Management System | Establish and maintain a documented risk management process for the High-Risk system throughout its lifecycle | 2 Dec 2027 |
| Art. 10 | Data and Data Governance | Ensure training, validation, and testing data practices meet quality and governance standards | 2 Dec 2027 |
| Art. 11 | Technical Documentation | Maintain technical documentation describing the system's design, intended purpose, and compliance measures | 2 Dec 2027 |
| Art. 12 | Record-Keeping | Enable automatic logging of events throughout the system's operational lifetime | 2 Dec 2027 |
| Art. 13 | Transparency to Users | Ensure sufficient transparency to enable deployers to interpret the system's output and use it appropriately | 2 Dec 2027 |
| Art. 14 | Human Oversight | Implement human oversight measures enabling effective monitoring, intervention, and override of the system's operations | 2 Dec 2027 |
| Art. 15 | Accuracy, Robustness, Cybersecurity | Achieve appropriate levels of accuracy and resilience, with cybersecurity measures throughout the lifecycle | 2 Dec 2027 |
| Art. 26 | Deployer Obligations | Use the system per provider instructions; assign competent human oversight; monitor operation; report incidents to the provider | 2 Dec 2027 |
| Art. 27 | Fundamental Rights Impact Assessment | Conduct a FRIA before deploying the High-Risk AI system in contexts affecting natural persons | Before deployment |
The 2 December 2027 deadline applies to standalone AI systems. AI systems embedded in regulated products that already require third-party conformity assessment have a later deadline of 2 August 2028. If a system is found to be High-Risk, both the classification date and the applicable deadline shall be recorded in the classification register immediately.
The EU AI Act defines an AI system as a machine-based system designed to operate with varying levels of autonomy that infers, for a given set of objectives, outputs such as predictions, content, recommendations, or decisions. This covers a broad range of tools in routine use across regulated organisations.
Generative AI drafting tools (GPT-4o, Claude, Gemini); productivity copilots (M365 Copilot, Google Workspace AI); matter or case management AI; HR screening or performance tools; client intake automation; document review and classification tools; AI-powered risk or credit scoring.
Rule-based automation (macro scripts, if-then workflows); traditional statistical models used solely for internal analytics without individual-affecting outputs; search engines; spam filters; basic spell-check. The boundary can be unclear — when in doubt, assess and document the scope decision.
Article 6(2) does not provide a de minimis threshold. A general collective statement — "none of our AI tools are high-risk" — is not an Annex III assessment. Each AI system in use by the organisation requires its own documented assessment row. Four systems in use means four assessments to complete and retain. The companion Excel template (REG-AIMS-CLASS-001) provides one row per system and stores the assessment outcome, rationale, and approval.
The classification process is a three-step structured test. The steps are sequential and gated — each step is only reached if the previous step produces a finding that requires further assessment. Each step produces a documented answer. The three answers together constitute the classification record.
Document which areas were screened and why none applies to this system's use case. The assessment closes as Not High-Risk at Step 1. Record the rationale, obtain sign-off, and set the next annual review date. Step 2 is not required.
Record which Annex III area was identified as potentially applicable and why. Document the most plausible area — even if you believe the Step 2 result will be Not High-Risk. Proceed to Step 2 to test the use-case specificity.
The area is identified but the specific use case does not match any listed use case within that area. Document the area screened, the listed use cases tested, and why the organisation's actual use case does not match. Assessment closes as Not High-Risk at Step 2. Obtain sign-off. Step 3 is not required.
The organisation's use case matches a listed Annex III use case. Record the specific listed use case and the factual basis for the match. Proceed to Step 3 to assess whether an Art. 6(3) exclusion is available.
Record the specific exclusion ground and the factual basis. The assessment closes as Not High-Risk under Art. 6(3). Obtain qualified legal review of the exclusion argument before finalising — Art. 6(3) exclusions are strictly interpreted and the factual basis must be robust.
No Art. 6(3) exclusion applies. The system is High-Risk. Open Arts. 9–15 obligation rows immediately. Prioritise Art. 27 FRIA before deployment. The 2 December 2027 (standalone) or 2 August 2028 (embedded) compliance deadline is now active. Obtain qualified legal advice before finalising.
If an AI system is used for multiple distinct purposes within the same organisation, each significant use case shall be assessed separately. A system that is Not High-Risk in one use case may be High-Risk in another. The classification register shall contain one row per system per assessed use case — not one row per system regardless of how it is deployed.
Annex III lists eight areas. For each area, assess whether your AI system's primary use case matches one of the specific listed use cases within that area. A general sector match is not sufficient — the specific use case must correspond. Work through all eight areas for every system, even if you expect the outcome to be Not High-Risk.
What it covers: AI systems used for real-time and post-remote biometric identification of natural persons; biometric categorisation systems inferring or deducing sensitive attributes (race, political opinions, trade union membership, religious beliefs, sexual orientation); and emotion recognition systems.
Classification logic: This area is not triggered by general-purpose AI tools such as drafting assistants, matter management software, email copilots, or document analysis tools. None of these tools perform biometric identification or emotion recognition. Assessment closes at Step 1 — Not High-Risk.
When to look closer: Organisations using AI-powered access control systems, video surveillance with automated identification, or any system that infers personal characteristics from physiological data should give Area 1 close attention. Facial recognition, voice biometric authentication, and gait analysis tools all fall within Area 1 scope.
Area 1 screened. Use case: legal document drafting and research summarisation. The system does not perform biometric identification, categorisation of sensitive attributes, or emotion recognition. Area 1: not applicable. Assessment closes — Not High-Risk at Step 1.
What it covers: AI systems intended to be used as safety components in digital infrastructure, road traffic, water supply, gas, heating, and electricity supply. The key qualifier is "safety component" — the AI must be integral to the safe operation of the infrastructure, not merely used by an organisation that is part of a supply chain.
Classification logic: An organisation using AI to draft documents, manage clients, or analyse data is not managing or operating critical infrastructure. The connection must be direct and operational. Assessment closes at Step 1 for all standard knowledge-economy AI tools — Not High-Risk.
When to look closer: Energy companies, water utilities, and telecoms providers using AI to manage operational safety systems — load balancing, pressure monitoring, network fault detection — should assess Area 2 carefully. The test is whether the AI system's failure or malfunction would create a safety risk to the infrastructure itself.
What it covers: AI systems used to determine access to or assignment within educational institutions; AI used to evaluate the learning outcomes of students and assess their competences; AI used to detect prohibited behaviour in assessment contexts (e.g., exam monitoring).
Classification logic: Regulated businesses are not educational institutions for this purpose. Internal training programmes that use AI to track completion or test knowledge do not constitute educational assessment systems under Area 3 — the use case must be one where the AI system's output directly determines access to or standing within an educational institution. Assessment closes at Step 1 for standard business AI deployments — Not High-Risk.
When to look closer: Universities, colleges, vocational training providers, and online education platforms using AI to assess student performance, allocate places, or monitor examinations shall give Area 3 specific attention. The scope is narrower than "any AI used in education" — it targets AI that determines access and outcomes.
What it covers: AI systems used for recruitment or selection — particularly CV sifting and interview screening; AI used to make decisions affecting working conditions, task allocation, and performance monitoring; AI used to assess, promote, or terminate employment relationships; AI used in access to self-employment or contracting.
Classification logic: Area 4 is the one area where most regulated organisations' internal AI deployments require closer scrutiny. Productivity copilots used for document drafting are not employment management systems. However, the same copilot configured to assist HR in producing performance ratings or disciplinary reports that directly influence employment decisions moves into Area 4 territory.
The test: Does the AI system's output directly inform a decision that affects an individual's employment status, working conditions, or access to work? If yes, the use-case specificity test in Step 2 applies and the use case shall be compared against the listed Area 4 use cases in detail.
Many HR platforms now include AI features as standard. Automated CV filtering, AI-assisted interview scoring, workforce analytics generating individual performance flags, and task allocation algorithms that affect shift patterns or workloads shall all be assessed against Area 4. The assessment shall be documented for each distinct HR AI use case, not for the HR platform as a whole.
Area 4 screened. Use case: draft assistance for legal documents and correspondence in matter delivery. The system does not produce or directly inform employment decisions — its output is legal text reviewed and deployed by solicitors. Area 4: not applicable to this use case. Assessment closes — Not High-Risk at Step 1. Conditional caveat: if the same tool is configured to assist HR in drafting performance assessments or disciplinary records, the Area 4 assessment shall be reopened for that use case.
What it covers: AI systems used to evaluate the creditworthiness of natural persons or to establish their credit scores; AI used to determine access to or the terms of public benefit schemes; AI used to assess risk in health insurance, life insurance, and similar products; AI used in emergency services dispatching.
Classification logic: This area is most commonly triggered for financial services firms, insurers, and public sector bodies. The core question is whether the AI system produces an assessment that directly determines a natural person's access to a financial product, benefit, or essential service — not merely whether the organisation operates in a regulated sector.
The test: Does the AI system generate a creditworthiness score, an insurance risk assessment, or an eligibility recommendation that affects whether a natural person receives a service or benefit? Matter management software that tracks billing and deadlines does not trigger Area 5. An AI tool that generates credit risk flags used in lending decisions does.
Area 5 screened. Use case: matter management, time recording, billing, and file workflow. The system does not produce creditworthiness assessments, determine client eligibility for services, or generate insurance risk ratings. Area 5: not applicable to this use case. Assessment closes — Not High-Risk at Step 1. Conditional caveat: if LEAP AI is configured to make autonomous recommendations on matter acceptance, fee arrangements, or client eligibility that are acted upon without human review, the Area 5 assessment shall be reopened.
What it covers: AI systems used by law enforcement authorities — police, prosecutors, intelligence services — to assess risk posed by individuals, detect emotions in criminal investigations, analyse CCTV or surveillance data, predict criminal activity, detect fraud in criminal contexts, and profile individuals in crime-related investigations.
Classification logic: The key qualifier is "competent law enforcement authorities." A law firm, financial institution, or healthcare provider is not a law enforcement authority. Acting as legal representative for police forces, defendants, or prosecuting authorities does not bring the organisation within Area 6. Use of AI to detect internal fraud or compliance breaches does not trigger Area 6 — that use case falls under the organisation's own risk management, not law enforcement. Assessment closes at Step 1 — Not High-Risk for all standard regulated business AI deployments.
What it covers: AI systems intended to be used by competent public authorities to assess risk, detect forgeries, assist in examination of applications, and make decisions in migration, asylum, and border control contexts. Includes AI used to verify travel documents and assess the reliability of information provided by applicants.
Classification logic: Again, the key qualifier is "competent public authorities." A law firm practising immigration law, or an employer managing visa sponsorships, is not a competent public authority. The organisation's use of AI in preparing immigration applications or advising clients on immigration status does not trigger Area 7 — the provision covers the authority making the decision, not the professional advising the applicant.
If an AI tool is used to assist in the preparation of asylum applications and its output is of a nature that directly influences how a competent authority determines status — for example, automated generation of country condition assessments or risk profiling of individual circumstances — this warrants specific legal analysis before concluding Not High-Risk. Borderline cases shall be recorded as "Requires Legal Advice" in the classification register and referred for external review before finalising the assessment.
What it covers: AI systems intended to be used by a judicial or quasi-judicial body to assist in researching and interpreting facts and the law, and in applying the law to a concrete set of facts. Also covers AI intended to influence electoral and voting processes.
The critical distinction — adjudication vs. advocacy: The phrase "judicial or quasi-judicial body" is the operative qualifier. A law firm, in-house legal team, or legal technology provider is not a judicial or quasi-judicial body. The provision covers AI systems used by courts, tribunals, arbitration panels, and similar bodies in the exercise of their adjudicative function. It does not cover AI used by solicitors, barristers, or legal teams to prepare material for those bodies.
A solicitor using GPT-4o to research case law, draft a skeleton argument, summarise a witness statement, or review contractual clauses is engaged in legal advocacy — preparing material for submission to a judicial body. The court's own use of AI in its processes may fall within Area 8. The solicitor's preparation of material for that court does not.
A firm that classifies its legal drafting AI as High-Risk under Area 8 on the basis that it "relates to" the administration of justice has misread the Annex. The obligations that flow from that misclassification — Arts. 9–15, Art. 26, Art. 27 FRIA — are substantial and disproportionate to a genuinely Not High-Risk tool. The correct classification for solicitor-side legal drafting, research, and document review tools is Not High-Risk, with a documented rationale that explains precisely why Area 8 does not apply — the firm is not a judicial body, and the AI system is not assisting in adjudication.
Area 8 screened. Use case: researching case law, drafting skeleton arguments, summarising evidence, and producing legal correspondence. The system is used by solicitors — not by a judicial or quasi-judicial body — and its outputs are prepared for submission to courts, not used in the exercise of judicial functions. Area 8 listed use case test: the Annex covers AI assisting bodies in applying law to facts in an adjudicative capacity. This use case is legal advocacy, not adjudication. Area 8: not applicable. Assessment closes — Not High-Risk at Step 2.
Under ITIL 4 Configuration Management, the area-by-area screen constitutes the Identification activity for each Configuration Item (AI system): establishing what properties the item has, what its compliance characteristics are, and what its regulatory status is. This screen is not a one-time event — it is a lifecycle record that is updated on material change and reviewed annually. Each area screened, and the outcome of that screen, is a field in the system's configuration record.
"Notwithstanding paragraph 2, an AI system referred to in Annex III shall not be considered high-risk if it does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision-making."
Article 6(3) is only reached if Steps 1 and 2 both produce affirmative results — the system falls within an Annex III area, and the specific use case matches a listed use case within that area. At that point, the exclusion grounds in Art. 6(3) are assessed.
The most common classification error at Step 3 is asserting Art. 6(3) on the basis of general human review of AI outputs. This conflates two distinct legal provisions. Art. 50(4) addresses the transparency labelling obligation for AI-generated content — it is satisfied by disclosing that content is AI-generated and by having editorial procedures. Art. 6(3) addresses classification — it requires that the system's output does not materially influence the outcome of a decision affecting natural persons. A firm that asserts Art. 6(3) because "our solicitors review all AI output" has not engaged with the legal standard. The question is not whether a human sees the output — it is whether the output, in practice, materially determines the decision.
If an Art. 6(3) exclusion is asserted, the following shall be documented in the classification record:
A High-Risk classification is a compliance trigger, not a compliance failure. The Act does not prohibit deploying High-Risk AI systems — it requires that the deployment be managed to defined standards by defined deadlines. The response to a High-Risk finding is structured, not reactive.
Omnibus VII (adopted 29 June 2026) postponed national AI regulatory sandboxes to 2 August 2027. Sandboxes allow supervised testing of AI systems under competent authority oversight — directly relevant where a High-Risk classification produces implementation questions that cannot be resolved internally. The sandbox pathway is a risk-managed route to regulatory clarity before the 2 December 2027 compliance deadline. Note this option in the obligation record for any system where the implementation path is uncertain.
The screening record is the document that converts the three-step assessment into a signed, dated, retainable compliance record. It is the answer to the client questionnaire, the regulatory inquiry, and the internal audit. Complete one record per system per assessment event. The companion REG-AIMS-CLASS-001 Excel template provides the structure — this chapter explains what goes in each field.
REG-AIMS-CLASS-001 is available as a companion Excel download with this guide. It provides the complete five-table structure: firm identification, per-system registry, assessment record, High-Risk obligation tracker, and audit log. The template generates one assessment row per system and includes conditional formatting that flags overdue reviews, missing sign-offs, and High-Risk systems with outstanding obligation rows. The Excel template is the working document — this chapter is the explanatory guide to completing it correctly.
Primary use case: Generative AI drafting and research tool used by solicitors to research case law, draft legal correspondence and skeleton arguments, summarise witness statements, and review contractual clauses. Output is reviewed by qualified solicitors before use.
Step 1: Eight Annex III areas screened. Area 8 (Administration of Justice) identified as the most plausible area on the basis that the tool is used in connection with legal proceedings. Area identified. Proceed to Step 2.
Step 2: Area 8 listed use case tested: AI systems intended to be used by a judicial or quasi-judicial body to assist in researching and interpreting facts and the law and in applying the law to a concrete set of facts. Organisation's use case: AI used by solicitors (a law firm) to prepare material for submission to courts. The firm is not a judicial or quasi-judicial body; it is an advocate before such bodies. The Act covers adjudication, not advocacy. Use case does not match the listed use case. Assessment closes — Not High-Risk at Step 2.
Classification rationale: GPT-4o is used at this firm by solicitors engaged in legal advocacy — preparing documents for submission to courts and tribunals. Annex III Area 8 covers AI systems assisting judicial and quasi-judicial bodies in the exercise of their adjudicative function. This firm is not a judicial body and does not exercise adjudicative functions. The AI system's use is preparatory advocacy, not adjudication. Area 8 does not apply. All other seven Annex III areas screened and found not applicable to this use case. Classification: Not High-Risk.
Conditional caveat: If the system is deployed within a judicial body, quasi-judicial body, or arbitration panel in an adjudicative capacity, this classification shall be reopened.
Approved by: COLP · July 2026 · Next review: July 2027
The classification record you have just completed is not a filing exercise. It is the answer to the question every panel client, regulator, and auditor will ask: how do you know your AI systems are not high-risk? The difference between a firm that has done this work and a firm that has not becomes visible the moment that question arrives.
If your classification exercise identifies one or more High-Risk systems, the Fundamental Rights Impact Assessment is your next obligation. The FRIA Workbook walks you through the statutory process, field by field, with explanatory commentary alongside every section.