Skip to main content
Praxikon

Can the list of high-risk AI still change? Articles 7 and 8 of the AI Act on amending Annex III and complying with the requirements

···34 min read

Anyone who wants to know whether an AI system is high-risk under the AI Act usually looks at Annex III first: is the use on the list or not? But that list is not fixed. Article 7 empowers the European Commission to add, modify and remove use cases. And for systems that are on the list, Article 8 is a short provision on what "complying with the requirements" means in practice.

Suppose a Dutch municipality uses a system that sorts applications for a hospitality operating licence. For each application the system gives a score for the applicant's reliability and advises whether a more extensive check is needed; an official decides. The supplier provides the same platform for assessing applications for special assistance benefits. The privacy adviser asks: "The licensing system is not in Annex III now. Can that change, and how quickly? And what exactly will the supplier have to do to comply with the requirements?" The example is fictitious.

In short

The Commission can amend Annex III by delegated act. Adding or modifying is only allowed where the AI system is intended to be used in one of the areas already listed in Annex III, and poses a risk equivalent to or greater than that of the systems already listed (Article 7(1)). In doing so the Commission weighs eleven criteria, such as the intended purpose, the type of data, how dependent the affected persons are and whether an outcome can be reversed (paragraph 2). Removal is possible where a system no longer poses a significant risk and the overall level of protection is not reduced (paragraph 3). A new area (in the Regulation: area heading) cannot be added under Article 7; that requires an amendment of the Regulation, and the Commission evaluates that question by 2 August 2028 and every four years thereafter (Article 112(2)). Every year the Commission assesses whether Annex III needs amending (Article 112(1)). In its first report of 20 May 2026 (COM(2026) 234) it saw no reason to amend yet, but it is monitoring, among others, systems for social welfare benefits and housing, tenant screening, insurance other than life and health insurance, and licensing by public authorities. When we checked on 18 November 2026, EUR-Lex listed no delegated act amending Annex III. Article 8 provides that a high-risk AI system complies with the requirements of Articles 9 to 15, taking into account its intended purpose and the generally acknowledged state of the art, with the risk management system as the basis (paragraph 1). For a product under Section A of Annex I, the provider ensures that the product is fully compliant with all applicable requirements under that harmonisation legislation and may integrate the testing and reporting processes and documentation into what that product legislation already requires (paragraph 2). The Digital Omnibus on AI (Regulation (EU) 2026/1744) did not amend Article 7, Article 8 or Annex III, but moved machinery from Section A to Section B of Annex I. The requirements apply to Annex III systems from 2 December 2027.

This piece is about whether the list can change and about what complying with the requirements involves. How you determine today whether a system falls under Annex III, and when the Article 6(3) exception applies, is covered in the Article 6(3) filter. What the requirements themselves ask is covered in the pieces on risk management under Article 9, technical documentation under Article 11, instructions for use under Article 13 and accuracy and robustness under Article 15.

Where these articles sit

Article 7 is in Section 1 of Chapter III, together with Article 6: the section on classifying AI systems as high-risk. Article 6(2) provides that the systems in Annex III are high-risk; Article 7 governs how that annex can change. Article 8 opens Section 2, with the requirements for high-risk AI systems. The requirements themselves are in Article 9 (risk management) to Article 15 (accuracy, robustness and cybersecurity). Article 8 is the heading above them: it says how those requirements are to be applied.

There are two routes to high-risk. Route 1 runs through Article 6(1) and Annex I: the AI system is a safety component of a product, or is itself a product, covered by EU product legislation, and that product must undergo a third-party conformity assessment. Route 2 runs through Article 6(2) and Annex III: the intended purpose falls under one of the use cases in the eight areas. Article 7 concerns route 2 only. Article 8(1) applies to both routes, except for products under Section B of Annex I (Article 2(2)); Article 8(2) concerns products under Section A of Annex I.

Article 7: how Annex III can change

The instrument. The Commission amends Annex III by a delegated act in accordance with Article 97. That is not an amendment by the European Parliament and the Council, but an act of the Commission that only enters into force if neither the European Parliament nor the Council objects within three months of notification; that period is extended by three months at their initiative (Article 97(6)). Before adopting it, the Commission consults experts designated by each Member State (paragraph 4). The power is conferred for five years from 1 August 2024 and is tacitly extended for periods of identical duration, unless the European Parliament or the Council opposes the extension (paragraph 2). The European Parliament or the Council may also revoke it at any time (paragraph 3).

Adding or modifying: two conditions. The Commission may add use cases to Annex III or modify existing ones where both conditions are fulfilled (Article 7(1)):

  1. the AI systems are intended to be used in any of the areas listed in Annex III (point (a));
  2. the AI systems pose a risk of harm to health and safety, or an adverse impact on fundamental rights, that is equivalent to or greater than the risk posed by the high-risk AI systems already referred to in Annex III (point (b)).

The first condition is a hard limit. The eight areas are: biometrics; critical infrastructure; education and vocational training; employment, workers' management and access to self-employment; access to and enjoyment of essential private services and essential public services and benefits; law enforcement; migration, asylum and border control management; and administration of justice and democratic processes. A use outside those areas cannot be put on the list under Article 7. The second condition is a comparison: not whether a system is new or controversial, but whether the risk is as great as that of what is already listed. The Commission puts it this way in its first report under Article 112: the threshold is not mere technological novelty or public concern, but a demonstrable and comparable risk (COM(2026) 234, point 21).

The eleven criteria. When assessing the second condition, the Commission takes into account (Article 7(2)):

CriterionWhat it is about
(a)The intended purpose of the AI system
(b)The extent to which the system has been used or is likely to be used
(c)The nature and amount of the data, in particular whether special categories of personal data are processed
(d)How autonomously the system acts, and whether a human can override a decision or recommendation that may lead to harm
(e)Whether use has already caused harm or given rise to significant concerns, for example as shown by reports or documented allegations submitted to national competent authorities
(f)The potential extent of the harm: intensity, number of persons affected, disproportionate impact on a particular group
(g)How dependent the affected persons are on the outcome, in particular where they cannot reasonably opt out for practical or legal reasons
(h)An imbalance of power or a vulnerable position in relation to the deployer, due to status, authority, knowledge, economic or social circumstances, or age
(i)Whether the outcome is easily corrigible or reversible; outcomes with an adverse impact on health, safety or fundamental rights are not considered easily corrigible or reversible
(j)The magnitude and likelihood of the benefit of the system for individuals, groups or society
(k)Whether existing Union law already provides effective measures of redress (excluding claims for damages) and effective measures to prevent or substantially minimise the risks

The Commission has a certain margin of discretion in applying these criteria, provided they are duly taken into account (COM(2026) 234, point 20). According to Recital 52, the same methodology and criteria also underlie the current list.

Removal. The Commission may remove a high-risk AI system from Annex III where both conditions are fulfilled: the system no longer poses any significant risk to fundamental rights, health or safety, taking into account the criteria in paragraph 2, and the deletion does not decrease the overall level of protection under Union law (Article 7(3)).

What Article 7 can and cannot do. The first row below falls under Article 7; the other three fall outside that power.

What changesInstrumentProvision
Adding, modifying or removing a use case within the eight areasDelegated act of the CommissionArticle 7(1) and (3), Article 97
Adding, modifying or deleting the conditions of the Article 6(3) filterDelegated act of the CommissionArticle 6(6) to (8), Article 97
Extending an existing area heading or adding a new area headingAmendment of the Regulation by the European Parliament and the Council, on a Commission proposalArticle 112(2)(a) and (10)
Amending the list of prohibited practices in Article 5Amendment of the RegulationArticle 112(1) and (10); COM(2026) 234, point 14

Amendments to the Article 6(3) filter must ensure consistency with the delegated acts adopted under Article 7(1) and may not decrease the overall level of protection (Article 6(8)). A delegated act under Article 7 and one under Article 6 can therefore together determine whether a system is high-risk in practice.

Article 112: the annual review and the first report

From the entry into force of the Regulation until the end of the delegation period in Article 97, the Commission assesses once a year whether the list in Annex III and the list of prohibited practices in Article 5 need amending, and submits its findings to the European Parliament and the Council (Article 112(1)). In addition, by 2 August 2028 and every four years thereafter, it evaluates whether existing area headings in Annex III should be extended or new ones added (paragraph 2(a)). The AI Office develops an objective and participative methodology for this (paragraph 11), and where appropriate the Commission submits proposals to amend the Regulation (paragraph 10).

The first report under paragraph 1 is COM(2026) 234 of 20 May 2026. The Commission consulted the Member States through subgroups of the AI Board (points 24 and 26, with footnote 7), ran a consultation from 6 June to 18 July 2025 that received 547 responses (points 38 and 39), and analysed 3,791 incidents from the OECD AI Incidents and Hazards Monitor (point 50). The outcome for Annex III:

  • No amendment now. Many Member States saw no need to amend Annex III, and seven confirmed this in writing; they considered it premature while the high-risk rules do not yet apply and the classification guidelines are not yet available (points 31 and 32). Proposed use cases that the Commission assessed did not meet the Article 7 threshold or were already covered by the annex (point 45).
  • Flagged for monitoring. Under point 5 of Annex III the report mentions public administration systems for social welfare benefits and social housing, tenant screening and eligibility for private services such as telecoms and utilities, insurance other than life and health insurance (such as motor vehicle, professional liability, homeowners' and disability insurance), and emergency response decisions (point 46). Further: self-help AI chatbots and AI companions, AI in debt collection, and a broader reading of critical infrastructure (points 34, 35 and 46).
  • Licensing by public authorities. One Member State suggested covering AI systems used by public authorities to handle applications for authorisations, such as a driving licence, a firearms and ammunition licence or working with explosives. According to the Commission there is, at this stage, scant evidence that such use cases justify high-risk classification under the Article 7 criteria; it will keep monitoring the risks (point 37).

The report leaves the evaluation of the areas themselves aside, because that is only due by 2 August 2028 (point 5). It announces the next annual assessment for 2026 (point 63). We did not find a second report when we checked on 18 November 2026. On that day the AI Act's document page on EUR-Lex listed one amending act: Regulation (EU) 2026/1744. No delegated act amending Annex III was listed.

What the Digital Omnibus on AI changed

Regulation (EU) 2026/1744, the Digital Omnibus on AI, lists in Article 1 the provisions of the AI Act it amends. Article 7, Article 8 and Annex III are not among them, and the consolidated text of 27 July 2026 shows no amendment markers at these provisions. What did change, and matters here:

  • Article 97. Paragraphs 2, 3 and 6 were replaced (Article 1, point (37)). For Article 7 nothing changes: the power still runs for five years from 1 August 2024, and the objection period is still three months, extendable by three. New are the powers for Article 2(13) and Article 30(2), from 27 July 2026.
  • Article 6 and Article 3, point (14). Three paragraphs were inserted in Article 6 that clarify route 1 (Article 1, point (8)). Paragraphs 1a and 1b concern the safety component: an AI system used solely for non-safety related aspects of user assistance, performance optimisation, service efficiency, automation, convenience or quality control is not a safety component, unless its failure or malfunctioning would endanger health and safety. Paragraph 1c concerns the condition in paragraph 1, point (b): a product that must undergo a third-party conformity assessment solely due to risks other than risks to health and safety, such as radio spectrum or electromagnetic interference, does not fulfil that condition. The definition of safety component in Article 3, point (14), was also amended (Article 1, point (4)). Annex III and the Article 6(3) filter were left as they were.
  • Annex I. Machinery was taken out of Section A: the old Machinery Directive 2006/42/EC was deleted, and the Machinery Regulation (EU) 2023/1230 is now in Section B (Article 1, point (41)). For Section B, only Article 6(1), Article 60a and Articles 102 to 112 of the AI Act apply, plus Articles 57 to 59 in so far as the high-risk requirements have been integrated in that product legislation (Article 2(2), as replaced by Article 1, point (2)). Article 8(2) therefore no longer applies to machinery. Instead, the Commission will add the AI requirements to Annex III of the Machinery Regulation by delegated acts, which are to apply by 2 August 2028 (Article 3, point (1), of Regulation (EU) 2026/1744).
  • Article 2(13). For systems under route 1, the application of requirements or obligations in Articles 9 to 15 and 17 to 25 may be limited, where the product legislation in Section A of Annex I provides an equivalent or higher level of protection and the overall level of protection is not reduced. The Commission adopts delegated acts for this by 2 August 2027 (Article 1, point (3)).
  • Guidelines on integration. The Commission publishes guidelines on the practical implementation of Article 8(2), Article 9(10) and Article 17(3) by 1 August 2027, to avoid duplication (Article 96(1), first subparagraph, point (g), added by Article 1, point (36)).
  • The dates of application in Article 113, third paragraph, point (c) (Article 1, point (40)), see below.

Article 8(1): what complying with the requirements means

The main rule. High-risk AI systems shall comply with the requirements of Section 2, "taking into account their intended purpose as well as the generally acknowledged state of the art on AI and AI-related technologies". The risk management system of Article 9 is taken into account when ensuring compliance (Article 8(1)).

That single sentence contains three things.

  1. All requirements, tailored to the system. The system must comply with Articles 9 to 15: risk management, data and data governance, technical documentation, record-keeping, transparency and instructions for use, human oversight, and accuracy, robustness and cybersecurity. The intended purpose determines what that asks in concrete terms. In our reading this does not allow a provider to skip requirements; it means the provider decides for each requirement how it is met for this purpose: what level of accuracy is needed, which groups must be represented in the data, what oversight fits.
  2. The state of the art, not perfection. The measures should take into account the generally acknowledged state of the art on AI, and be proportionate and effective (Recital 64). Harmonised standards are the usual way to capture that state of the art: a system that complies with a harmonised standard whose reference has been published in the Official Journal is presumed to comply with the requirements of Section 2, to the extent the standard covers them (Article 40(1)). Implementing Decision (EU) 2026/2274 of 12 October 2026 published the reference of EN 18286:2026; the decision appeared in the Official Journal on 14 October 2026 and entered into force that day, and the presumption of conformity applies from that publication (Article 2 and Recital 6 of the decision). The decision refers to Article 40 in Recital 1, but does not say which requirements of the AI Act the standard covers. The standard concerns quality management systems for providers, including processes for post-market monitoring (Recital 3 of the decision), the subject of Article 17. Do not assume that it covers the requirements of Articles 9 to 15. In our reading the state of the art is a moving target: what is acknowledged today may be outdated in two years. In our reading, continuous risk management (Article 9(2)) and post-market monitoring (Article 72) therefore require the provider to recalibrate its measures periodically.
  3. Risk management as the backbone. The risk management system is a continuous iterative process planned and run throughout the entire lifecycle of the system, requiring regular systematic review and updating (Article 9(2)). Article 8 makes it the starting point: the choices under the other requirements follow from the risks the risk management identifies.

Article 8(2): interplay with product legislation

Full compliance with everything. Where a product contains an AI system to which both the AI Act and the product legislation in Section A of Annex I apply, the provider is responsible for ensuring that the product is fully compliant with all applicable requirements under that product legislation (Article 8(2), first sentence). After the Omnibus, Section A covers, among others, toys, lifts, radio equipment, pressure equipment, personal protective equipment, medical devices and in vitro diagnostic medical devices.

The choice to integrate. To ensure consistency, avoid duplication and minimise additional burdens, the provider may integrate the testing and reporting processes, information and documentation required by the AI Act into the documentation and procedures that the product legislation already requires (paragraph 2, second sentence). It is a choice, not an obligation. The AI Act adds requirements for risks the product legislation does not cover; according to Recital 64, machinery or medical devices incorporating an AI system might present risks not addressed by the essential health and safety requirements of the product legislation. Integrating therefore means combining, not replacing: the AI requirements must be traceable in the integrated process.

Article 8(2) does not stand alone. Three related provisions work out the same idea:

ProvisionWhat is allowedFor whom
Article 9(10)Making the AI Act risk management part of, or combining it with, risk management procedures under other Union lawProviders subject to requirements on internal risk management processes under other Union law
Article 11(2)A single set of technical documentation with the information required by the AI Act and by the product legislationProviders of systems related to a product under Section A of Annex I
Article 17(3)Including the quality management aspects in the quality management system under sectoral Union lawProviders with quality management obligations or an equivalent function under sectoral Union law
Article 17(4)The quality management system is deemed in place by complying with the governance rules of financial services law, except Article 17(1)(g), (h) and (i) (risk management, post-market monitoring, serious incident reporting)Providers that are financial institutions subject to internal governance requirements under Union financial services law

The Digital Omnibus on AI expressly names Article 8(2), Article 9(10) and Article 17(3) in Recital 37 as the mechanisms that allow economic operators to integrate an assessment of AI-specific risks into existing risk and quality management systems. How this works for medical devices is covered in medical AI under the AI Act and the MDR; how the conformity assessment works in conformity assessment: internal control or notified body.

Main example: the municipality, the licensing system and the supplier

Is the licensing system in Annex III now? In our reading, no. The system does not assess eligibility for essential public assistance benefits and services (Annex III, point 5(a)), and the municipality is not acting here as a law enforcement authority in relation to criminal offences (point 6). That a Member State suggested adding such systems to Annex III also indicates that they are not covered now; the Commission does not say so in so many words (COM(2026) 234, point 37).

Can that change? Yes, but the route depends on a question the Regulation does not answer: does licensing by a municipality fall within one of the eight areas, for example point 5, "access to and enjoyment of essential private services and essential public services and benefits"?

  • If so, the Commission can add the use case by delegated act, where the risk is equivalent to or greater than that of the current use cases. Criteria that in our reading count for this system: the applicant cannot avoid the outcome if they want to run a business (Article 7(2)(g)), the municipality has authority over the applicant (point (h)), the type of data processed, such as data on reliability and behaviour (point (c)), and the number of applicants affected (point (f)). On the other side are the human decision (point (d)) and the existing objection and appeal procedures (point (k)). For comparable licences, such as a driving licence or a firearms licence, the Commission currently sees scant evidence (point 37).
  • If not, a new area heading is needed, and that requires an amendment of the Regulation. The first evaluation of the area headings is due by 2 August 2028 (Article 112(2)(a)); any proposal then follows the ordinary legislative procedure.

How quickly? A delegated act only enters into force once the three-month objection period (extendable by three) has expired without objection, or if the European Parliament and the Council say earlier that they will not object (Article 97(6)). Before that comes the preparatory work and the consultation of Member State experts (paragraph 4). The AI Act contains no separate transitional rule for a use case added to Annex III later. Article 111(2) concerns systems placed on the market or put into service before the date of application of Chapter III; whether and how that rule applies to a system that only becomes high-risk through a later amendment is an open question. In our reading a municipality does well to set the system up now so that complying later does not become impossible.

A risk of its own: changing the purpose. There is a faster way to high-risk, and it lies with the municipality itself. The supplier already provides the platform for assessing applications for special assistance benefits; that use falls under Annex III, point 5(a), and for that purpose the supplier is the provider. Article 25(1)(c) only comes into play if the municipality uses the licensing system outside the purpose determined by the supplier. If it modifies the intended purpose of a system not classified as high-risk in such a way that it becomes high-risk, for example by using the licensing module itself for benefit applications, it becomes a provider itself (Article 25(1)(c)). In our reading the municipality should therefore record the purpose for which it uses each part, and agree with the supplier that the benefits module is supplied as a separate high-risk system.

What already applies: the GDPR. Whether the system is high-risk under the AI Act is separate from whether a DPIA is needed. In our reading a score for the reliability of applicants will readily fall under item 15 (profiling) of the Dutch Data Protection Authority's list of processing operations for which a DPIA is mandatory, in so far as natural persons are scored; the list gives reliability and behaviour as examples (Stcrt. 2019, 64418, item 15, in Dutch; Article 35(4) GDPR). Because the licensing decision produces legal effects, Article 35(3)(a) GDPR is also relevant: a systematic and extensive evaluation of personal aspects based on automated processing, on which decisions producing legal effects are based. Whether the evaluation is that extensive depends on how the system is set up. If Annex III is amended later, that DPIA already exists. For a high-risk system, such as the benefits module, the municipality, as a body governed by public law, must also carry out a fundamental rights impact assessment; where the DPIA already meets one of its obligations, it may cross-reference those sections or include them (Article 27(1) and (4)). If the benefits module is already in use before 2 December 2027, the high-risk obligations, including the fundamental rights impact assessment, apply by 2 August 2030 at the latest, or earlier if the system is subject to significant changes in its design before then (Article 111(2)). See DPIA and FRIA side by side.

The supplier: building the requirements into existing processes. For the benefits module the supplier is the provider of a high-risk system. It already works with a quality system and an information security management system under international standards, and wants to build the AI Act requirements into them rather than set up a second system alongside. What is allowed?

  1. Article 8(2) does not help. Its software is not a product under Section A of Annex I. The integration option in paragraph 2 is not available to it.
  2. In our reading, neither do Article 9(10) and Article 17(3), as long as it is only voluntarily certified. Those provisions concern requirements under Union law, and a voluntary certification is not Union law. If Union law does impose risk management requirements on it, Article 9(10) can help for that part.
  3. But it may reuse in practice. Nothing in the AI Act prevents it from placing risk management, quality management and documentation within the processes it already has. In our reading the test is then not where the processes sit, but whether each requirement of Articles 9 to 15 and of Article 17 can demonstrably be found in them. For quality management it can use EN 18286:2026, the standard whose reference was published by Implementing Decision (EU) 2026/2274. The decision refers to Article 40 in Recital 1, but does not say which requirements of the AI Act the standard covers; the standard concerns quality management systems, including post-market monitoring. It should not assume that the standard also covers the requirements of Articles 9 to 15.
  4. Tailored to the purpose. The intended purpose, assessing eligibility for special assistance benefits, determines what Article 8(1) requires. In our reading that includes testing on groups that more often have fluctuating incomes, instructions for use that show the caseworker where the system is less reliable, and human oversight that can genuinely override an outcome (Article 14).

What the municipality agrees in the contract with the supplier: the intended purpose of each part of the platform; that the supplier notifies it if a use case becomes high-risk through an amendment of Annex III and then cooperates on compliance; and what information the municipality receives for its DPIA and, for the benefits module, its fundamental rights impact assessment. See also AI procurement contracts and the FRIA for municipalities.

Example 2: an insurer with its own pricing model

An insurer develops its own model for risk assessment and pricing of motor insurance for consumers and uses it in its own underwriting process. This example is also fictitious.

Not Annex III now. Annex III, point 5(c), only covers risk assessment and pricing in the case of life and health insurance. Motor insurance is not covered. Stakeholders asked the Commission to include other essential insurance, with motor vehicle insurance as an example; the Commission is monitoring this (COM(2026) 234, point 46).

But within an area. Unlike the municipality's case, no new area is needed here: insurance already falls under point 5. The Commission can therefore modify point 5(c) or add a use case by delegated act, where the risk is equivalent to or greater than in life and health insurance. Criteria that in our reading weigh in: how many people are affected ((b) and (f)), whether customers can avoid the outcome ((g)), and what protection insurance law already offers ((k)).

If it changes. The insurer is the provider, because it develops the model itself and puts it into service under its own name, and is also the deployer. As a financial institution it has integration routes that the municipality's supplier does not. Recital 158 lists Directive 2009/138/EC (Solvency II) among Union financial services law; in our reading an insurer therefore falls under the rules for financial institutions. Its quality management system is then deemed in place if it complies with the governance rules of financial services law, except for Article 17(1)(g), (h) and (i): the risk management system, post-market monitoring and the procedures for reporting serious incidents (Article 17(4)). It must set up those three elements itself. It may make the risk management part of, or combine it with, its existing risk management procedures under financial services law (Article 9(10)). It keeps the technical documentation and the logs as part of its documentation under financial services law (Article 18(3) and Article 19(2)). For the fundamental rights impact assessment it depends on the form of the amendment. Article 27(1) refers to Annex III, points 5(b) and (c). If the Commission includes motor insurance in point 5(c) itself, Article 27(1) also applies to the insurer as deployer. If it adds a new point, the obligation does not follow from Article 27(1) as currently worded.

Compare the bank with its own credit model: there the question is not whether the list changes, because creditworthiness is already in point 5(b). See AI in credit decisions and, for premiums, fair dynamic insurance premiums.

When does it apply?

Sections 1, 2 and 3 of Chapter III apply from 2 December 2027 for AI systems classified as high-risk under Article 6(2) and Annex III, and from 2 August 2028 for AI systems under Article 6(1) and Annex I. Article 6(5), on the guidelines, is excepted (Article 113, third paragraph, point (c), as replaced by Article 1, point (40), of Regulation (EU) 2026/1744). For Article 8 that means: from 2 December 2027 for Annex III systems and from 2 August 2028 for systems related to products under Section A of Annex I. Article 8 does not apply to systems related to products under Section B of Annex I (Article 2(2)). Article 7 concerns Annex III only; for that provision only the date of 2 December 2027 is relevant.

The power to adopt delegated acts is conferred on the Commission for five years from 1 August 2024 (Article 97(2)), and the annual assessment under Article 112(1) runs from the entry into force of the Regulation. Whether the Commission could already amend Annex III by delegated act before 2 December 2027 is not expressly stated in the Regulation. In practice the question does not arise now: in May 2026 the Commission saw no reason for an amendment.

Where a high-risk system was placed on the market or put into service before the date of application of Chapter III referred to in Article 113, the high-risk obligations apply to its operators only if it is subsequently subject to significant changes in its design; providers and deployers of systems intended to be used by public authorities must comply by 2 August 2030 (Article 111(2), as replaced by Article 1, point (39)). That date of application is 2 December 2027 for Annex III systems and 2 August 2028 for systems under Annex I. Under the same provision, Article 5 continues to apply in full.

What to do now

As a deployer, such as a municipality:

  1. Record the intended purpose of each system and check it against the use cases in Annex III, not just the areas.
  2. Keep a short list of systems that currently fall just outside Annex III but come close, such as the use cases the Commission monitors in COM(2026) 234.
  3. Read the annual report under Article 112(1) and the evaluation of the areas due by 2 August 2028, and then check whether your list changes.
  4. Carry out the DPIA now, if the processing is on the Dutch DPA's list or is otherwise likely to result in a high risk (Article 35(1) GDPR), regardless of Annex III.
  5. Do not change a system's purpose unnoticed. A new purpose can make you the provider (Article 25(1)(c)).
  6. Agree in the contract that the supplier informs you and cooperates if a use case becomes high-risk through an amendment of Annex III.

As a provider:

  1. Explain for each requirement of Articles 9 to 15 how the intended purpose shapes it, and record this in the technical documentation.
  2. Follow the state of the art and the publication of harmonised standards, and schedule a periodic recalibration of your measures within the risk management.
  3. Integrate where the law allows it (Articles 8(2), 9(10), 11(2), 17(3) and (4)), and elsewhere reuse existing processes without dropping requirements.
  4. Follow the guidelines on integration once the Commission publishes them, by 1 August 2027 (Article 96(1)(g)).
  5. If you have a product with AI under Annex I, check whether it is still in Section A; machinery has been in Section B since 27 July 2026.

Not sure whether your system currently falls under Annex III and which role you have? With the case check you describe the use in a few questions and see which AI Act and GDPR obligations come into view.

Frequently asked questions about amending Annex III and complying with the requirements

Newsletter

Every Tuesday, the AI Act week ahead in 5 minutes

A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.

Practical and short · No spam · One-click unsubscribe

Written by

Zahed Ashkara, freelance AI and privacy consultant and legal adviser

Zahed Ashkara

Freelance AI and privacy consultant and legal adviser · AI governance & safety

Zahed Ashkara is a freelance AI and privacy consultant and legal adviser for the GDPR and the EU AI Act. Organisations hire him as AI governance consultant, privacy counsel, external DPO or AI governance lead, have him carry out a DPIA, FRIA or AI Act classification, or book an in-company training. He focuses on AI governance and safety, including AI agents. He works exclusively as a freelancer, through Embed AI (trade name Praxikon). He is a Certified AI Compliance Officer (CAICO) and has been a member of the NEN standards committee Artificial Intelligence & Big Data since July 2026.

Follow on LinkedInView expert profile

Use these sources in your AI assistant too: Praxikon in Claude and other assistants

Real-World Stories

Discover how organizations apply this in practice

Related articles

When a high-risk AI system does not comply: corrective actions, documentation keeping and cooperation under Articles 18, 20 and 21 of the AI Act

What must a provider do when its high-risk AI system does not (or no longer) comply with the AI Act, and how long does it keep the documentation? Article 18 (ten years of documentation keeping), Article 20 (corrective actions and the duty to inform the supply chain and the supervisor) and Article 21 (cooperation with the authorities, access to logs, confidentiality). With what the Digital Omnibus did and did not change, when it applies and which fines apply, and with a municipality that receives a notice from its supplier and a bank with its own credit model as examples.

EU AI ActArticle 18Article 20

EU declaration of conformity and CE marking for high-risk AI: what Articles 47 and 48 of the AI Act require

After the conformity assessment, the provider of a high-risk AI system draws up an EU declaration of conformity for each system and affixes the CE marking (Articles 47 and 48). What the declaration must contain under Annex V, why the signature means taking responsibility, how a digital CE marking works, what authorised representatives, importers and distributors check, what a municipality verifies when it buys, when it applies and what happens when it is missing, with a vendor of a welfare model and a bank as examples.

EU AI ActArticle 47Article 48

Importers and distributors of high-risk AI: what Articles 23 and 24 of the AI Act require

Between the provider and the deployer there is often an importer and a distributor. Articles 23 and 24 of the AI Act give them their own obligations for a high-risk AI system: check before supply, do not supply when in doubt, add your own name, mind storage and transport, keep records and cooperate with the authorities. When a reseller or integrator becomes a provider itself (Article 25(1)), what the Digital Omnibus changed, when it applies and which fines apply, with a municipality buying through a Dutch reseller and a bank working with an integrator as examples.

EU AI ActArticle 23Article 24