EU AI Act · Legal logic
Updated 6 October 2026 · Digital Omnibus included
AI Act obligations by role: the legal logic
Which obligations of the AI Act apply to an AI system does not follow from a single classification but from a fixed order of questions: AI system, scope, role, prohibitions, high risk, transparency and general-purpose AI models. This page sets out that logic, the role matrix and the application dates.
Assessment order in eight steps · Role matrix from provider to distributor · Role switch under Article 25 · Obligations finder with start dates
Risk Classification
AI-based Application Screening
Section A: High-risk triggers
18/21 answeredDoes the system make or influence decisions about natural persons?
Does the outcome have legal or similarly significant effect?
Are special categories of personal data processed?
Risk level
High
Annex III · Employment & workforce management
Requirements
- Risk management systemArt. 9 EU AI ActFulfilled
- Data governanceArt. 10 EU AI ActPartial
- Human oversightArt. 14 EU AI ActMissing
Short answer
Your obligations under the AI Act depend on two things: your role (provider, deployer, importer or distributor) and how the system is classified (prohibited, high-risk, transparency case or none of these). Check in this order: AI system under Art. 3(1), scope under Art. 2, role, Art. 5, Art. 6, Art. 50 and Art. 51 onwards. Art. 4 always applies to providers and deployers.
Why order matters
An obligation never applies to “the system” – it applies to a role in relation to a system
The AI Act (Regulation (EU) 2024/1689) is built like product safety law: it allocates obligations to operators along the supply chain. The same recruitment software triggers a quality management system, a conformity assessment and registration for the company that makes it, while the HR team that uses it has to organise human oversight, keep logs and inform staff. Asking only “Is our system high-risk?” therefore gives you half an answer.
The steps also build on each other. If a program is not an AI system within the meaning of Art. 3(1), the role question never arises. If a practice falls under Art. 5, the high-risk assessment is pointless because the use is not allowed at all. And the exception in Art. 6(3) can take an Annex III system out of the high-risk regime, but not out of Art. 50 or Art. 4. Keeping to the order avoids duplicate work and, above all, false all-clears.
This page explains the logic step by step and turns it into a role matrix and an obligations finder. Individual topics have their own pages: risk classes under AI risk classification, the Annex III categories, the high-risk checklist and the Article 50 labelling rules. Here the focus is on the framework that connects them.
Legal basis
The provisions that make up the list of obligations
Consolidated version of the AI Act as amended by Regulation (EU) 2026/1744 (the “Digital Omnibus”). The table follows the assessment order.
Scroll the table sideways →
| Provision | Content | Addressee | What it means for the assessment |
|---|---|---|---|
| Art. 3(1) | AI system: a machine-based system designed to operate with varying levels of autonomy, which may adapt after deployment and infers from its input how to generate outputs such as predictions, content, recommendations or decisions. | all | The entry gate. Conventional rule-based software that does not infer is outside; record your reasoning. |
| Art. 2 | Scope: providers, deployers, importers, distributors, product manufacturers, authorised representatives; exclusions include military and national security (para. 3), research (paras. 6, 8), purely personal use (para. 10) and open-source systems (para. 12). | all | Providers in third countries are covered where the output is used in the Union (para. 1(c)). The open-source exclusion does not cover high-risk, Article 5 and Article 50 cases. |
| Art. 3(3)–(8) | Roles: provider (3), deployer (4), authorised representative (5), importer (6), distributor (7), operator as the umbrella term (8). | all | Roles are determined per system and per activity. One company can be provider and deployer of the same system at once. |
| Art. 4 | AI literacy: take measures to support the development of AI literacy of staff and others acting on your behalf; no particular level has to be guaranteed. | providers, deployers | Applies since 2 February 2025 regardless of risk class, including simple assistant tools. |
| Art. 5 | Prohibited practices, e.g. manipulative techniques, social scoring, emotion recognition in the workplace and in education, untargeted scraping of facial images. | providers, deployers | Since 2 February 2025. The prohibitions added by the Omnibus (para. 1(ba) and (bb)) apply from 2 December 2026. |
| Art. 6 | High risk: para. 1 safety component or product under Annex I requiring third-party conformity assessment; para. 2 use cases in Annex III; para. 3 exception where there is no significant risk; para. 4 documentation and registration. | providers (classification) | Profiling of natural persons rules out the para. 3 exception. The provider’s classification shapes the obligations of every other role. |
| Arts. 16–21 | Provider obligations for high-risk AI: requirements of Arts. 8–15, quality management system, keeping documentation and logs, conformity assessment, declaration of conformity, CE marking, registration, corrective action, cooperation with authorities. | providers | The longest list. Applies from 2 December 2027 (Annex III) or 2 August 2028 (Annex I). |
| Arts. 22–24 | Authorised representative for providers established in third countries (Art. 22); checks by importers (Art. 23) and distributors (Art. 24) before placing on the market or making available. | authorised representatives, importers, distributors | Verification duties: CE marking, declaration of conformity, instructions for use, technical documentation; where in doubt, do not place on the market. |
| Art. 25 | Role switch: distributors, importers, deployers or other third parties are considered providers if they put their name on a high-risk system, substantially modify it, or change the intended purpose so that it becomes high-risk. | deployers, importers, distributors, third parties | In these cases the original provider must in principle supply documentation, information and technical access (para. 2). |
| Arts. 26, 27 | Deployer obligations for high-risk AI (use in line with the instructions, human oversight, input data, monitoring, logs, information) and the fundamental rights impact assessment for certain deployers. | deployers | The fundamental rights impact assessment applies to bodies governed by public law, private entities providing public services and deployers under Annex III point 5(b) and (c). |
| Art. 50 | Transparency: disclosure of AI interaction (para. 1) and machine-readable marking (para. 2) as provider duties; information on emotion recognition and biometric categorisation (para. 3) and disclosure of deepfakes and certain texts (para. 4) as deployer duties. | providers, deployers | Applies since 2 August 2026 alongside all other obligations. For generative systems placed on the market before 2 August 2026, para. 2 applies from 2 December 2026 (Art. 111(4)). |
| Arts. 51–55 | General-purpose AI models: documentation, information for downstream providers, copyright policy, summary of training content (Art. 53); additional obligations for systemic risk (Art. 55). | providers of AI models | Since 2 August 2025. A separate track next to the system assessment; models placed on the market before 2 August 2025 have until 2 August 2027 (Art. 111(3)). |
| Arts. 111, 113 | Entry into application and transitional rules, as amended by Regulation (EU) 2026/1744. | all | Decide from when each obligation on the list actually applies. Legacy high-risk systems only fall under Chapter III after a significant change in design (Art. 111(2)). |
Wording checked on 6 October 2026 on ai-act-law.eu (consolidated version). The version published in the Official Journal on EUR-Lex is authoritative.
Assessment order
How to work out the obligations in eight steps
Each step answers one question whose result prepares the next. Record the answer and the reasoning for each system in writing – that is the basis for any later review.
- 1
Is it an AI system?
Decision pointCheck the elements of Art. 3(1): machine-based, some degree of autonomy, outputs inferred from inputs. Inference is what matters – through machine learning or logic- and knowledge-based approaches, for instance – as opposed to hard-coded rules. A spreadsheet macro with if-then rules is not an AI system; a model that classifies incoming invoices is.
- Who
- Business unit, IT
- Output
- Classification with reasoning
- Basis
- Art. 3(1)
- 2
Does the regulation apply to this use?
Decision pointCheck the exclusions in Art. 2: exclusively military purposes, scientific research, testing before placing on the market (testing in real-world conditions is not excluded), purely personal use, open-source systems outside high-risk, Article 5 and Article 50 cases. For a company in normal operation, usually no exclusion applies.
- Who
- Compliance, legal
- Output
- In scope yes/no
- Basis
- Art. 2
- 3
What is your role?
Determine the role per system. If you develop it or have it developed and place it on the market or put it into service under your own name, you are the provider – own use included. If you use it under your authority, you are the deployer. Importers and distributors are links in the supply chain. Also check the role switch under Art. 25.
- Who
- Compliance, procurement
- Output
- Role(s) per system
- Basis
- Art. 3(3)–(7), Art. 25
- 4
Is it a prohibited practice?
Decision pointCompare the purpose with the list in Art. 5. If a prohibition applies, the assessment ends there: the system may not be placed on the market or used in that form. Typical borderline cases are emotion recognition at work and systems that deliberately exploit the vulnerabilities of certain groups.
- Who
- Compliance, legal
- Output
- Stop or continue
- Basis
- Art. 5
- 5
Is the system high-risk?
Decision pointTwo routes lead into the high-risk regime: as a safety component or product under Annex I requiring third-party conformity assessment (Art. 6(1)), or as a use case in Annex III (Art. 6(2)). For Annex III, check the exception in para. 3 – and document the result before the system is placed on the market.
- Who
- Compliance, business unit
- Output
- High-risk yes/no, route
- Basis
- Art. 6, Annex I, Annex III
- 6
Does the use trigger transparency duties?
Whatever the outcome of step five, check Art. 50: direct interaction with people, synthetic content, deepfakes, AI text on matters of public interest, emotion recognition or biometric categorisation. These duties apply on top of the high-risk obligations; they do not replace them.
- Who
- Product, marketing, compliance
- Output
- Transparency cases
- Basis
- Art. 50
- 7
Do you provide an AI model?
If you offer a general-purpose AI model – for example your own language model that others integrate – the model obligations apply as well. Anyone who merely builds a third-party model into a system is a downstream provider of that system (Art. 3(68)); under Art. 53(1)(b) the model provider must make information and documentation available to it.
- Who
- Engineering, legal
- Output
- Model obligations yes/no
- Basis
- Arts. 51–55
- 8
Combine the list and add dates
Merge the results into a list of obligations per system and role, and give every item its start date. Art. 4 always belongs on the list for providers and deployers. Decide when the assessment is repeated – for example when the system, its purpose or the supplier changes.
- Who
- Compliance
- Output
- Dated list of obligations
- Basis
- Art. 4, Art. 111, Art. 113
Recommendation: keep negative results too (“not an AI system”, “exception under Art. 6(3)”). These are the decisions that supervisory authorities or customers question first.
Role matrix
Which role carries which areas of obligation
The matrix shows which areas of the AI Act are addressed to which role. It does not replace a case-by-case assessment, but it shows why the same classification means very different workloads for different companies.
Scroll the table sideways →
| Area | Provider | Deployer | Importer | Distributor |
|---|---|---|---|---|
| AI literacy (Art. 4) | yes | yes | not addressed – but a deployer where it uses AI itself | as importer |
| Prohibited practices (Art. 5) | no placing on the market, no putting into service | no use | no placing on the market | not named separately; Art. 5 covers placing on the market, putting into service, use |
| High-risk requirements (Arts. 8–15) | meet and demonstrate | use in line with instructions, human oversight | check the provider’s evidence | check CE marking and documents |
| Quality management, conformity assessment, CE (Arts. 16, 17, 43, 47, 48) | yes | – | check that it was done (Art. 23(1)) | check CE marking (Art. 24(1)) |
| Registration in the EU database (Art. 49) | Annex III except point 2; also with the Art. 6(3) exception | public authorities and Union bodies only (para. 3) | – | – |
| Keep logs | at least 6 months (Art. 19) | at least 6 months (Art. 26(6)) | – | – |
| Keep documents available | 10 years (Art. 18) | – | certificate, declaration, instructions for 10 years (Art. 23(5)) | information on reasoned request (Art. 24(5)) |
| Post-market monitoring, incidents (Arts. 72, 73) | set up the system, report serious incidents | inform provider, authority where needed (Art. 26(5)) | inform where there is a risk (Art. 23(2)) | inform where there is a risk, corrective action (Art. 24(4)) |
| Fundamental rights impact assessment (Art. 27) | – | public bodies, providers of public services, Annex III point 5(b)/(c) | – | – |
| Transparency (Art. 50) | disclosure of AI interaction, machine-readable marking | deepfakes, AI text, emotion recognition, biometric categorisation | not addressed | not addressed |
| General-purpose AI models (Arts. 53–55) | model providers only | – | – | – |
How to read it: “–” means the AI Act assigns no obligation of its own to that role in that area. The authorised representative (Art. 22) has no column; it carries out the tasks of a provider established in a third country towards the authorities on that provider’s behalf.
Determining roles
Provider or deployer? Drawing the line in practice
Provider
Under Art. 3(3), a provider is anyone who develops an AI system or has it developed and places it on the market or puts it into service under its own name or trade mark – whether for payment or free of charge. Two points are often missed. First, commissioning the development is enough. Second, under Art. 3(11), putting into service expressly includes supply for own use. A company that has an agency build an internal recruitment tool and then uses it is therefore both provider and deployer.
Deployer
Under Art. 3(4), a deployer is anyone who uses an AI system under its authority, except in the course of a personal, non-professional activity. That describes most companies that buy software with AI features: the CRM with lead scoring, the accounting tool that reads receipts, the chat assistant in the browser. Employees normally act on behalf of the company, so the deployer obligations sit with the company.
Importer
Under Art. 3(6), an importer is a person established in the Union who places on the market an AI system bearing the name or trade mark of a company established in a third country. Under Art. 3(7), a distributor is anyone else in the supply chain who makes a system available on the Union market. A German systems integrator that sells US software exclusively in the EU is typically the importer; a reseller who sells the same software on is a distributor.
Role per system
A role is not a property of the company but of an activity relating to a specific system. A software house is the provider of its own product, the deployer of the translation AI it buys in and perhaps the distributor of a partner’s solution it resells. Record the role in the AI inventory per entry, not once per company.
Article 25 AI Act
Role switch: when deployers and distributors become providers
Art. 25(1) turns distributors, importers, deployers or other third parties into the provider of a high-risk AI system in three situations – with all the obligations of Art. 16. The examples are simplified and for illustration only.
Point (a)
Own name or trade mark
Example: a recruitment services firm sells a bought-in high-risk screening module under its own product brand. It is considered the provider – without prejudice to contractual arrangements that allocate the obligations differently.Point (b)
Substantial modification
Example: a deployer retrains a high-risk credit-scoring system on its own data in a way that was not foreseen in the original conformity assessment and affects compliance with the requirements (Art. 3(23)). If the system remains high-risk, the deployer becomes its provider.Point (c)
New intended purpose
Example: a company uses a general chat assistant to pre-sort applications and assess candidates. A system that was not high-risk becomes a use case under Annex III point 4 – and the company becomes the provider of that high-risk system.Para. 2
What it means for the original provider
The initial provider is no longer the provider of that specific system but must cooperate closely: supply technical documentation, known limitations and targeted technical access – unless it clearly specified that its system is not to be changed into a high-risk system.Para. 3
Product manufacturers under Annex I
Where the high-risk system is a safety component of a product under Annex I Section A, the product manufacturer is considered the provider if the system is placed on the market with the product under its name, or put into service under its name afterwards.Para. 4
Contracts with suppliers
Anyone supplying components, models or services for a high-risk system concludes a written agreement with its provider on information, capabilities and technical access. Suppliers of open-source components that are not general-purpose AI models are exempt.
Recommendation: check Art. 25 whenever something significant changes – a new purpose, retraining, reselling under your own brand. A role switch does not happen by contract; it happens through what you do with the system.
Obligations finder
Which obligations apply to your AI system?
Answer the questions for one specific system in your role. The result lists the relevant obligations with article and start date. Everything is evaluated in your browser; no answers are transmitted.
List of obligations
0 of 6 questions answered
Answer the questions – the list builds up with every answer.
Simplified guidance, not legal advice. The finder does not cover every special case.
Timeline
From when each obligation applies
The AI Act entered into force on 1 August 2024 but applies in stages. The Digital Omnibus (Regulation (EU) 2026/1744, in force since 27 July 2026) moved the high-risk dates.
Scroll the table sideways →
| Date | What applies | Basis |
|---|---|---|
| 2 February 2025 | Chapters I and II: definitions, AI literacy (Art. 4), prohibited practices (Art. 5) | Art. 113(3)(a) |
| 2 August 2025 | Obligations for general-purpose AI models (Chapter V), governance, penalties | Art. 113(3)(b) |
| 2 August 2026 | General date of application, including the transparency duties in Art. 50 | Art. 113(2) |
| 2 December 2026 | New prohibitions in Art. 5(1)(ba) and (bb); Art. 50(2) for generative systems placed on the market before 2 August 2026 | Art. 113(3)(a), Art. 111(4) |
| 2 August 2027 | Deadline for providers of AI models placed on the market before 2 August 2025 | Art. 111(3) |
| 2 December 2027 | High-risk rules (Chapter III Sections 1–3) for systems under Art. 6(2) and Annex III | Art. 113(3)(c)(i) |
| 2 August 2028 | High-risk rules for systems under Art. 6(1) and Annex I | Art. 113(3)(c)(ii) |
Context: post-market monitoring (Art. 72), serious incident reporting (Art. 73) and the right to an explanation (Art. 86) are tied to high-risk AI systems and therefore only become relevant in practice with the high-risk dates. A full overview is on the AI Act deadlines page.
Common mistakes
Seven shortcuts that lead to the wrong answer
“We only use ChatGPT, so nothing applies.”
As a deployer you are at least subject to Art. 4. If you publish deepfakes, or text on matters of public interest without editorial responsibility, Art. 50(4) applies as well. And if you use the assistant to select candidates, you may become the provider of a high-risk system under Art. 25(1)(c).“The provider is always whoever built the model.”
The provider of the AI system is whoever places it on the market or puts it into service under its own name. The model maker is the provider of the model, with its own obligations under Art. 53. If you build a third-party model into your own customer service bot, you are the provider of the bot.“High risk is a question of technology.”
The intended purpose decides. The same language model is harmless when summarising meeting notes and an Annex III point 4 use case when it is used to assess employees.“The Art. 6(3) exception settles everything.”
The exception requires that there is no significant risk and that one of four conditions is met. It is ruled out for profiling. It must be documented before placing on the market, registration under Art. 49(2) still applies, and Art. 50 and Art. 4 continue to apply.“We are only a distributor, so this does not concern us.”
Before making a system available, distributors must check the CE marking, the declaration of conformity and the instructions, and hold back where in doubt (Art. 24). Selling under your own brand makes you the provider under Art. 25.“We have until 2027.”
Prohibitions and AI literacy have applied since February 2025, model obligations since August 2025 and transparency duties since August 2026. The Omnibus only postponed the high-risk rules.“Assessed once, assessed for good.”
A new intended purpose, retraining or a change of supplier can change both role and classification. The assessment therefore needs a fixed review cycle and has to be repeated after every change.
Borderline cases
Situations that need a closer look
- Providers in third countries: the AI Act also applies where only the output of the system is used in the Union (Art. 2(1)(c)). For high-risk systems and AI models, an authorised representative in the Union must be appointed (Art. 22, Art. 54).
- Products under Annex I Section B (e.g. civil aviation, motor vehicles, marine equipment, rail, machinery under Regulation (EU) 2023/1230): under Art. 2(2), only Art. 6(1), Art. 60a and Arts. 102 to 112 apply; the requirements are taken over through the sector legislation.
- Safety components: systems used exclusively for non-safety purposes such as user assistance, performance optimisation or quality control are not safety components under Art. 6(1a); systems whose failure would endanger health and safety are (para. 1b).
- Legacy systems: high-risk systems placed on the market before the relevant date only fall under Chapter III if their design is significantly changed afterwards (Art. 111(2)). Systems intended for public authorities must comply by 2 August 2030.
- Open-source systems: excluded under Art. 2(12) – unless they are placed on the market or put into service as high-risk systems or as systems falling under Art. 5 or Art. 50.
- SMEs and small mid-caps: the obligations apply to them in principle; relief exists for the quality management system (Art. 17(2)) and technical documentation, and for fines the lower amount applies (Art. 99(6), (6a)).
Recording the result
What a traceable assessment of obligations contains
Recommendation from practice: record the assessment for each system so that a third party can follow it without asking questions.
- System description with intended purpose, user group and affected persons
- Reasoning on Art. 3(1) and on possible exclusions under Art. 2
- Role per system, including the check for a role switch under Art. 25
- Result on Art. 5, with the prohibitions that were checked
- High-risk route (Annex I, Annex III or none) and, where relevant, the Art. 6(3) assessment
- Transparency cases under Art. 50 and how they will be implemented
- List of obligations with article, responsible person and start date
- Date of the assessment, people responsible and next review date
Fines for context (Art. 99): up to €35 million or 7% of worldwide annual turnover for breaches of Art. 5; up to €15 million or 3% for breaches of, among others, Arts. 16, 22–26 and 50; up to €7.5 million or 1% for incorrect or incomplete information to authorities.
Legal Logic in SimpleAct
Generate the list of obligations for each AI system in your inventory
The Legal Logic module answers, for every AI system you record, which obligations of the AI Act apply. It is based on a questionnaire on role (provider or deployer), risk category and use case. The result lists the relevant obligations by article – with the reasoning for why each one applies or does not.
- Questionnaire on role, risk category and use case per system
- Result sorted by article, with reasoning for “applies” and “does not apply”
- Each obligation linked to the matching module in SimpleAct
- Set a review cycle, e.g. every 12 months
- The review owner is reminded automatically when the next review is due
- Result as the basis for the audit playbook and governance
- Before: rule-based risk assessment with three assessment blocks
- After: compliance checklist per risk class with evidence and status
Legal Logic is available from the Professional plan (€349 per month, €279 with annual billing, plus VAT) and is marked PRO in the app. The assessment is a structured aid and is not legal advice.
FAQ
Frequently asked questions on working out obligations
More questions? We're happy to help. Send email · Try 7 days (credit card required)
Sources and status
Last checked on · SimpleAct editorial team
- Regulation (EU) 2024/1689 (AI Act), EUR-Lex
- Regulation (EU) 2026/1744 (Digital Omnibus), EUR-Lex
- Article 2 AI Act – Scope
- Article 3 AI Act – Definitions
- Article 5 AI Act – Prohibited practices
- Article 6 AI Act – Classification as high-risk
- Article 25 AI Act – Responsibilities along the value chain
- Article 26 AI Act – Obligations of deployers
- Article 50 AI Act – Transparency obligations
- Article 53 AI Act – Obligations of model providers
- Article 111 AI Act – Systems and models already on the market
- Article 113 AI Act – Entry into application
This page is for guidance only and is not legal advice. Examples are simplified. The text published in the Official Journal of the EU is authoritative.
Changes
- 2026-10-06 – Page rebuilt: assessment order in eight steps, role matrix, role switch under Art. 25, timeline after the Digital Omnibus and obligations finder.
- 2026-04-01 – First published.
Work out obligations for your whole AI inventory
Record your AI systems, assess risk and role and derive the list of obligations for each system – with a review cycle and named owners.