Staff use an off-the-shelf copilot for drafts
Usually no special risk tier on those facts, but the organisation still needs contextual AI-literacy measures and must check privacy, confidentiality, copyright and output quality.
Baseline duties applyBuild · EU AI Act · Regulation plus enacted 2026 amendment
The EU AI Act is more than labels.
A support bot, applicant ranker, medical-device safety component and general-purpose model do not follow one checklist. Start with what the system does, where it reaches, which role you hold and when that route applies.
Operational guide, not legal advice. Legal position checked 2 August 2026. This page reads Regulation (EU) 2024/1689 together with Regulation (EU) 2026/1744, which entered into force on 27 July 2026.
5routes can stackPROHIBITED · HIGH-RISK · GPAI · TRANSPARENCY · OTHER
6+roles can carry dutiesPROVIDER ≠ DEPLOYER ≠ SUPPLY CHAIN
2024→30phased legal timelineENTRY INTO FORCE ≠ APPLICATION DATE
01 · Situation first
Start with the situation, not the tool.
The same model can sit in an ordinary drafting workflow, a prohibited practice, a high-risk decision system or several routes at once.
Software ranks job applicants
Recruitment and candidate evaluation are Annex III use cases. Prepare the provider and deployer controls before the amended high-risk rules apply on 2 December 2027.
High-risk routeAI performs a medical-device safety function
If the Section A product route and third-party health-or-safety conformity trigger are met, it is Annex I high-risk. The corresponding AI Act rules apply from 2 August 2028.
Product high-risk routeA customer talks to a support bot
The interactive-system provider normally designs an AI notice into the first interaction. Article 50 applies from 2 August 2026.
Transparency routeA company trains and releases a GPAI model
Model documentation, downstream information, copyright policy and a public training-content summary are separate from the rules for an app built on that model.
GPAI routeAI manipulates vulnerable people or creates social scores
Do not jump to disclosure. Screen the Article 5 prohibition first; the original prohibited-practice families have applied since 2 February 2025.
Stop and classifyA realistic CEO video or public-interest article is generated
A professional deployer may need a deepfake or text disclosure. Substantive review with an accountable publisher matters for the public-interest-text exception.
Article 50 routeAI materially influences a benefit, credit or other decision
The system may be high-risk, the deployer may owe notice, and the affected person may have complaint and explanation routes. Other equality and data-protection rights continue to apply.
Rights and high-risk routes02 · Act-wide pathfinder
Find the routes worth opening.
Choose every situation that fits. The result is a linked reading path, not a verdict that the system is lawful or compliant.
Every route remains available without the interactive pathfinder. Start with scope, roles and the prohibited-practice screen.
03 · Scope
First ask whether the Act reaches the system.
“AI” in marketing copy is not the test. The Act defines an AI system by how a machine-based system infers outputs that can influence physical or virtual environments.
AI system: a machine-based system designed to operate with varying autonomy, which may adapt after deployment and infers from inputs how to generate predictions, content, recommendations or decisions for explicit or implicit objectives.
Market, establishment or output
The Act can reach non-EU providers placing systems or GPAI models on the EU market, EU deployers, and third-country providers or deployers whose system output is used in the EU.
Purpose and phase matter
AI systems or models specifically developed and put into service solely for scientific R&D, and research, testing or development activity before market placement or putting into service, can be excluded. Other systems merely used in research remain covered; real-world testing is regulated, not generally excluded.
Only deployer duties drop out
A natural person’s purely personal, non-professional use is excluded from deployer obligations. It does not legalise fraud, harassment, privacy violations or unlawful content.
Not a blanket exemption
The system-level exemption does not cover Article 5, Article 50 or high-risk systems. GPAI models have a separate, narrower open-source treatment.
Other law continues
GDPR, ePrivacy, equality, employment, copyright, consumer, product-safety, sector and national law continue alongside the AI Act.
Document the inference test
Simple deterministic rules may fall outside the definition; sophisticated optimisation or learned prediction may not. Record the architecture, inputs, outputs, autonomy and inference method.
04 · Actors
Assign the role before the obligation.
One organisation can be provider, deployer and product manufacturer at the same time. Put the role in the system register instead of assuming the vendor owns every duty.
Provider
Develops or commissions a system or GPAI model and places it on the market or puts the system into service under its name, whether paid or free.
Deployer
Uses an AI system under its authority for professional or organisational activity. Employees acting under the organisation’s instructions are normally part of that deployment.
Importer and distributor
An importer brings a third-country provider’s branded system to the EU market; a distributor makes a system available elsewhere in the supply chain.
Product manufacturer
Places a product on the market or into service with an AI system under its own name. Product-law and AI Act routes can meet in one conformity process.
Authorised representative
An EU-established representative accepts a written mandate for a non-EU provider and performs the specified documentation, cooperation and contact duties.
Affected person
A person in the EU may receive notices and can use complaint or explanation routes. “Affected person” is not an operator role, but it changes the control design.
Role-transfer trap: an importer, distributor, deployer or other third party can become the provider by putting its name on the system, substantially modifying it, or changing the intended purpose so that it becomes high-risk. Contract for documentation and technical access before that happens.
05 · Horizontal duties
“Minimal risk” does not mean “nothing applies.”
General duties, other law and voluntary controls sit underneath the special risk routes.
Support practical AI literacy
Tailor measures to people’s knowledge, experience, training, use context and affected groups. The amended rule does not require one certificate or guarantee a fixed level for every person.
- Map roles, systems and foreseeable mistakes.
- Teach limits, escalation, data handling and review controls.
- Keep the materials, audience, date and refresh trigger.
Bias work does not waive data law
High-risk providers may exceptionally process special-category personal data where strictly necessary for Article 10 bias detection and correction. Providers and deployers of other AI systems or models, and deployers of high-risk systems, may do so only for biases likely to affect health or safety, harm fundamental rights or cause discrimination prohibited by Union law, and only with every Article 4a safeguard. This creates no duty to conduct bias work.
- Show why anonymised, synthetic or other data cannot work.
- Restrict access, reuse, sharing and retention.
- Document necessity, security and deletion.
Scale controls to the real risk
For systems outside high-risk rules, voluntary codes can still cover environmental performance, accessibility, inclusion, risk testing and governance. Other binding law may demand the same controls independently.
- Set a named owner and acceptable-use boundary.
- Measure quality, harm, cost and incident signals.
- Reclassify after purpose, model or audience changes.
06 · Route map
Risk tiers are routes, not a single score.
Check from the top down. A high-risk system can also have Article 50 duties, and an application can inherit obligations from a GPAI model without becoming the model provider.
01 · Stop screenProhibited practicesDo not proceed unless a precise statutory exception applies.
02 · Controlled routeHigh-risk systemsAnnex I products and Annex III uses take different routes on the enacted 2027 and 2028 application dates.
03 · Human-facing routeArticle 50 transparencyInteraction, synthetic-output marking, biometrics, deepfakes and public-interest text.
04 · Baseline routeOther or minimal riskLiteracy and other law still apply; voluntary controls remain useful.
Separate model axisGPAI modelsModel-provider duties can stack with any system-level route downstream.
07 · Article 5
Stop prohibited practices before designing controls.
Eight original practice families have applied since 2 February 2025. Two enacted additions apply from 2 December 2026; other law may already prohibit the same conduct.
- Harmful manipulation or deception
Subliminal, purposefully manipulative or deceptive techniques that materially distort an informed decision and cause, or are reasonably likely to cause, significant harm.
- Exploiting vulnerability
Using age, disability or a specific social or economic situation to materially distort behaviour in a significantly harmful way.
- Social scoring
Scoring people over time from social behaviour or personal characteristics where detrimental treatment is unrelated, unjustified or disproportionate.
- Individual criminal prediction
Assessing or predicting criminal-offence risk based solely on profiling or personality traits, subject to the narrow objective-facts and human-assessment boundary.
- Untargeted facial-image scraping
Creating or expanding facial-recognition databases by indiscriminately scraping the internet or CCTV footage.
- Emotion inference at work or school
Inferring emotions in workplaces or education institutions, except for a use intended for medical or safety reasons.
- Sensitive biometric categorisation
Using biometric data to infer race, political opinions, trade-union membership, religion, philosophical beliefs, sex life or sexual orientation, subject to narrow statutory boundaries.
- Real-time remote biometric identification
Law-enforcement use in publicly accessible spaces, except tightly limited victim, imminent-threat and serious-offence cases with legal, necessity, proportionality and authorisation safeguards.
Non-consensual intimate depictions
Realistic intimate or sexually explicit material of an identifiable person without the specified explicit consent. The amendment defines when provider and deployer conduct is caught and how safeguards matter.
Child sexual abuse material
AI systems used or designed for generation or manipulation of covered material, subject to the amendment’s purpose, foreseeability, safeguard and “without right” provisions.
08 · Chapter III
High-risk depends on product and intended use.
The enacted 2026 amendment moved Annex III duties to 2 December 2027 and Annex I product duties to 2 August 2028. Those are application dates, not permission to ignore applicable sector, data or equality law.
Regulated-product route
AI is a safety component of, or is itself, a product listed in Annex I and the product must undergo third-party conformity assessment for health or safety risks under that product law. Non-safety convenience or efficiency functions are excluded unless failure could endanger health or safety. Section A products take the direct Chapter III route, subject to any Article 2(13) delegated limitation; for Section B products, Article 2(2) routes most requirements through sector law.
Applies 2 August 2028Listed-use route
The intended use falls within a specific Annex III case in biometrics, infrastructure, education, employment, essential services, law enforcement, migration, justice or democratic processes.
Applies 2 December 2027Narrow non-high-risk finding
An Annex III system may be found not high-risk only where it does not pose a significant risk of harm to health, safety or fundamental rights—including by not materially influencing a decision outcome—and performs at least one listed narrow procedural, preparatory, completed-work or pattern-detection task. Profiling of natural persons remains high-risk. Document and register the finding.
Document and register
BiometricsCritical infrastructureEducation and trainingEmployment and workersEssential services and benefitsLaw enforcementMigration and bordersJustice and democracy
Prove the system across its lifecycle
- Continuous risk management and pre-market testing.
- Data governance where training, validation or test data are used.
- Technical documentation, automatic logs and instructions.
- Human-oversight design, accuracy, robustness and cybersecurity.
- Quality management, conformity assessment, declaration, CE marking and registration.
- Post-market monitoring, corrective action and serious-incident reporting.
Control the real use
- Follow instructions and assign competent, trained, authorised oversight.
- Keep controlled input data relevant and representative.
- Monitor, suspend and report risks or serious incidents.
- Retain controlled logs for at least six months unless other law says otherwise.
- Give workplace and affected-person notices where required.
- Complete registration, DPIA and fundamental-rights impact work where applicable.
Do not flatten Annex I: for Section B products, Article 2(2) directly retains only Article 6(1), Article 60a and Articles 102–112; Articles 57–59 follow only as sector law integrates the high-risk requirements. For Section A, Article 2(13) permits delegated limits on duplicated duties where product law provides equal or greater protection. Verify any applicable delegated act rather than assuming those duties are switched off.
Supply-chain control: importers and distributors verify provider, conformity, documentation, marking and storage or transport duties before making a high-risk system available. A rebrand, substantial modification or changed high-risk purpose can transfer provider responsibility.
Legacy high-risk transition: apart from specified large-scale Union IT systems, Article 111 generally brings a high-risk system placed on the market or put into service before its relevant Chapter III date into the Regulation only if its design is significantly changed from that date. High-risk systems intended for public-authority use must comply by 2 August 2030 in any case. Article 5 remains unaffected.
09 · Chapter V
A GPAI model is a separate compliance layer.
Calling a model through an API does not make every customer its provider. Training, adapting, releasing and integrating a model can create different roles along the value chain.
Document and enable downstream compliance
- Maintain model technical documentation.
- Give downstream system providers capability, limitation and integration information.
- Maintain a policy for EU copyright and rights reservations.
- Publish the required training-content summary using the AI Office template.
- Appoint an EU representative when required and cooperate with authorities.
Add evaluation, mitigation and security
A model is presumed to have high-impact capabilities above 1025 training FLOPs, or can be designated on equivalent capability or impact. Notify the Commission within two weeks after the threshold is met or known, subject to the rebuttal process.
- State-of-the-art evaluations and adversarial testing.
- Union-level systemic-risk assessment and mitigation.
- Serious-incident tracking and reporting.
- Model and physical-infrastructure cybersecurity.
The exception is limited
Qualifying free and open-source GPAI models can be exempt from some technical and downstream documentation and representative duties. Copyright policy and training-summary duties remain, and systemic-risk models do not receive the same exemption.
New GPAI models
Chapter V obligations apply to providers placing covered models on the EU market from that date.
Commission enforcement
The AI Office can enforce Chapter V and the GPAI fine regime. Code adherence is voluntary, not immunity.
Legacy GPAI models
Providers of models placed on the market before 2 August 2025 must bring them into compliance.
10 · Article 50 · applies from 2 August 2026
Transparency is one chapter, not the whole Act.
Each card names a different actor and trigger. A disclosure never makes a prohibited, unsafe or otherwise unlawful system lawful.
Interactive AI systems
Providers design systems for direct human interaction so people are told they are interacting with AI by the first interaction, unless that fact is genuinely obvious in context.
- Actor
- Provider
- Keep
- First-screen, message or audio capture; wording; context; version; accessibility test.
Machine-readable synthetic-output marking
Providers of systems generating synthetic text, audio, images or video make outputs marked and detectable using effective, interoperable, robust and reliable technical solutions as far as feasible.
- Actor
- Provider, including covered GPAI systems
- Boundary
- Standard editing, non-substantial alteration and narrow law-enforcement cases.
Emotion and biometric exposure
Deployers inform people exposed to emotion-recognition or biometric-categorisation systems and comply with applicable personal-data law. Check Article 5 first because some uses are prohibited.
- Actor
- Deployer
- Keep
- Exposure map, notice, purpose, legal assessment, data-protection record and complaint route.
Deepfake image, audio and video
Deployers disclose AI-generated or manipulated media that resembles existing people, objects, places, entities or events and could falsely appear authentic or truthful. Evident creative works receive a less disruptive disclosure accommodation, not silence.
- Actor
- Professional deployer
- Keep
- Authenticity assessment, expected audience, labelled final asset and placement.
Public-interest text
Deployers disclose AI-generated or manipulated text published to inform the public on matters of public interest, unless it receives qualifying human review or editorial control and a person or entity holds editorial responsibility.
- Actor
- Professional deployer
- Keep
- Public-interest decision, sources, substantive changes, final approval and accountable publisher.
Human review
“Someone looked at it” is not enough.
For the public-interest-text exception, a competent person reviews meaning, facts and completeness, can change or reject the text, approves the final published version and sits inside a process with identifiable editorial responsibility.
Meaning, accuracy and completeness—not only spelling or style.
Verify factual claims and correct unsupported content.
Approve, change or reject the substance.
Prevent unreviewed AI changes after approval.
Clear wording
Say what happened at first contact.
The Act does not mandate one sentence. Make the information clear, distinguishable, accessible and available no later than the first interaction or exposure.
You are chatting with an AI assistant.
This call is handled by an AI voice assistant.
This video contains AI-generated or AI-manipulated people or events.
This article was generated or altered with AI and has not received substantive human editorial review.
Official EU labels · optional
Use the symbols as labels—not as a compliance stamp.
The Commission makes three human-visible labels freely reusable without attribution. They support Article 50(4) disclosure; they are not the provider-side machine-readable marking required by Article 50(2).
Basic icon
AI involved
Use with a nearby plain-language label or accessible second layer that explains what AI did.
Fully AI-generated
AI generated
Use when the entire in-scope item was generated by AI, apart from prompting, with no human-created elements or qualifying editorial control.
Partially AI-modified
AI modified
Use when pre-existing human-made content was changed with AI into an in-scope deepfake or public-interest text.
Visible label ≠ machine-readable watermark. For Code signatories, provider-side marking uses signed metadata and, where applicable, an imperceptible watermark; there is no single universal Article 50(2) watermark file. These symbols are the human-perceptible disclosure layer. Non-signatories may use them, but must not imply that they signed the Code.
11 · Rights and remedies
Design the route back to a responsible human.
The AI Act does not create one universal right to a model explanation or compensation. It creates specific notices and remedies while other EU and national rights continue.
Notice of high-risk decision support
Where an Annex III high-risk system makes or assists decisions about people, the deployer informs them that they are subject to its use, subject to the Act’s scope and dates.
Meaningful explanation
A person subject to a deployer decision based on output from an Annex III high-risk system—excluding point 2 critical-infrastructure systems—can obtain a clear, meaningful explanation of the system’s role and the main decision elements where the decision has legal or similarly significant effects which they consider adverse to health, safety or fundamental rights. Statutory and parallel-Union-law limits apply.
Complaint to the authority
Any natural or legal person with grounds to suspect an infringement can complain to the relevant market-surveillance authority.
Protected reporting
EU whistleblower protections apply to reporting AI Act infringements. Downstream GPAI providers also have a specific complaint route to the AI Office.
GDPR and equality law remain
Access, information, objection, human intervention, non-discrimination and judicial remedies may arise under other law even where the AI Act route is unavailable.
Make contact findable
Give the affected person a decision owner, channel, response process, evidence-preservation rule and escalation path—not only a generic privacy inbox.
12 · Innovation and governance
Test under supervision, then keep owning the risk.
Sandboxes can improve legal certainty and evidence. They do not waive Article 5, data law, safeguards or liability for harm.
Controlled, time-limited development
Each Member State must have at least one national or jointly equivalent sandbox operational by 2 August 2027. The AI Office may establish a Union-level sandbox within its competence; the EDPS may do so for Union bodies.
- Agree a sandbox plan, scope, safeguards and supervision.
- Use written proof and exit reports as compliance evidence—not certification.
- Mitigate significant risks or expect testing to be suspended.
People are not an ungoverned test set
Covered high-risk testing outside a sandbox needs a plan, authority route, registration and oversight. Informed consent, reversibility, incident handling, time limits and data safeguards apply subject to precise exceptions.
- Separate laboratory validation from regulated real-world testing.
- Preserve consent, approvals, versions, incidents and outcomes.
- Keep a prompt recall, suspension and deletion process.
National authorities
Market-surveillance authorities enforce most system duties; notifying authorities oversee conformity-assessment bodies. Member States provide a single contact point.
European AI Office
The Commission, acting through the AI Office, exclusively enforces Chapter V and the specified Article 75(1) system categories, including same-undertaking GPAI-based systems and VLOP/VLOSE systems, subject to statutory sector and public-authority exceptions. This system competence reaches deployers only where they are also the provider or belong to the same undertaking; other deployers remain under national supervision.
AI Board and expert bodies
The European AI Board coordinates national application. The Scientific Panel and Advisory Forum add technical and stakeholder expertise.
EDPS and notified bodies
The EDPS supervises Union institutions. Designated notified bodies perform the third-party conformity work required for covered high-risk systems.
Sandbox safe harbour has limits: participants remain liable under applicable Union and national law for damage to third parties. Prospective providers who follow the sandbox plan and terms and act in good faith on competent-authority guidance are protected from AI Act administrative fines for sandbox infringements, but supervisory and corrective powers remain.
13 · Dates
Entry into force and application are different dates.
The Act entered into force on 1 August 2024. The timeline below reflects the enacted July 2026 amendment, not the earlier proposal.
Position from 2 August 2026
General provisions and Article 50 apply; GPAI enforcement is active. Annex III and Annex I high-risk lifecycle duties remain on the later dates enacted below.
Act enters into force
Regulation (EU) 2024/1689 enters into force, with phased application.
Definitions, literacy and original bans
Chapters I and II apply, including Article 4 and the eight original Article 5 practice families.
GPAI, governance and penalties
Chapter III Section 4, Chapter V, Chapter VII, Chapter XII and Article 78 apply, except Article 101; providers of GPAI models placed before this date retain the 2 August 2027 transition.
AI Omnibus enters into force
Regulation (EU) 2026/1744 enacts the revised dates and targeted changes; amended Articles 102–110 also apply from this date.
General date and Article 50
Most remaining provisions, Article 50 duties and Commission GPAI enforcement apply.
New bans and marking transition
The intimate-depiction and CSAM prohibitions apply. Providers of synthetic audio, image, video or text systems, including GPAI systems, placed on the market before 2 August 2026 must comply with Article 50(2)’s machine-readable marking duty by this date; the other Article 50 duties already apply from 2 August 2026.
Sandboxes and legacy GPAI
National sandboxes must be operational; pre-2 August 2025 GPAI models must comply.
Annex III high-risk duties
Chapter III Sections 1–3, except Article 6(5), apply to Article 6(2) listed-use systems, subject to Article 111’s legacy-system transition.
Annex I product high-risk duties
Chapter III Sections 1–3, except Article 6(5), apply to Article 6(1) systems, subject to Section B sector-law treatment, Article 2(13) delegated limitations and Article 111’s legacy-system transition.
Public-authority legacy systems
Providers and deployers of pre-existing high-risk systems intended for use by public authorities must comply by this date. AI components of Annex X large-scale IT systems placed on the market or put into service before 2 August 2027 must comply by 31 December 2030.
Up to €35m or 7%
Whichever ceiling is higher for an undertaking, subject to proportionality and the special SME treatment.
Up to €15m or 3%
Covers the specified duties in Articles 16, 22–26, 31, 33, 34 and 50, including the amended Article 25(2) and (4) value-chain cooperation duties.
Up to €7.5m or 1%
For incorrect, incomplete or misleading information supplied to a notified body or competent authority in reply to a request.
Up to €15m or 3%
A separate Article 101 ceiling for intentional or negligent Chapter V and enforcement failures: whichever is higher for an undertaking.
AI Office operator enforcement: for Article 75(1) operators, Article 75c can apply the €15m/3% band to any applicable AI Act infringement, even one not enumerated in Article 99(4), and to failures to comply with enforcement decisions, measures or binding commitments. Misleading replies use the €7.5m/1% band. Periodic payments can reach 5% of average daily income or worldwide annual turnover in the preceding financial year per day.
Union institutions have separate ceilings: Article 100 allows the EDPS to impose up to €1.5m for Article 5 infringements and up to €750,000 for other infringements.
14 · Operating contract
One register, then route-specific evidence.
Do not build separate spreadsheets for labels, literacy, GPAI and high-risk projects. Keep one source of truth and attach the evidence each route requires.
| Register field | Record | Acceptance check |
|---|---|---|
| System and purpose | Owner, version, model, inputs, outputs, users, affected groups and intended purpose | Describes the actual workflow, not “AI-powered” |
| Scope and EU link | AI-system or GPAI rationale, market, establishment, output use and any exclusion | Every exclusion cites facts and a legal basis |
| Roles and supply chain | Provider, deployer, importer, distributor, manufacturer, representative and affected people | Contracts provide required documents and technical access |
| Route classification | Article 5 screen, Annex I/III test, GPAI layer, Article 50 triggers and other law | Concurrent routes remain visible |
| Date and owner | Applicable date, transition, accountable role and decision authority | No bare “2026 deadline” or other relative-date language |
| Controls and evidence | Literacy, data, testing, oversight, notices, conformity, review, logs and approvals | Each control has an artefact and a pass condition |
| Monitoring and change | Incidents, complaints, performance, model or purpose changes and review date | A trigger reopens classification before release |
Know every system
Include shadow use, vendor features, embedded product AI and retired versions still affecting people.
Stop before deployment
No release until scope, role, prohibition, route, date and control owner are recorded.
Reclassify material changes
New purpose, model, audience, data, autonomy, integration or brand can change risk and role.
Official sources · checked 2 August 2026
Use the law for the final decision.
Guidance and codes explain implementation but do not replace the Regulation. Check EUR-Lex for a consolidated text; where it does not yet include the amendment, read both Regulations together.
Regulation (EU) 2024/1689 — original binding AI Act →
Regulation (EU) 2026/1744 — enacted AI Omnibus amendment →
AI Act Service Desk — official implementation timeline →
AI Act Service Desk — official detailed compliance checker →
Commission — scope, high-risk, GPAI and governance Q&A →
Commission — AI-system definition guidelines →
Commission — draft high-risk classification guidelines as of 2 August 2026 →
Commission — Article 4 AI-literacy Q&A →
Commission — prohibited-practice guidelines →
Commission — GPAI provider guidelines →
Commission — final Article 50 transparency guidelines →
Commission — governance and enforcement →
The practical default
Classify the situation. Follow every route it triggers.