Docs

Regulations

AI Act

AI Act — Practical guide for companies and developers 🤖

The AI Act is not “the law that stops AI”. It’s the list of use cases where you can’t pretend nothing is happening. Knowing which box your system falls into is 90% of the work.

What it is and why it matters 🔍

Regulation (EU) 2024/1689 — known as the AI Act — is the world’s first comprehensive legal framework for artificial intelligence. It’s a regulation, so it’s directly applicable in all Member States without national implementation.

It was published in the Official Journal of the European Union on 12 July 2024 and entered into force on 2 August 2024 (the twentieth day after publication). It applies in full from 2 August 2026, but with a phased rollout that has already activated many provisions (see Timeline and deadlines).

The stated goal is twofold: to promote trustworthy, human-centric and innovative AI, while also protecting health, safety and fundamental rights against harmful uses of AI systems. The Regulation is built on a risk-based model: the riskier the use case for people, the stricter the obligations. No more, no less.

The key principle is proportionality by risk: not all AI systems are treated the same. Most everyday uses have light obligations; the bulk of the paperwork is concentrated on a few high-risk categories.

Who it applies to 🎯

The AI Act has an extraterritorial scope: what matters is not only where you are established, but where your systems’ output ends up.

It applies to (Art. 2):

Providers that place on the market or put into service AI systems or general-purpose AI (GPAI) models in the Union, wherever they are established;
Deployers (users) of AI systems established or located in the EU;
Providers and deployers in third countries if the output is used in the EU;
Importers and distributors of AI systems;
Product manufacturers that integrate an AI system into their product under their own name or trademark;
Authorised representatives of providers not established in the EU.

What stays out (Art. 2):

Systems used exclusively for military, defence or national security purposes;
Scientific research and development activities (and systems under research/testing before market placement, unless real-world testing);
Systems released under open-source licenses, unless placed on the market as high-risk or falling under prohibited practices or transparency obligations;
Purely personal, non-professional use.

Dual-use always lurks: a system developed for civilian purposes that ends up in national security can still fall under the Regulation depending on actual use.

The roles in the value chain 🧑‍🤝‍🧑

The AI Act distinguishes operators with different obligations:

Role	Who it is	Where the obligations are
Provider	The one who develops or has a system developed and places it on the market or puts it into service	Art. 16
Deployer (user)	The one who uses an AI system under their own authority, outside personal use	Art. 26
Importer	The one who places on the EU market a system from a third-country provider	Art. 23
Distributor	The one who makes a system available on the market without being its provider or importer	Art. 24
Product manufacturer	The one who integrates an AI system into their own product	Art. 25
Authorised representative	The mandatee of a non-EU provider	Art. 22

An operator can hold multiple roles at once (e.g. importer and distributor): in that case they accumulate all the related obligations.

The prohibited practices: the line you can’t cross ⛔

Art. 5 categorically prohibits certain AI practices, with no “good intention” exception. Non-compliance carries the highest penalty of all (up to €35,000,000 or 7% of worldwide annual turnover). In summary, these are prohibited:

Subliminal or manipulative techniques that materially distort a person’s behaviour, impairing their ability to make an informed decision;
Exploitation of vulnerabilities due to age, disability or a specific social or economic situation;
Social scoring based on behaviour or personal characteristics;
Predictive criminal-risk assessment based only on profiling (with exceptions for supporting human assessment grounded in objective facts);
Creation of facial-recognition databases via untargeted scraping of images from the internet or CCTV;
Emotion inference in the workplace and in education (except medical or safety reasons);
Biometric categorisation of individuals based on biometric data to infer race, political opinions, religion, sexual orientation, etc.;
Real-time remote biometric identification in publicly accessible spaces for law enforcement, except narrow cases (trafficking victims, terrorist threats, serious crimes) with prior authorisation.

Note: the use of biometric verification (authentication: “are you really you?”) is not prohibited. It’s real-time remote identification for surveillance purposes that is heavily restricted.

High-risk systems: where the real work is 📋

The practical heart of the AI Act is the high-risk AI system category (Art. 6). Almost all the documentation and technical obligations concentrate here.

A system is high-risk if:

Art. 6.1: it’s a safety component (or itself the product) of a product already subject to specific EU harmonisation legislation (machinery, medical devices, toys, lifts, etc. — list in Annex I);
Art. 6.2 + Annex III: it falls within one of the 8 critical sectors of Annex III (see below).

The Annex III sectors 🗂️

Biometrics (remote biometric identification, categorisation, emotion recognition) — if permitted by applicable law;
Critical infrastructure (safety components of critical systems: energy, traffic, water, gas, digital);
Education and vocational training (assessment and admission of students, exams, anti-cheating surveillance);
Employment and workers management (recruitment, selection, performance monitoring);
Access to essential private and public services (credit, health/life insurance, emergency response);
Law enforcement (credibility assessment, deepfake detection, crime analysis);
Migration, asylum and border control management (document verification, risk assessment);
Administration of justice and democratic processes (judicial assistance, legal research, influencing elections/referenda).

The exception: non-significant risk ⚖️

Art. 6.3 clarifies that an Annex III system is not high-risk if it does not pose a significant risk of harm to health, safety or fundamental rights (e.g. it does not materially influence decisions). This exception only applies if one of the following holds:

it performs a limited procedural task;
it improves the result of an already-completed human activity;
it detects patterns without replacing or influencing an already-completed human assessment;
it performs a preparatory task for a relevant assessment.

But beware: if the system profiles natural persons, it is always high-risk, no exceptions. And if you declare that your system is not high-risk, you must document the assessment before placing it on the market and register in the EU database.

The requirements for high-risk systems 🛡️

For those developing high-risk systems, the obligations are structured and documentation-heavy. Here are the main requirements for the provider (Art. 8-15):

Requirement	What it means in practice
Risk management system (Art. 9)	Assess and mitigate risks across the whole lifecycle, iteratively
Data governance (Art. 10)	Training/validation/test datasets that are relevant, representative, error-free and free from bias
Technical documentation (Art. 11)	Full description of the system, purpose, architecture, development, testing
Record-keeping (Art. 12)	Automatic logs to trace the system’s operation over time
Transparency and information to deployers (Art. 13)	Clear instructions on capabilities, limitations, risks and how to use
Human oversight (Art. 14)	Measures to let people monitor and intervene
Accuracy, robustness, cybersecurity (Art. 15)	Reliable performance, resilience to errors and attacks

The provider must also: set up a quality management system (Art. 17), draw up the EU declaration of conformity (Art. 47) and affix the CE marking (Art. 48), keep all documentation for 10 years (Art. 18-19) and register the system in the EU database (Art. 49) before placing it on the market.

The deployer, for its part, must (Art. 26): use the system according to instructions, ensure human oversight, monitor and report malfunctions/incidents, and when the system makes decisions affecting people, inform them they are subject to the use of an AI system and ensure the right to explanation.

Cost isn’t pointless bureaucracy: for a product already subject to CE marking (e.g. a medical device or a machine), many requirements integrate with the documentation you already produce. The AI Act adds a layer, it doesn’t replace it.

Transparency obligations for everyone else 🤝

For non-high-risk but still “interactive” systems, Art. 50 imposes lighter but still mandatory transparency obligations:

Chatbots and direct interaction: inform people they are interacting with an AI system (unless it’s obvious);
Synthetic content: mark outputs generated or manipulated artificially (audio, images, video, text) in a mechanically detectable format;
Deepfakes: make it known that the content is generated or manipulated artificially;
Emotion recognition / biometric categorisation: inform the people exposed;
Generated text on matters of public interest: reveal it is AI-generated, unless subject to human editorial review.

This is the part that most affects everyday web/app development: any chatbot, any image or text generator that publishes content for informational purposes falls here.

General-purpose AI (GPAI) models 🧠

General-purpose AI models (GPAI) — like the large language models — have a dedicated regime (Chapter V, Art. 50-56):

Basic obligations for all GPAI providers (Art. 53): technical documentation, information for downstream providers, a copyright-compliance policy (respecting the right-reservation under Directive 2019/790) and a public summary of the content used for training;
Open source: open-source GPAI models with public weights are exempt from the basic obligations, unless they present systemic risks;
Authorised representative: GPAI providers in third countries must appoint an EU representative (Art. 54), with an open-source exception (unless systemic risk).

Models with systemic risk (Art. 51) — presumed when training compute exceeds 10^25 FLOPs — have additional obligations (Art. 55): model evaluation, adversarial testing, mitigation of systemic risks, tracking and reporting of serious incidents, and adequate cybersecurity.

The 10^25 FLOPs threshold is the concrete way of saying “the biggest models in the world”. The Commission can update it via delegated acts and can also designate models below the threshold based on the criteria in Annex XIII.

Impact for companies 🏢

If your company uses or develops AI, the compliance path is:

Inventory of systems: list all AI systems in use or in development (chatbots, recommendation engines, recruiting tools, content generation, etc.);
Classification: determine for each system whether it is prohibited (Art. 5), high-risk (Art. 6+Annex III), transparency (Art. 50), GPAI, or low-risk;
Gap analysis: compare the current state against the obligations of the category;
Compliance plan: documentation, data governance, human oversight, security, for high-risk systems;
Role assignment grid: clarify whether you are a provider, deployer, importer or distributor (roles accumulate).

Penalties (Art. 99, for violations committed by undertakings):

Violation	Maximum penalty
Prohibited practices (Art. 5)	€35,000,000 or 7% of worldwide turnover
Provider/deployer/transparency obligations	€15,000,000 or 3% of worldwide turnover
Inaccurate information to notified bodies/authorities	€7,500,000 or 1% of worldwide turnover

For SMEs and start-ups, penalties are reduced to the lower of the amount or percentage (Art. 99.6). And the Regulation provides for regulatory sandboxes and innovation-support measures for small players.

Impact for developers 💻

What really changes in daily work:

Documentation as a requirement: for high-risk systems, technical documentation (data, architecture, testing) is not a side dish, it’s part of the deliverable;
Data governance: you must be able to demonstrate that training datasets are relevant, representative and free from bias. The “dataset README” becomes de facto mandatory;
Robustness and cybersecurity testing: accuracy, resilience to errors and attacks become verifiable requirements;
Traceability: automatic logs and records for monitoring high-risk systems;
Transparency: for chatbots and synthetic content, labelling and technical marking of outputs;
Copyright (GPAI): if you work on foundation models, respecting the rights reservation and the training-content summary.

A concrete example of a hidden edge: a GPAI provider’s API put into service for customer support falls under the chatbot transparency obligations; if the same API feeds a CV-screening system, you classify the system as high-risk (employment). The same model, two different regimes depending on the use case.

Timeline and deadlines 📆

The AI Act enters into force and applies in a phased manner:

Date	What applies
2 August 2024	Entry into force (20 days after publication in the OJ of 12.7.2024)
2 February 2025	Chapters I and II (general provisions + prohibited practices): the Art. 5 prohibitions are already in force
2 August 2025	Chapter III sec. 4 (notified bodies), Chapter V (GPAI, except art. 101), Chapter VII (governance), Chapter XII (penalties, except art. 101) and Art. 78
2 August 2026	Full application of the Regulation
2 August 2027	Art. 6.1 (high-risk classification for products in Annex I) and related obligations

In practice: the prohibitions have been operational since 2025, the GPAI obligations since 2025, and most requirements on Annex III high-risk systems expire with the full application of 2026. For products subject to other harmonisation legislation (machinery, medical devices…) the further 2027 deadline applies.

Compliance checklist ✅

A concise checklist to get oriented.

Scoping and inventory

Inventory of all AI systems (in use and in development)
Classification by risk category (prohibited / high / transparency / GPAI / low)
Identified your own roles (provider, deployer, importer, distributor)

Prohibited practices (Art. 5)

No manipulative system / exploitation of vulnerabilities
No social scoring, no crime prediction based only on profiling
No biometric scraping, no emotion inference at work/study
No biometric categorisation on sensitive attributes
Real-time remote identification only in allowed and authorised cases

High-risk systems (only if applicable)

Risk management system (Art. 9)
Data governance and bias-free datasets (Art. 10)
Technical documentation (Art. 11)
Record-keeping / automatic logs (Art. 12)
Instructions for deployers / transparency (Art. 13)
Human oversight (Art. 14)
Accuracy, robustness, cybersecurity (Art. 15)
Quality management and CE declaration + marking (Art. 17, 47, 48)
Retention of documentation for 10 years (Art. 18-19)
Registration in the EU database (Art. 49)

Transparency (Art. 50)

Chatbot: user informed that they’re facing an AI
Synthetic content: outputs marked/detectable
Deepfake: made known
Text on public interest: declaration of artificial generation

GPAI (if applicable)

Technical documentation and downstream data (Art. 53)
Copyright compliance + training-content summary (Art. 53)
EU authorised representative if third-country provider (Art. 54)
If systemic risk: evaluation, adversarial testing, incident reporting (Art. 55)

Useful links 🔗

AI Act text (EUR-Lex) — Regulation (EU) 2024/1689
AI Act — European Commission — Official page and guidance
AI Office — The European office for AI Act implementation
EU AI action plan — AgID — Italian AI strategy
EU recommendations on trustworthy AI — AI ethics