IA

AI Act: what obligations for your business in 2026?

Prohibitions already in force, the high-risk systems regime applicable from 2 August 2026, fines of up to €35m: a practical briefing on the AI Act and a 7-step action plan to bring your business into compliance.

Contents
Schedule a discussion

Reading time:

12 min

The European regulation on artificial intelligence, known as the AI Act (Regulation (EU) 2024/1689), entered into force on 1 August 2024. It applies in successive waves. 2026 is the pivotal year: on 2 August 2026, most of the obligations become applicable, including those that bear on high-risk AI systems. If your business develops, integrates or uses AI, now is the time to act. Here is how.

Who is covered by the AI Act?

The regulation does not only target AI solution vendors. It distinguishes several roles, each with its own obligations:

  • The provider: the party that develops an AI system or has it developed, then places it on the market under its own name or trade mark. This is the most exposed role.
  • The deployer: any business or body that uses an AI system under its own authority in a professional context. An SME using a CV screening tool is a deployer.
  • The importer and the distributor: those who place on the European market systems developed outside the EU, or who sell them.

A key point: the text has extraterritorial reach. A US provider whose system is used in the Union, or whose AI-generated output is used in the Union, falls within the scope of the regulation. French businesses buying foreign solutions must therefore check their suppliers' compliance, because in practice the deployer remains liable within its own scope.

Watch out, too, for a change of role: a deployer that substantially modifies a high-risk system, or sells it under its own brand, becomes a provider with all the associated obligations. This is a classic trap for integrators and SaaS vendors that customise third-party models.

The application timeline: where do things stand in summer 2026?

Four deadlines structure the roll-out:

  • 2 February 2025: prohibited practices + AI literacy.
  • 2 August 2025: GPAI.
  • 2 August 2026: transparency obligations (Article 50), governance and several other provisions remain applicable.
  • 2 December 2027: "stand-alone" high-risk AI systems (Annex III: recruitment, credit, HR, etc.).
  • 2 August 2028: high-risk AI systems embedded in regulated products (Annex I: medical devices, machinery...).

In other words: as you read this, the prohibitions and the GPAI rules already apply, and the bulk of the high-risk regime kicks in within a few weeks. The time for watching is over; operational compliance has begun.

The four risk levels

The AI Act takes a risk-based approach. Four categories, four regimes:

1. Prohibited practices

Article 5 bans certain practices outright, whatever the sector. Among them:

  • subliminal or deceptive manipulation causing significant harm;
  • exploitation of vulnerabilities linked to age, disability or social situation;
  • general social scoring;
  • emotion recognition in the workplace and in educational institutions, except for medical or safety reasons;
  • untargeted scraping of facial images from the internet to build facial recognition databases;
  • biometric categorisation aimed at inferring sensitive data (origin, opinions, sexual orientation).

These prohibitions cover use cases that are more widespread than one might think. An HR tool claiming to analyse employees' emotional engagement on video calls potentially falls under the ban. The risk is not theoretical: the highest penalty ceiling in the regulation applies.

2. High-risk systems

This is the heart of the regulation. There are two routes to qualification: systems embedded in regulated products (Annex I) and systems in the sensitive areas listed in Annex III. This second list directly concerns service businesses:

  • recruitment, worker management and access to employment;
  • creditworthiness assessment and pricing in life and health insurance;
  • access to essential public and private services;
  • education and vocational training;
  • biometrics, critical infrastructure, law enforcement, justice, migration.

A nuance worth knowing: Article 6(3) allows a system to escape high-risk classification if it only performs a narrow procedural task or improves the result of a previously completed human activity, without materially influencing the decision. This exemption must be documented and the system registered. It never applies if the system performs profiling of natural persons.

3. Limited-risk systems: transparency

Article 50 imposes transparency obligations: informing people that they are interacting with an AI (chatbots), marking AI-generated or manipulated content (including deepfakes), and disclosing the use of emotion recognition or biometric categorisation systems where these are lawful. These obligations apply from 2 August 2026.

4. Minimal risk

Everything else: spam filters, AI in video games, most productivity tools. No specific obligation, but voluntary codes of conduct are encouraged, and the GDPR continues to apply as soon as personal data is processed.

High-risk systems: the concrete obligations

If you are a provider

The provider of a high-risk system must, before placing it on the market:

  • set up a documented risk management system, maintained throughout the life cycle (Article 9);
  • ensure the quality and governance of the data used for training, validation and testing: relevance, representativeness, handling of bias (Article 10);
  • compile complete technical documentation in line with Annex IV (Article 11);
  • ensure automatic logging of events (Article 12);
  • provide deployers with clear instructions for use: capabilities, limitations, performance, human oversight measures (Article 13);
  • design the system to allow effective human oversight (Article 14);
  • ensure accuracy, robustness and cybersecurity (Article 15);
  • carry out the conformity assessment, affix the CE marking and register the system in the EU database before placing it on the market;
  • set up post-market monitoring and report serious incidents to the authorities.

If you are a deployer

The deployer's obligations (Article 26) are lighter but very real:

  • use the system in accordance with its instructions;
  • assign human oversight to people who are trained and have the necessary authority;
  • ensure the relevance of input data where it controls that data;
  • keep the logs generated by the system for at least six months;
  • inform employees and their representatives before deploying a high-risk system in the workplace;
  • inform the people concerned when a decision about them is taken or assisted by the system;
  • for public bodies and certain private players providing essential services: carry out a fundamental rights impact assessment (FRIA, Article 27).

General-purpose AI models (GPAI)

Since 2 August 2025, providers of general-purpose AI models must maintain technical documentation, provide information to downstream integrators, put in place a copyright compliance policy and publish a summary of training content. Models presenting a systemic risk bear stronger obligations: evaluations, adversarial testing, incident reporting, cybersecurity. A code of practice supports implementation.

For a French business, the practical issue is contractual: if you build a product on a third-party LLM, demand the documentation provided for by the regulation and check your model provider's commitments. Your compliance file depends on it.

AI Act and GDPR: two texts that apply together

The AI Act does not replace the GDPR; it adds to it. As soon as an AI system processes personal data, both bodies of law apply in parallel:

  • Legal basis and information: each AI-related processing operation (training, inference, continuous improvement) must rest on a legal basis under Article 6 GDPR and be disclosed to the individuals concerned.
  • Automated decision-making: Article 22 GDPR governs fully automated decisions producing legal or similarly significant effects. Human intervention, the right to express one's point of view, the right to contest: these safeguards combine with the human oversight required by the AI Act.
  • DPIA and FRIA: the GDPR impact assessment (Article 35) remains mandatory for high-risk processing. The AI Act FRIA can build on the existing DPIA, but the two exercises are not the same.
  • CNIL: since 2024 the CNIL (French data protection authority) has published a series of recommendations on the development of AI systems (building training datasets, legitimate interest, informing individuals, exercising rights). They are the practical reference for reconciling innovation and compliance in France.

On national governance, the framework is taking shape: the CNIL is expected to play a central supervisory role for systems involving personal data, alongside other sector regulators. As the precise designation of the competent French authorities is still being finalised, have this point checked when you make your decisions.

Penalties: a ceiling above the GDPR

The AI Act penalty regime goes further than the GDPR's:

  • up to €35 million or 7% of worldwide annual turnover for using a prohibited practice;
  • up to €15 million or 3% for breaching the other obligations (high risk, transparency, GPAI);
  • up to €7.5 million or 1% for supplying incorrect information to the authorities.

SMEs benefit from an adjusted cap (the lower of the two amounts applies). But the most immediate risk is not only the administrative penalty: it is the commercial risk. Large clients are already writing AI Act compliance clauses into their tenders and contracts. Being unable to document your compliance means losing contracts.

The 7-step action plan

Here is the roadmap we recommend to our clients:

  1. Map: list every AI system developed, bought or used in the business, including AI features embedded in your SaaS tools. This is the most common blind spot.
  2. Classify: for each system, determine your role (provider, deployer) and the risk level. Document the reasoning, especially if you rely on the Article 6(3) exemption.
  3. Rule out the prohibited: check immediately that no use falls under Article 5. If in doubt about a behavioural or emotional analysis tool, suspend it and have it reviewed.
  4. Close the gaps: for high-risk systems, start the upgrade (documentation, human oversight, logging, informing employees and the CSE (works council) where applicable).
  5. Secure the contracts: audit your supplier contracts. Require compliance warranties, technical documentation, instructions for use and commitments to cooperate in the event of an inspection. Pass the obligations on in your T&Cs if you are a software vendor.
  6. Train: the AI literacy obligation in Article 4 has applied since February 2025. Train the teams that use or supervise the systems, and keep records of that training.
  7. Govern over time: appoint an AI lead, align their role with the DPO's, set up an intake procedure for every new AI tool, and monitor the implementing acts and harmonised standards to come.

How do you prove compliance?

Complying is not enough; you must be able to demonstrate it. Three tools structure the evidence:

  • Harmonised standards: the European standardisation bodies (CEN-CENELEC) are drafting the technical standards that will implement the regulation's requirements. A system that complies with a harmonised standard published in the Official Journal of the EU will benefit from a presumption of conformity. Follow their publication closely: they will set the standard expected by the authorities and by your clients.
  • Documentation: register of AI systems, classification analyses, technical documentation, instructions, training records, logs. In an inspection, this file is the first thing examined. A business that cannot produce its mapping starts at a serious disadvantage.
  • The contract: between provider and deployer, the allocation of obligations must be in writing. Compliance warranties, the obligation to supply documentation, assistance in an inspection, audit clause, treatment of system updates: all points to negotiate before signing, not after the incident.

An often overlooked point: updates. A system that is compliant on the day it is deployed may cease to be so after a substantial change to the model. Provide contractually for prior notice of any significant modification and build the compliance review into your product cycle.

Sector by sector: the friction points

Some sectors stack regulations. In healthcare, AI software may fall simultaneously under the AI Act, the Medical Devices Regulation and the specific rules on health data: cross-classification is an exercise in its own right, which we will cover in a forthcoming article. In human resources, Annex III places recruitment and employee assessment in the high-risk category, with, in France, the additional interplay of the French Labour Code and consultation of the CSE. In finance and insurance, creditworthiness scoring and health or life insurance pricing are expressly targeted. If your business touches one of these areas, generic compliance will not be enough: the difficulty lies in where the texts intersect, and that is where legal support really pays off.

Key takeaways

The AI Act is no longer a distant prospect. The prohibitions have applied since February 2025, the GPAI rules since August 2025, and the high-risk systems regime becomes fully applicable on 2 August 2026. Compliance is built in months, not weeks: mapping, classification, contracts, documentation, training.

Mirabile Avocat supports businesses, software vendors and start-ups in their AI Act and GDPR compliance: classification audits, contract drafting, compliance documentation, dealings with the authorities. Want to know where you stand? Take the test below, then let's talk.

To learn more

We only use ChatGPT or an equivalent internally. Does the AI Act apply to us?

Yes, at the very least through the AI literacy obligation applicable since February 2025, and through the GDPR if personal data is entered into the tool. An internal usage policy and team awareness training form the minimum baseline. If the tool is used for sensitive HR or customer decisions, high-risk classification must be examined.

Our AI solution is developed by a supplier. Who is liable?

The supplier is in principle the provider, you are the deployer, and each bears its own obligations. But if the solution is sold under your brand or substantially modified by you, you switch to the provider role with all the associated obligations. The contract must set out this allocation precisely.

What penalties does the AI Act provide for?

Up to €35 million or 7% of worldwide annual turnover for using a prohibited practice, €15 million or 3% for breaching the other obligations, and €7.5 million or 1% for supplying incorrect information to the authorities. SMEs benefit from an adjusted cap.

Are there support measures for SMEs?

Yes. The regulation provides for regulatory sandboxes in which innovative systems can be tested under supervision, priority access for SMEs and reduced conformity assessment fees. This is a lever to build into your product strategy.

When does the AI Act apply?

The regulation entered into force on 1 August 2024. The prohibitions and the AI literacy obligation have applied since 2 February 2025, the rules on general-purpose models since 2 August 2025, most of the high-risk systems regime from 2 August 2026, and systems embedded in regulated products from 2 August 2027.

Still have questions?

Our team is available!

Have a question?

Your information will remain strictly confidential.
Thank you! We will get back to you shortly. If you'd like to speed things up, schedule a time with me directly here:
Schedule a 15-minute call
Oops! Something went wrong while submitting the form.
Homme en costume bleu foncé avec cravate et pochette blanche, bras croisés, regardant vers l'avant.

Ressources

Aller plus loin

00
article(s) affiché(s) sur
00

7 min

GDPR DPO: duties, responsibilities and real cost for a business
Since the General Data Protection Regulation (GDPR) came into force in May 2018, the role of Data Protection Officer (DPO) has established itself as a central pillar of digital compliance. Yet many directors of micro-businesses, SMEs and sta

6 min

Drafting and posting your legal notices on Shopify
Mandatory in France, notably on e-commerce sites built on Shopify, a website's legal notices are information that allows internet users to know who they are dealing with and how they can get in contact with the website's owners. These

4 min

Software licence agreement - Lawyer
Software licences play an essential role in protecting the rights of software publishers and in regulating the use of the software by users. They define the conditions under which a developed software may be used, modified, or distributed, while

4 min

Discounts, rebates, commercial cooperation: what are the differences and what is at stake?
What is at stake and what are the differences between discounts, rebates and commercial cooperation?

4 min

Regulatory developments in franchising in 2025: what franchisors and franchisees absolutely need to know
The world of franchising is constantly experiencing legal developments that can significantly impact the relationships between franchisors and franchisees. In 2025, several major regulatory changes have reshaped the landscape of this popular business model.

17 min

GTC of Use and GTC of Sale: differences, articulation and mistakes to avoid to secure your digital business
For any director of a digital company, e-merchant, platform publisher or online service provider, the General Terms of Use (GTU) and the General Terms and Conditions of Sale (GTC) constitute the daily contractual foundation of the business. Yet these two documents are the subject of
Book an appointment
Book an appointment📆 15-Min Meeting