AI & data · 20 November 2026 · 7 min

EU AI Act compliance: what companies must do, and the real risks

EU AI Act compliance after the Digital Omnibus: what applies now, what has been postponed, and the risks the regulation does not cover (shadow AI, lock-in, security).

AI Act : obligations des entreprises, illustration d'un réseau de nœuds bleus et verts reliés sur fond clair

In short. For a company that uses AI (a “deployer”), EU AI Act compliance currently comes down to three things: building AI literacy across its teams, staying clear of prohibited practices, and being transparent about certain generated content. The heavy “high-risk” obligations have been pushed back to 2 December 2027, but they already cover everyday uses such as recruitment. Some of the most expensive risks sit outside the regulation: unsanctioned AI use, decisions handed over without review, supplier lock-in and data leaks. A simple governance framework, owned by senior management, addresses both.

EU AI Act compliance: where things stand

As of 19 November 2026. The EU Artificial Intelligence Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024. It applies in stages, and its timeline was amended in 2026 by the “Digital Omnibus on AI”, Regulation (EU) 2026/1744 of 8 July 2026, which entered into force on 27 July 2026. For a leadership team, EU AI Act compliance boils down to a timeline and a role.

Date What applies
2 February 2025 Prohibited AI practices and the AI literacy obligation (Article 4)
2 August 2025 Governance rules and obligations for providers of general-purpose AI models
2 August 2026 General application of the regulation, including transparency rules, with some exceptions
2 December 2026 New ban on systems generating non-consensual intimate content or child sexual abuse material (“nudification” apps); marking of generated content, for providers of systems placed on the market before 2 August 2026
2 December 2027 Obligations for Annex III high-risk systems (recruitment, education, access to essential services…)
2 August 2028 Obligations for high-risk systems embedded in regulated products (Annex I)

The postponements were agreed politically on 7 May 2026 and given final approval by the Council on 29 June 2026. A delay is not a pause: what has applied since February 2025 and August 2026 remains in force.

What the regulation expects from a company using AI

The AI Act distinguishes the provider, who develops and places an AI system on the market, from the “deployer”, defined in Article 3 as the person or body using an AI system under its authority, outside personal use. A mid-sized company that gives its staff a chat assistant, a CV screening tool or a writing copilot is a deployer. Its obligations depend on the use case, not on the tool.

For every use case

  • AI literacy (Article 4). Providers and deployers take measures to support the AI literacy of their staff and of people operating these systems on their behalf. The Digital Omnibus reworded this obligation without removing it (recital 8 of Regulation 2026/1744).
  • Prohibited practices (Article 5). Social scoring, manipulation and emotion recognition in the workplace, among others. They are rare in business, but you need to be able to spot them in a tool you buy.
  • Transparency (Article 50). The deployer of a system that generates or manipulates image, audio or video deep fakes must disclose that the content is artificial. The same applies to text published to inform the public on matters of public interest, unless it has undergone human review under editorial responsibility.

For high-risk use cases

Annex III classifies as high-risk, among others, systems intended for recruiting or selecting candidates, and those used to make decisions about employment relationships or to evaluate staff performance and behaviour. A CV pre-screening tool therefore falls into this category. From 2 December 2027, for systems put into service from that date or significantly modified afterwards (Article 111), Article 26 requires the deployer in particular to use the system according to its instructions, assign human oversight to people with the necessary competence and authority, monitor its operation, keep logs for at least six months and, as an employer, inform workers’ representatives and affected employees before putting it into service.

Penalties

Article 99 provides for fines of up to EUR 35 million or 7% of worldwide annual turnover for prohibited practices, and up to EUR 15 million or 3% for the breaches listed in Article 99(4), including the obligations of deployers of high-risk systems (Article 26) and transparency (Article 50). For SMEs, the lower of the two amounts applies; the Digital Omnibus extends this rule to small mid-caps, except for prohibited practices, and asks Member States to also take their economic viability into account.

Shadow AI: the risk leadership does not see

France’s national cybersecurity agency, ANSSI, defines shadow AI as the use of consumer generative AI tools without the approval or oversight of IT. In its AI cyber-risk analysis of February 2025, it notes that this practice can lead to leaks of confidential data and regulatory breaches.

It is the direct heir of shadow IT, with one difference: whatever is pasted into an online assistant leaves the company. A contract, a customer file, a piece of code or a bid response can end up with a service whose terms nobody has read. For an aerospace or space subcontractor bound by its customers’ confidentiality clauses, the stakes go well beyond compliance.

Banning is not enough: usage simply moves to personal phones. What works is to provide an approved tool, set out on one page what may and may not go into it, and explain why. It is also a practical way to support the AI literacy required by Article 4.

Misuse: when the tool decides instead of a person

The second risk is quieter: letting AI settle a question with no one checking. A contract summary that misses a clause, a plausible but wrong customer reply, a briefing note that becomes the basis for a decision. ANSSI’s analysis recommends building in human validation wherever necessary and matching a system’s level of autonomy to the risk analysis.

This becomes pressing with agents, systems that chain actions across other software. In its guide on securing agentic AI systems, ANSSI suggests sorting every action into three categories: autonomous execution, mandatory human approval, and forbidden. It asks that the approval mechanism stay out of the agent’s own reach.

Our view: the question is not “is the tool reliable?” but “who signs off?”. Any AI output that commits the company towards a customer, an employee or an authority needs a named human owner.

Maintainability and dependency: a supplier like any other

An AI use case that works today can degrade without anything changing on your side: the provider updates its model, changes its terms or prices, or retires a service. The prompts, automations and connectors built around it then become debt that is hard to take over, especially if one person built them.

ANSSI advises mapping the supply chain of an AI system (components, software, datasets) and questioning its dependency chain. It also recommends planning a degraded mode for business services without the AI system. For an executive, that means three plain questions:

  • Which processes would stop tomorrow if the AI tool became unavailable?
  • Can we recover our data and configurations in a usable format?
  • Who in-house can explain and maintain what has been built?

These are the questions you already ask about an ERP or a hosting provider. AI belongs in the same supplier management framework, with an exit clause once the use becomes critical. This is one strand of our AI governance work.

Security: what AI changes for attackers and for you

In its threat summary published on 4 February 2026, CERT-FR observes that attackers already use generative AI for social engineering, reconnaissance and writing malicious code. It has not identified any system able to run an attack end to end on its own. The report also mentions that more than 100,000 ChatGPT accounts were compromised between 2022 and 2023 by credential-stealing malware: an account’s history can hold whatever staff put into it.

CERT-FR also notes that AI tools open new doors: booby-trapped open models, flaws in the servers connecting assistants to applications (MCP), and software packages with AI-hallucinated names later registered by attackers. ANSSI further flags indirect prompt injection, hidden in a document or web page the tool reads, to make it disclose information or run commands.

Two principles from ANSSI’s agentic AI guide are worth putting on the wall: assume any data sent to a hosted model may be accessible to its provider, and deny by default any agent action that has not been explicitly authorised.

Where to start

AI Act compliance and risk control can be handled with one and the same set-up. Here is a realistic roadmap for the next six months.

  1. Inventory the use cases, including unofficial ones: a ten-minute anonymous survey of teams is enough to start. Record the tool, the data involved and the decision it supports.
  2. Classify each use: prohibited, high-risk (recruitment, staff evaluation…), subject to transparency, or routine.
  3. Select and roll out an approved tool, with contract terms actually read (data reuse, location, reversibility).
  4. Write a one-page policy: what can go into the tool, what must be reviewed, who signs off. Build awareness and training on that basis: these are the measures expected under Article 4.
  5. Name a human owner for every use that commits the company, and define a degraded mode for critical processes.
  6. Prepare for December 2027 if a use is high-risk: provider instructions, oversight, logs, informing employee representatives.

If you would like support, our AI strategy, governance and security service starts from this inventory and delivers a policy and a roadmap. For the wider picture, see also the evolving CIO and digital transformation, or get in touch.

Written with the help of AI tools, then reviewed, fact-checked and approved by the founder of Beneficis, who takes editorial responsibility for it.

Sources

Let's talk about your situation

Describe your context in a few lines: the founder replies to you personally, within 2 business days.

Tell us about your situation →Book a 30-minute call