The ceiling, for prohibited practices
This is the maximum for prohibited uses of AI. For transparency the ceiling is €15M or 3%. Penalties are imposed by national authorities and must be proportionate, with attention to SMEs.
The EU AI Act applies to every company that uses AI systems for work. Staff training and transparency are already mandatory; the high-risk obligations apply from 2027. We map the systems in use, prepare the documents and train the team: the technical side that lawyers don't cover.
Not only for those who develop AI. Also for those who use it in the company: ChatGPT, Copilot, a CRM with predictive features, a chatbot on the website. The 2026 Digital Omnibus eased some obligations and postponed the high-risk ones, but did not remove them.
This is the maximum for prohibited uses of AI. For transparency the ceiling is €15M or 3%. Penalties are imposed by national authorities and must be proportionate, with attention to SMEs.
Anyone using AI systems in the company must take measures to build their staff's skills. After the Omnibus there is no need to guarantee a precise level, but the obligation has been in force since February 2025.
The four documents we recommend to show what you have done: AI system inventory, risk classification, internal usage policy, transparency notice for customers and users.
If you answer yes to at least one of these, under the AI Act you are a deployer, that is a company using AI systems, and you have obligations. There are no size thresholds: it also applies to a 3-person SME.
Even occasionally. Even with a personal account ("shadow AI"). Yes → you are a deployer.
Lead scoring, suggested responses, candidate screening, call summaries. Yes → you are a deployer.
Even if it comes from a platform or a WordPress plugin. People must know they are talking to an AI: if you had the chatbot custom-built, the obligation is yours.
These are high-risk uses: human oversight, log retention, informing workers and the people affected. They apply from 2 December 2027.
A map of all the AI tools you use officially: enterprise licences, plugins in your SaaS, models embedded in CRMs, agents we have built for you.
AI systems the team uses without telling you: personal ChatGPT accounts, browser plug-ins, custom GPTs. Team survey + network/log audit.
For each system: minimal / limited / high risk according to the AI Act criteria. Documentation of the reasons for the choices made.
Audit-ready register + internal usage policy (what you can and cannot do with ChatGPT, sensitive data, escalation). Versioned.
A 2-3 hour session for the team, tailored to roles: what an LLM is, what it can and cannot do, risks, data not to enter. With the written record needed for Article 4.
From €250/month: register updates when you add tools, training refreshers, regulatory monitoring and policy updates.
The AI Act deadlines relevant to an Italian or Swiss SME, updated for the Digital Omnibus (EU Reg. 2026/1744).
Some uses are prohibited, such as emotion recognition in the workplace. Anyone using AI systems must take measures to train their staff: after the Omnibus without a set level, but with a record of what has been done.
Obligations for those who develop the models (OpenAI, Anthropic, Google and others): documentation and transparency. Nothing changes directly for those who use them.
People talking to a chatbot must know it, and deepfakes and certain AI-generated texts must be disclosed. From this date national authorities supervise and can impose penalties; in Italy, Law 132/2025 entrusts supervision to ACN.
Providers of generative systems already on the market must mark their content. The Omnibus also adds new prohibitions, including non-consensual intimate images.
Obligations for high-risk systems used, for example, to select staff, assess creditworthiness or access to essential services. It was 2 August 2026: the Omnibus postponed it.
AI systems that act as safety components of regulated products, such as medical devices and toys. Relevant if you develop products with AI inside.
A practical list to understand where you stand and what to do now, together with the deadline calendar. Designed for SMEs. No jargon, just things to do. We email it to you after the Quick Check.
Yes. The AI Act has no size thresholds: it applies to anyone using AI systems for work. For most SMEs the obligations are few and concrete: staff training, transparency, care with prohibited uses. 3-6 weeks are usually enough to put the basics in place.
Yes, if you use it for work. It's called "shadow AI" and it's one of the most critical points for SMEs: the company is still responsible for AI systems used for business purposes, even if it doesn't know they exist.
Because 80% of the work is technical, not legal: understanding what a CRM "with AI" really does, classifying a custom agent, taking inventory of browser plugins, writing policies the team understands. A lawyer is needed for the grey areas of interpretation. We do the rest and work with your lawyer where needed.
Since 2 August 2026 national supervisory authorities can carry out checks and impose penalties. In Italy, Law 132/2025 entrusts the supervision of AI systems to ACN, with AgID handling promotion and notification. Penalties must be proportionate and take the size of the company into account; they become more likely when a problem arises from using AI without rules or training.
The AI Act is an EU regulation, but it also applies to those outside the Union when they place AI systems on the European market or when the outputs of the systems are used in the EU. Switzerland is following its own regulatory path. In practice: if you are a Swiss company working with European customers, it makes sense to align.
The initial setup takes 3-6 weeks. Then it needs maintaining: every time you add a new AI tool (which happens often), the register must be updated. That's why we offer a light retainer (from €250/month) covering updates, training refreshers and regulatory monitoring.
We tell you where you stand on the AI Act obligations, how much work is needed, and whether it is worth it for your company. Honestly.