When you talk about AI Act, in the boardroom one thing usually happens: the discussion shifts to OpenAI, Google, Meta, Anthropic. European AI regulation is perceived as a problem of the giants. Italian SMEs reassure themselves: "we don't develop models, so it doesn't affect us".
That was true in the early phases — entry into force on August 1, 2024, prohibitions and AI literacy from February 2, 2025, obligations on general-purpose models from August 2, 2025. It's different now, because in less than three months obligations on high-risk systems enter force.
The phase 2 of the AI Act — the package of obligations entering force on August 2, 2026 for high-risk systems — shifts the weight of the regulation from producers to users. In technical terms: companies that put an AI system into production for internal use or for their clients become deployers, and enter the perimeter of obligations. Not at producer-level risk, but at a level that requires assessments, documentation, and in some cases notification.
In Italian terms: if you've activated a chatbot, a hiring tool, a credit scoring system, a product recommendation based on AI, or a surveillance/security system using AI, you're in. And you have something to do.
This piece is a pragmatic guide. Not a legal manual — for that you need specialized lawyers, whom you should still consult before closing processes. It's an orientation on what changes, what to evaluate, what to defer, and what to do now.
The four risk categories (in plain language)
The AI Act classifies AI systems into four risk tiers. No frills:
Unacceptable risk: forbidden. Real-time biometric recognition in public spaces, state social scoring, subliminal manipulation. Doesn't concern you, except in edge cases.
High risk: hiring and personnel selection, credit scoring, student evaluation, critical infrastructure management, medical diagnostics, justice or public security applications. Many Italian SMEs are here: HR-tech, fintech, EdTech, MedTech, and all those who use AI to make decisions impacting people.
Limited risk: chatbots, content generation (text, images), deep-fakes. Obligations mainly of transparency: the user must know they're talking to an AI or seeing generated content.
Minimal risk: spam filters, generic product recommendations, automatic translation. No specific obligations, but recommendations of good practices.
Most AI systems in production in Italian SMEs today are in the limited or high tiers. Let's see them separately.
If you're in "limited risk" (chatbots, content generation)
Good news: obligations are light. Bad news: they must still be respected.
Obligation 1 — Transparency to the customer. If you operate a chatbot on the site or a conversational system via WhatsApp/email, you must declare it's an AI. A note at the bottom of the page isn't enough. It must be clear at the start of the interaction. In Italy, the orientation of the Privacy Authority has been clear: an immediate and understandable declaration is required.
In practice: an initial message like "Hi! I'm [Brand]'s virtual assistant. I'll answer you, but if you'd rather talk to a person, write 'human' at any time."
Obligation 2 — Transparency on generated content. If you publish AI-generated content (text, images, video) on public channels — website, blog, social, advertising — you must mark it as such in cases where it can be mistaken for human content in a misleading way. The practical rule: if the content represents opinions, facts, or testimonies, it must be marked. If it's purely illustrative (a background image, a decorative animation), no.
Obligation 3 — Anti-abuse systems. If your chatbot can be used to do dangerous things (prescribe medicines, give legal advice, collect sensitive data), you must have documentable guardrail systems. Nothing esoteric: system prompts preventing certain behaviors, logs of problematic conversations, escalation to human in specific cases.
What you do NOT have to do if you're in this tier: no registration, no advance notification, no mandatory audit. Document for yourself, be ready to demonstrate the choices made if someone asks. End.
If you're in "high risk" (HR, scoring, decisions impacting people)
Here things get serious.
If you use AI systems to:
- select candidates (CV screening, pre-interview automation)
- evaluate clients for credit or insurance
- monitor employees
- make decisions impacting access to essential services
…you're a high-risk deployer. Obligations include:
1. Fundamental Rights Impact Assessment (FRIA). A document — a European template exists — that describes how the system decides, on which data, with what potential bias, with what human review. Must be done before putting the system into production, and updated with every significant modification.
2. Significant human oversight. Decisions cannot be made autonomously by the AI system: there must be a competent person who can review and modify the decision. Attention: an employee who "clicks ok" for five thousand CVs a day is not enough. Required is informed oversight, with access to the reasons for the decision, with real power to change it.
3. Technical documentation. Which model you use, how it was trained (or if it was fine-tuned), which data it saw, which accuracy and bias tests were done. If you use a third-party service (the most common case), the vendor must give you this documentation — and if you don't ask for it or they don't have it, the legal problem is yours.
4. Registration on European database. For some specific cases, registration on a public EU database is required. It's something to verify case by case with the lawyer.
5. Transparency toward impacted user. The person suffering a decision (the non-hired candidate, the customer who received a credit denial) has the right to know AI was involved, and to request an explanation.
If you're thinking "we had no idea all this existed", you're in good company. Most Italian SMEs using AI systems in the "high risk" zone haven't yet understood. Sanctions and full obligations for high-risk systems start from August 2, 2026. But concrete cases — not sampling checks, but lawsuits from injured parties — start already now.
The part disguised as bad news that's also an opportunity
There's a counterintuitive aspect of AI Act phase 2.
All this mandatory documentation — the FRIA, human oversight, transparency to the customer — is exactly what serious companies should be doing anyway, AI Act or not. They're practices that reduce the risk of operational errors, improve customer trust, and (often) improve the quality of the AI itself in production.
Companies equipping themselves for compliance now will be better off commercially in the coming years. Companies postponing it hoping that "nobody will check anyway" will collect problems on two fronts: legal and reputational. A lawsuit for algorithmic discrimination in personnel selection, even if it doesn't lead to direct sanctions, destroys a B2B brand for years.
We've seen it happen to a client of ours. Without naming them, it was a wake-up call that changed the whole team's priorities.
What to do Monday morning
For those who want to start, an operational checklist in five points:
- Inventory of AI systems in production. All of them. Chatbots, marketing automations, scoring, CV screening. Nothing is "too small".
- Risk classification. For each system, which of the 4 tiers is it in? If you're not sure, default to the highest tier (it's safer).
- Ask the vendor for documentation. If you use a third-party AI service, write an email today to their DPO/Legal asking for AI Act technical documentation. Their response tells you a lot: those who are organized respond in 48 hours with a PDF. Those who aren't respond in three weeks with "we're working on it".
- Involve internal (or external) legal. No, don't replace legal with this article. This article gets you to the legal meeting with precise questions — not to avoid it.
- Document the choices. Even if you don't have to yet, start tracking why you activated a certain system, how you tested it, who supervises it. Build a track record.
A final note. On OpificioAI and in the products we're launching, we had to do this work internally in the past months. It was less expensive than we feared, but more expensive than we hoped. And it was, above all, a good opportunity to rethink some product choices that without regulatory pressure we would have left as is. Regulation isn't only a brake: it's also a second reading of the product, done from a different perspective. Sometimes that's exactly what's needed.
Monday, June 29 we close June with a pillar long-form: From SEO to GEO — the new map of organic visibility. Three different engines, three different strategies. How we're reconfiguring client content to stay visible in the era of ChatGPT, Perplexity, and Gemini.
Have AI systems in production and don't know what risk tier they're in? Under artificial intelligence for business we do AI compliance audits in half a day. Let's talk before an angry customer asks us.