The EU AI Act (Regulation EU 2024/1689) classifies AI systems into 4 risk levels: unacceptable (banned), high (heavy obligations), limited (transparency) and minimal (no obligations). Most of it has applied since 2 August 2026; high-risk obligations start on 2 December 2027. It covers both those who build AI and those who use it.
The dates that matter
The EU AI Act, Regulation (EU) 2024/1689, has been law since 2024 and applies in stages. Most of it has applied since 2 August 2026, but not all of it: the obligations for high-risk systems were postponed by Regulation (EU) 2026/1744, the so-called Digital Omnibus, in force since 27 July 2026.
- 2 February 2025: bans on unacceptable-risk practices and the AI literacy duty.
- 2 August 2025: obligations for providers of general-purpose AI (GPAI) models, national authorities and penalties.
- 2 August 2026: transparency obligations (Article 50) and Commission fines for general-purpose model providers.
- 2 December 2026: two new bans introduced by the Digital Omnibus, and the deadline for marking content generated by systems already on the market.
- 2 December 2027: obligations for high-risk systems in Annex III (recruitment, credit, education and similar).
- 2 August 2028: obligations for high-risk systems embedded in products already regulated (Annex I).
The problem is that most Italian companies, especially SMEs, are still approaching this regulation as if it were for big tech companies. It's not. The Regulation also applies to those who use AI systems in a professional context, not just those who develop them. Anyone using a predictive CRM, a chatbot or HR software with automated functions is already subject to the regulation.
Let's clarify what is already in force, what has applied since August 2026 and what arrives in 2027 and 2028, and what it means concretely for those developing AI software or using it in business.
What's already in force (and many don't know)
The AI Act didn't arrive all at once. The first critical deadline was 2 February 2025, when the absolute ban on AI systems with unacceptable risk kicked in, like social scoring, subliminal manipulation and mass biometric surveillance. The list of banned practices is wider than the examples usually quoted: it also covers emotion recognition in the workplace and untargeted scraping of facial images. The second deadline, 2 August 2025, concerned obligations for general-purpose AI (GPAI) models. These are obligations of the model providers, such as those who build GPT, Claude or Gemini; whoever uses them in production through an API is affected indirectly, because those providers supply the documentation to keep.
There's also an obligation that passed almost unnoticed, in force since 2 February 2025 like the bans: the AI Act requires that all employees using AI systems have an AI literacy level proportionate to their role. It's not an advanced technical course, it's an awareness obligation. Those who haven't done anything on this front yet are already late.
The point is these obligations are not theoretical. They are already active.
The distinction everyone must understand: provider vs deployer
This is perhaps the most important thing to internalise, and also the one most confused about.
The AI Act mainly distinguishes between providers and deployers, that is, those who develop the AI system and place it on the market under their own responsibility, and those who use it. On paper the distinction seems clear, but in concrete application the boundary between the two figures is often hard to identify.
For a software house like us, the distinction is anything but academic. An agency or software house that develops a virtual assistant for a client, using an existing model's API and adding code, interface and branding, in the eyes of the Regulation is no longer a simple AI user: it's a provider, with much heavier obligations than someone using ChatGPT as an operational tool.
In practice: if you integrate an AI model into a product you then deliver to a client, you're probably a provider. If you use AI tools internally to work better, you're a deployer. Beware though: if you customise a model via fine-tuning for a specific use case, you can be considered a provider even if you didn't develop the base model.
And for deployers? Obligations are lighter but concrete: human oversight, staff training, risk management and transparency toward end users. It's not an exemption, it's a different perimeter.
The four risk levels: where do you fit
The AI Act classifies systems into four categories. The risk level determines which obligations apply, and here the difference is substantial.
Systems with unacceptable risk are banned from February 2025: behavioural manipulation, government social scoring, real-time biometrics in public spaces. No adjustment possible, you simply can't do them.
High-risk systems are those used in sensitive sectors: HR selection, credit, education systems, safety components. For these systems the operational deadline is 2 December 2027 (2 August 2028 for those embedded in products already regulated), after the delay set by Regulation (EU) 2026/1744. From then on they need technical documentation, conformity assessment, registration in the European database and post-market monitoring.
Systems with limited risk, like chatbots that must declare they are AI, mainly have transparency obligations. Those with minimal risk, like spam filters or generic content recommendation systems, have no specific obligations.
The practical question every company should ask now: which category does the AI software I develop or use fall into?

Transparency: who has to declare what
Transparency obligations are set out in Article 50, which has applied since 2 August 2026 and splits the work between provider and deployer. If a system interacts directly with people, the provider has to design it so that it is clear there is a machine on the other side. If it generates or manipulates images, audio or video that look real, the output has to be marked in a machine-readable way, and whoever publishes it has to disclose that to users. At the end of July 2026 the European Commission published guidelines on how to apply these obligations, technical marking included. For generative systems already on the market before 2 August 2026, the amending regulation gives providers until 2 December 2026 for the marking.
In practice, for a company: if you run a chatbot on your site, visitors have to understand immediately that they are not writing to a person. If you use synthetic voices or faces in marketing material, you say so. If your software analyses emotions or biometric data, you have a bigger problem than a disclaimer, and it is worth talking it through before development starts.
What applies from August 2026 and what comes next
Since August 2026 most of the AI Act applies; high risk follows from 2 December 2027 (Annex III) and 2 August 2028 (Annex I). Anyone developing or using AI systems in their processes must comply with the obligations that already apply, with sanctions up to 35 million euros or 7% of annual global turnover.
Sanctions follow precise three-tier logic:
- Violating bans (unacceptable-risk systems): up to 35 million euros or 7% of global turnover
- Violating obligations for high-risk systems: up to 15 million or 3% of turnover
- False information to authorities: up to 7.5 million or 1% of turnover
For SMEs, the lower amount always applies. But financial penalties aren't the worst risk. Authorities can order suspension or withdrawal from the market of the non-compliant system, with operational impacts potentially more serious than the fine itself.
High risk slipped to 2027, it did not disappear
Regulation (EU) 2026/1744, in force since 27 July 2026, pushed the heaviest part back. Obligations for standalone high-risk systems, the ones listed in Annex III such as recruitment, credit scoring, education and access to essential services, start on 2 December 2027. Those for systems embedded in products already covered by EU product rules, Annex I, start on 2 August 2028. The same regulation adds two bans that apply from 2 December 2026: systems that generate intimate or sexually explicit images of identifiable people without their consent, and systems that generate child sexual abuse material. Providers of general-purpose models already on the market before 2 August 2025 have until 2 August 2027 to comply.
The delay is not a reason to wait. The time is there for the slow part, which is not writing documentation but working out whether a system is high risk and renegotiating contracts with whoever supplies it. A candidate scoring model becomes, under that classification, a system with requirements on data quality, logging, human oversight and conformity assessment, and that work does not compress into a quarter.
Who supervises, in Italy
The rules on national authorities and penalties have applied since 2 August 2025, together with the obligations of providers of general-purpose AI (GPAI) models. Since 2 August 2026 the European Commission can also fine those providers directly.
Italy has its own AI law, 132/2025, layered on top of the European regulation, and it names the national authorities: AgID as notifying authority for the bodies that will assess conformity, and ACN as market surveillance authority, with inspection and sanctioning powers. For a software developer that changes little in substance and a lot in practice: your counterpart in an inspection is Italian, and ACN is the same agency that already receives NIS2 incident notifications.
What to do now, concretely
No need to panic, but you need to move. Three practical actions we recommend to anyone developing or using software with AI components:
- Map your AI systems in use or in development. Identify what you use internally and what you deliver to clients. For each, ask: which risk category does it fall into? Am I provider or deployer?
- Review contracts with vendors and clients. Many current SaaS contracts don't clarify provider/deployer roles. Before December 2027 is the time to align them, and it pays not to wait for the last quarter. Who is responsible for what, in case of dispute?
- Start an internal AI literacy plan. Not a 40-hour course, but basic awareness of what's used and why. It's already mandatory, not optional.
On the inventory, one practical tip: the right list is not the one IT gives you, it is the one you put together by walking past the desks. Every time we run one, a personal subscription used on company data turns up that nobody had recorded. Training should not only happen but be documented, starting from what can and cannot go into a prompt. And in supplier contracts, besides the roles, you need a clause that guarantees you the information you need to stay compliant yourself.
What changes in the projects we build
On the technical side these duties turn into things you wanted anyway. Chatbot disclosure is one line of interface. Traceability is a log of model calls, with prompt version, input and output kept for as long as they are useful. Human oversight is an approval step before an irreversible action fires, which in our agent projects is already there for operational reasons rather than legal ones.
That is why the place to slot these requirements in is the analysis document, not the acceptance test. How we approach this kind of work is on our AI projects page.
At Redergo we're already integrating these assessments into projects with AI components, from design phase to technical documentation delivered to the client. If you're developing or evaluating a system with AI functions and want to understand where you stand regarding the regulation, tell us about your project.



