AI Act, August 2, 2026: the deadline nobody is explaining to you properly
There are 75 days to go until the EU regulation on artificial intelligence fully applies. A guide written by people who put AI into production every day, not by people who talk about it on panels.

This week Milan is hosting AI Week 2026 at Fiera Milano Rho. 25,000 attendees, 700 speakers, 17 themed stages. Social media is already full of photos of packed halls, the most shareable slides, the event buzz that grows bigger every year.
And that's great, because the noise helps. It helps bring AI into boardrooms, into management committees, into the thinking of people who used to see it as "something for the developers."
But there's one date that, scrolling through the event agenda, we struggle to see mentioned often enough: August 2, 2026.
In 75 days, the AI Act, Regulation (EU) 2024/1689, becomes fully applicable. That is, its most operational part, the one that binds anyone who develops or uses AI systems considered "high-risk." And here, in Italian companies, there's a fairly serious problem that almost nobody is shouting about.
The study you should read instead of yet another article on generative AI
In 2023, before the AI Act had even been formally adopted, the appliedAI Institute for Europe published an analysis of 106 AI systems already in use at European companies. The result:
- 18% could be classified unambiguously as high-risk
- 42% could be classified unambiguously as low-risk
- 40% were in a gray area: it was not possible to determine with certainty which of the two categories they fell into
That 40% is the number every CIO should pin above their desk. It means that, in any European company today, roughly one AI system in every two or three sits in a zone of regulatory ambiguity that will no longer be tolerated from August 2, 2026. Either you know where that system falls, or you have to treat it as if it were high-risk. Period.
Which brings us to the first observation we feel we have to make, looking at how things are going in the field: most Italian companies don't even know how many AI systems they already have in-house. Let alone classify them.
What changes on August 2, 2026 (seriously, no frills)
Regulation (EU) 2024/1689 entered into force on August 1, 2024, with a phased application timeline. The milestones already in effect are:
- February 2, 2025 — ban on AI practices posing an unacceptable risk (social scoring, subliminal manipulation, mass scraping for facial recognition, emotion recognition in the workplace and in education, and others) and the AI literacy obligation for companies
- August 2, 2025 — application of the rules for general-purpose AI (GPAI) models, EU governance, and the penalty regime becoming operational
August 2, 2026 is the date of so-called "full applicability." It's the day when:
- Obligations for high-risk AI systems classified under Article 6 and Annex III of the Regulation become operational.
- The transparency obligations of Article 50 take effect (chatbots disclosed as such, AI content that can be labeled, deepfakes recognizable in certain contexts)
- National supervisory authorities begin exercising concrete inspection and sanctioning powers
In other words: from August 3 onward, if a supervisory authority, an enterprise customer running its own vendor audit, or a labor union worried about your AI résumé screening calls you to account, you need answers that are documented, traceable and open to inspection. "We introduced it because it worked well" no longer cuts it.
Annex III: the list that decides your fate
Annex III of the Regulation lists the areas in which an AI system is automatically considered high-risk. Here it is in the form in which it makes sense to read it — not the way you'll find it in other guides, but translated into the use cases you actually come across in Italian companies:
- Biometric identification of people — facial recognition, biometric categorization, emotion recognition (the latter is outright prohibited in the workplace and in education, except for medical or safety reasons)
- Critical infrastructure — AI systems that manage or support power, water and gas networks, road traffic, transportation
- Education and vocational training — systems that decide admissions, assess students, assign learning paths, monitor exams
- Human resources — résumé screening, performance evaluation, promotion decisions, worker monitoring (this is the area where most Italian SMEs will discover they have a problem)
- Access to essential services — credit scoring, insurance decisions, access to welfare benefits, emergency services dispatch
- Law enforcement, migration, border management
- Justice and democratic processes
An important technical note. Article 6, paragraph 3, provides that a system falling under Annex III is not considered high-risk if it "does not pose a significant risk of harm." Nice. But the burden of proving it, with a documented assessment kept available to the authorities, lies with the provider, before the system is placed on the market. And if the system performs profiling of natural persons, it is always high-risk, no matter what, with no exceptions.
In practical terms: the burden of proof is on you. If you don't document it, it's high-risk. Period.
"But we only use AI, we don't develop it"
This is probably the line we hear most often when the topic comes up in company meetings. It's also the most dangerous one.
The Regulation identifies four categories of parties involved:
- Provider — whoever develops an AI system, or has one developed, and places it on the market under their own brand (yes, even if you had it built entirely by an outside software company)
- Deployer — whoever uses the AI system professionally within their own business (this is where you'll find the 80% of Italian SMEs that think they're not involved)
- Importers and distributors — those who bring systems developed elsewhere to the European market
So: if your company uses an AI system for résumé screening, for scoring customers' credit risk, as an assistant supporting clinical decisions in healthcare, or as an agent that routes customer support cases by priorities that affect customers — you're in. In different capacities, with different obligations for providers and deployers, but you're in.
And yes, this applies even if you're outside the EU: the Regulation has extraterritorial reach, along the lines of the GDPR. If the systems, or even just their outputs, are used in Europe, you have to comply.
The invisible integration trap
This is the part that always holds the nastiest surprises in assessments. Try asking your HR manager whether your company uses "artificial intelligence." They'll almost certainly say no.
Then look under the hood:
- Your HR platform (Workday, SAP SuccessFactors, Oracle HCM, BambooHR, Personio…) has almost certainly switched on AI features in the past year
- Your CRM (Salesforce with Einstein, HubSpot, Dynamics 365…) uses AI for lead scoring, next-best-action suggestions, automatic classification
- Customer support probably has a triage or routing layer based on language models
- Marketing runs dynamic personalization with predictive engines
- Cybersecurity has anomaly detection systems that are, for all intents and purposes, AI
- Finance has some forecasting tool running on top of a model
You didn't develop any of this. But in many of these cases you're a deployer under the Regulation. And your upstream vendors won't tell you anything until you explicitly ask them. In fact, most of them are still figuring out how to structure their own compliance documentation.
Paradoxically, that's exactly the problem: AI came into the company through the back door. A toggle switched on by default by the vendor, a feature that shipped in release notes nobody read, a POC never formally promoted to production, but in production for nine months now.
The obligations, explained the way a colleague would explain them
Let's step outside the Regulation's jargon and translate what it actually means to be compliant as of August 2, 2026.
For those who develop or customize high-risk AI systems (Providers):
- Set up a risk management system that runs throughout the system's entire lifecycle (not a static document, but an ongoing process)
- Demonstrate that the training, validation and testing data are of good quality, representative, and don't produce obvious bias
- Keep structured technical documentation: what the system does, how it does it, what its known limitations are, under what conditions it struggles (Annex IV of the Regulation lists the minimum items)
- Implement an automatic logging system: every relevant operation must leave an inspectable trail
- Ensure transparency toward deployers: whoever uses the system downstream must understand what it does and, above all, what it doesn't do
- Design real human oversight mechanisms (not oversight for show)
- Ensure accuracy, robustness and cybersecurity against measurable standards
- Carry out a conformity assessment before placing the system on the market (for many sectors a self-assessment is enough; for others you need a notified body)
- Maintain post-market monitoring: monitor the system throughout its operational life and report serious incidents to the authorities
For those who use AI systems (Deployers):
- Use the system according to the provider's instructions (sounds trivial; it isn't when you turn on advanced features "because they're there")
- Ensure real human oversight in critical processes
- Ensure input data quality (garbage in, regulatory garbage out)
- Monitor how the system behaves over time
- Inform workers when the system affects them (HR is the most sensitive case in Italy, partly because of the Workers' Statute and union agreements)
- Where applicable, carry out a DPIA (your old friend from the GDPR) and a Fundamental Rights Impact Assessment
Put bluntly: the regulator wants every relevant AI system in your company to have an instruction manual, an identifiable person in charge, a black box that can be inspected, and someone able to explain after the fact what the system did and why.
Penalties: round numbers that speak for themselves
Article 99 sets out a three-tier penalty system:
| Type of violation | Maximum fine (large companies: whichever is higher) |
|---|---|
| Prohibited AI practices (Art. 5) | €35 million or 7% of global annual turnover |
| Non-compliance with other obligations (including high-risk) | €15 million or 3% of global annual turnover |
| Incorrect/incomplete information to authorities | €7.5 million or 1% of global annual turnover |
For SMEs and startups, the reverse principle applies: you take the lower amount, not the higher one. A startup with €2 million in revenue risks up to €60,000 for non-compliance involving high-risk systems (3% of €2 million). For many businesses that are significant but not huge, you can see how sums like these can still be devastating.
And the point is that fines are only the beginning. Under Article 79, the supervisory authority can order the withdrawal of the system from the market, a halt to its use, or corrective action within tight deadlines. If your operations depend on an AI system, a withdrawal order isn't a compliance problem: it's a business continuity problem.
What to do in the next 75 days
We know the temptation to put it off is strong — "let's do it after the summer break" is the most-uttered phrase of all time in Italian companies. But let's put it this way: August 2 falls on a Sunday. August 3 is a Monday. Let's just say the 2026 summer break will be short for anyone who isn't ready.
The bare-minimum sequence, to start now:
- Inventory. A list of every AI system in use or in development in the company. It includes SaaS tools with AI features enabled by default. It includes POCs that have turned into production. It includes internal scripts that use generative model APIs. Everything.
- Classification. For each item: high, limited (transparency obligations only) or minimal risk. If you can't decide with certainty, treat it as high for now and work it out as you go.
- Roles. For each system: are you the provider, the deployer, or both? Who did you get the system from? What have you customized on top of it?
- Gap analysis. Compare the obligations expected by August 2 with what you actually have. Is there technical documentation? Is there logging? Is there risk management? Has AI literacy training been done?
- Governance. Define who is responsible for what: AI Officer, internal committee, an owner for each system. Without this, you won't get anywhere.
- Compliance roadmap. Whatever you can't close by August 2, fit it into a credible plan. The supervisory authority penalizes negligence, not imperfection.
- Training. The AI literacy obligation has been in force since February 2, 2025. If no one in your company has received structured AI training, you're technically already non-compliant.
The mistake we see most often (and wish we'd never see again)
Treating the AI Act as a legal problem.
We say this with all due respect for our colleagues in legal: AI Act compliance isn't written in a law office. It's built into technical processes, architecture choices, how training data is managed, the way human validation workflows are designed. Even the best lawyer can't document an AI system nobody has mapped out for them, and can't do post-market monitoring on a model nobody has set up logging for.
The work is technical and organizational, and it has to be done with people who understand how a model works, how its robustness is measured, and how to build an MLOps pipeline so that compliance is a by-product of the process, not an after-the-fact burden.
Seven years ago, the GDPR taught (those willing to learn) one specific lesson: privacy by design works when privacy is built into the system's architecture. The AI Act asks for the same thing with an AI twist — reliability by design, not at the end of the sprint.
Conclusion: 75 days isn't much, but it isn't zero
In a few days, AI Week will close its halls. The keynotes will end up on YouTube, the slides on LinkedIn, and some people will have made great business contacts. Good: that's the beauty of events.
What's left are the 75 calendar boxes between today and August 2, 2026. Those boxes will decide whether your company — whatever it is, whatever its size — reaches that date with the clarity of someone who knows what they're doing, or in the confusion of someone who finds out in September that something should have been done.
The good news is that the work is doable. It's not a revolution: it's inventory, classification, governance, a few serious technical choices, a bit of training. Nothing exotic for anyone who has already been through GDPR, NIS2 and CE marking.
The bad news is that, once again as with the GDPR, anyone who starts in the last week of July won't make it.
So, have you done your AI inventory yet?
If the answer is "no," or "yes, but I'm not sure it's complete," let's talk. It's the starting point of practically every AI project we've set up since artificial intelligence stopped being a proof of concept and became real production.
Sources and legal references:
- European Commission — "AI Act — EU regulatory framework for artificial intelligence" (updated May 2026) – digital-strategy.ec.europa.eu
- Regulation (EU) 2024/1689 of the European Parliament and of the Council of June 13, 2024 – Official Journal of the European Union, July 12, 2024 – eur-lex.europa.eu
- European Commission AI Act Service Desk — Article 99: Penalties – ai-act-service-desk.ec.europa.eu
- appliedAI Institute for Europe gGmbH – "AI Act: Risk Classification of AI Systems from a Practical Perspective" (white paper, March 2023) – appliedai.de
- EU digital simplification package — "AI Omnibus", political agreement of May 7, 2026 (proposed extension to August 2, 2028 for high-risk systems embedded in regulated products)