EU AI Act Explained
Transcript
Made automatically from the video's sound, so it may contain mistakes. Click a line to play from there.
Let's dive right in. Welcome back to our Privacy, Data and AI Law Explainer. Today we're taking on a massive but absolutely essential task: translating the hugely complex EU AI Act into straightforward practical triage for builders, founders, and product teams.
If you're building software today, you're almost certainly touching AI, and you need to know exactly where you stand. This is general information for product teams, not legal advice. And we're anchoring everything strictly to the laws as they stand right now in November 2026.
Last time we mapped out the US privacy picture, but today we're crossing the Atlantic to tackle the European Union's landmark framework. Here is our agenda. We're looking at roles, provider versus deployer, the four risk tiers, general purpose AI rules, the November 2026 timeline, and finally a product team action plan.
Section one, roles. Provider versus deployer. Figuring out your legal identity. The absolute foundation of the Act is Article 3, which explicitly defines what actually counts as an AI system.
It's all about inferential systems versus deterministic code. According to the Commission guidelines, if your feature infers how to generate predictions, content, or recommendations from inputs, rather than just following a rigid set of pre-programmed if-then rules, well, then you're operating an AI system under the Act.
So the crucial thing here is understanding the supply chain split laid out in Regulation EU 2024-1689. Are you a provider or are you a deployer? A provider develops an AI system or model and puts it into the market under their own name or trademark.
A deployer, though, integrates or uses someone else's AI system under their own authority within a business workflow. Your legal obligations are going to look vastly different depending on which side of that line you're on.
Now, here's a massive pitfall for startups. Can a downstream SAS deployer suddenly flip into being a provider? Absolutely. Under the Commission's GPAI guidelines, if you take an upstream API and heavily fine-tune it, or more importantly, change its intended purpose, like taking a general language model and tweaking it specifically to be an HR screening tool, you legally flip into becoming a provider.
And guess what? You've just inherited a much heavier set of legal duties. Section 2. The four risk tiers.
Use case regulation, not a blanket tech ban. Right, so the EU AI Act isn't a blanket ban on tech, but it absolutely bans specific use cases. At the top of the pyramid are the Article 5 prohibited practices.
These are total red lines. We're talking about deploying subliminal techniques to cause harm, social scoring, emotion inference in the workplace or schools, untargeted scraping of facial images, predictive policing profiling, and limits on real-time public biometric ID.
These bans have actually been live since February 2025. And sitting right next to those is Article 4, which makes staff AI literacy mandatory. If your team touches AI, they legally must be trained on it.
High-risk AI, which falls under Article 6 and Annexes 1 and 3. We're looking at standalone domains like hiring, education, credit scoring, critical infrastructure, law enforcement, and biometrics. If you're a provider in these spaces, Articles 8 through 25 hit you with massive conformity hurdles, strict data governance and extensive logging.
If you're merely a deployer, integrating these high-risk tools, Article 26 means you have to maintain human oversight, retain logs for six months, and formally notify your workers if they're subjected to the system. But listen, don't panic.
For the vast majority of you building, standard SaaS features like customer service chatbots or image generators, you fall into the transparency tier, or even minimal risk. But under Article 50, transparency tier features carry immediate live duties as of August 2026.
Got a chatbot? You must clearly disclose to users that they're talking to a machine. Generating synthetic audio, video or text, you've got to use machine-readable watermarking under the Commission Code of Practice, and deepfakes need clear, visible labels.
Section 3. General Purpose AI rules, the foundational model layer. 10 to the power of 25 flops. That is a truly mind-bending amount of computing power, and it's the specific threshold that turns a standard general-purpose AI model into a systemic risk model under the Act.
Articles 51 through 55 outline these GPAI rules, and they've been live since August 2025, under the European AI Office and GPAI Code of Practice. Article 53 requires all model makers, aside from some very narrow open source exemptions, to maintain EU copyright policies, provide detailed downstream documentation, so deployers know what they're building on, and publish a summary of their training data using the mandatory template.
But if a model crosses that massive compute threshold into systemic risk, Article 55 adds strict red teaming, serious incident reporting, and advanced cybersecurity duties. Section 4.
The November 2026 Timeline. Anchoring Your Compliance Dates. If you look at this timeline, you'll see exactly why people get confused. The July 2026 Digital Omnibus on AI, that's regulation EU 2026-1744, delayed standalone Annex 3 high risk enforcement to the 2nd of December 2027, and Annex 1 regulated product enforcement to the 2nd of August 2028.
That was to let harmonised standards finalise. But please, do not fall into the trap of thinking you have nothing to do today. The Article 5 bans and Article 4 literacy rules, the GPAI rules, and crucially the Article 50 transparency labels and Article 57 sandboxes are all actively enforced today, right now, in November 2026.
Section 5. Product Team Action Plan. Triage for Builders. We've covered the complex legal mechanics, so let's translate that into concrete actions for your product roadmap.
It's a simple three-step triage process. First, define your role. Are you a deployer or a provider? Use the official EU compliance checker to be sure.
Second, audit your current features today against the live Article 5 bans. And if you have chatbots or generative AI, get your Article 50 transparency labels sorted immediately. Third, screen your long-term roadmap.
Are you planning a feature that touches an Annex 3 high-risk domain? Start planning for that December 2027 enforcement date right now. Taking these steps is absolutely vital because the Article 99 fines are incredibly steep.
Overseen by the European AI Office and national authorities, who, by the way, are actively coordinating under EDPB statement 3 2024, violating a prohibitive practice could cost up to 35 million euros, or 7% of your global turnover. Even providing incorrect information carries fines up to 7.5 million euros, or 1%.
Yes, proportionality is applied for SMEs, but the financial risk is still massive. While you and your product teams can probably handle the basic Article 50 transparency labels yourselves, we have a very strict rule here. If your feature touches an Annex III domain or borders on an Article 5 prohibition, stop building and call specialist legal conformity counsel immediately.
High risk compliance is just not a DIY project. And that wraps up our EU AI Act triage, but it leaves a massive open question. If your product feeds highly sensitive user data into these heavily regulated language models, who actually holds the liability when the model hallucinates a GDPR breach?
Join us for our next explainer AI, privacy and consent to find out.