The EU AI Act
Podcast · The EU AI Act

Updated EU AI Act Compliance Deadlines

by ZumiLabs · 23m
0:00 23m

Transcript

Made automatically from the audio, so it may contain mistakes. Click a line to play from there.

Host 1

You're sitting in this glass walled meeting room right now. It's November 2026, and you are just staring at a deploy button.

Host 2

Yeah, for two shiny new AI features your SAS team just spent months building.

Host 1

Exactly. I mean, the first is this generative AI assistant that faces your customers. And the second is an automated CV ranking tool for recruiters.

Host 2

Which sounds great on paper, but instead of celebrating, everyone is terrified.

Host 1

Right, because clicking that button might trigger some massive find. Half your team is pulling up, you know, these blog posts from 2024.

Host 2

Oh, the 2024 blogs screaming that every single AI app needs immediate high-risk EU certification.

Host 1

Yeah. And then the other half of the room is pointing to conflicting articles claiming, well, the rules have been entirely delayed by recent amendments, so you have absolutely nothing to worry about.

Host 2

So you're frozen. Total regulatory paralysis.

Host 1

Which is the worst place to be when you need answers before you ship. Well, consider this deep dive your regulatory map.

Host 2

Just a quick reminder before we jump in, though. This deep dive provides general information, not legal advice.

Host 1

Right, absolutely. Which it is grounded in primary sources, current as of today, November 2026. If you're building products, you still need your own legal counsel.

Host 2

Definitely. Which is incredibly necessary because the landscape shifted dramatically just a few months ago.

Host 1

It really did. So welcome back to the deep dive. Last time we mapped out the fragmented US privacy picture, navigating the patchwork of state level data laws.

Host 2

But today we are taking you directly across the Atlantic.

Host 1

Yes. Today is exclusively about unpacking the EU AI Act. Specifically regulation 2024 1689, as it was amended by the July 2026 digital omnibus on AI.

Host 2

We're just regulation 2026, 1744 for those keeping track of the exact citations.

Host 1

The goal today is to correct those outdated 2024 timelines. We are going to clarify exactly what is required, of whom, and on what exact statutory timeline as of right now.

Host 2

Because you you really can't afford to navigate European regulations using old Intel.

Host 1

Okay, let's unpack this. Before we get into the heavy compliance burdens, we have to start at the absolute foundation. Like, how does the law actually define what an AI system even is?

Host 2

Right. Because they aren't just regulating, you know, a fancy Excel spreadsheet with a macro in it.

Host 1

I would hope not.

Host 2

The definition is anchored in Article 3, specifically Article 3, paragraph one. To fall under the scope of this regulation, a system has to be machine-based and designed to operate with varying levels of autonomy.

Host 1

Okay. Autonomy makes sense.

Host 2

And this is the linchpin. It must infer from the input it receives how to generate outputs.

Host 1

Like predictions or content.

Host 2

Exactly. Predictions, content, recommendations, or decisions that influence physical or virtual environments.

Host 1

What's fascinating here is that word infer. It feels like it's doing a lot of heavy lifting in that definition.

Host 2

It's doing all the heavy lifting. The mechanism of inference is what separates true AI from deterministic traditional rules-based software.

Host 1

So what's a good example of deterministic software?

Host 2

Think about a standard payroll application. You input an employee's hourly rate. You input the hours worked, and the software applies a hard-coded mathematical formula.

Host 1

Right to determine the paycheck.

Host 2

Exactly. It's entirely deterministic. If you input those same numbers a million times, you get the exact same paycheck a million times.

Host 1

Because the software isn't learning.

Host 2

Right. It isn't inferring anything. It's merely executing the developer's rigid instructions.

Host 1

But if I build like an AI expense management tool where a user just dumps a folder of messy, unstructured receipts and handwritten notes.

Host 2

Oh, yeah. That's completely different.

Host 1

Because the system starts recognizing patterns, right? It categorizes what is a valid business denure versus a personal expense. That system is making a leap.

Host 2

Exactly. It's inferring relationships in the data. The system is operating with a degree of autonomy to generate an output based on pattern recognition.

Host 1

Not explicit rules.

Host 2

Right. And that inference mechanism is the exact trigger that pulls your software under the umbrella of the EU AI Act.

Host 1

So once you realize you've built or are using a system that infers, your immediate next step is figuring out your legal role, right?

Host 2

Entirely. Because the entire law is structured around a supply chain split.

Host 1

Your liability changes depending on where you sit in the ecosystem.

Host 2

Yes. All regulatory obligations flow from whether you are classified as a provider or a deployer. And getting this wrong is a massive legal trap.

Host 1

Let's lay out the boundaries then. Article 3, paragraph 3 defines the provider.

Host 2

Right. This is the entity that develops the AI system, or has it developed for them, and places it on the market under its own name or trademark.

Host 1

Okay. And then Article 3, paragraph 4, defines the deployer as the entity using the AI system under its authority in a professional capacity.

Host 2

Aaron Powell Like the SAS team from our Cold Open. The key differentiator there is the introduction to the market versus the operational use.

Host 1

Is this somewhat like like building a car versus driving a car?

Host 2

Aaron Powell That's a great way to look at it.

Host 1

Like the provider is the manufacturer of the semi-truck, right? They design the engine, they conduct the crash tests, they ensure it meets safety standards before it leaves the factory.

Host 2

They bear the burden of technical conformity.

Host 1

Right. But the deployer is the logistics company that just buys a fleet of those trucks.

Host 2

Exactly. They aren't building engines, but they are responsible for deciding which routes the trucks drive, ensuring the drivers are awake and have licenses.

Host 1

So they manage their real-world operational safety.

Host 2

Yes. That analogy holds up perfectly. If we connect this to the bigger picture, this split ensures that the regulatory burden falls on the entity with the most technical control.

Host 1

The provider.

Host 2

Right. Well, the operational burden falls on the entity with the contextual control, the deployer.

Host 1

But wait, what happens if that trucking company decides they want more horsepower? So they pull the engine out of the truck, completely rebuild it with custom parts, and paint their own logo on the grill.

Host 2

Well, at that point, they fundamentally altered the safety profile of the vehicle, right? And claimed it as their own.

Host 1

So they aren't just a deployer anymore.

Host 2

Exactly. In the eyes of the AI Act, they've crossed a very dangerous threshold into becoming a provider.

Host 1

This is that flipping concept, right?

Host 2

Right.

Host 1

We'll dig into the mechanics of that later, but knowing your role is really only step one.

Host 2

Step two is knowing your risk level. Because the EU AI Act is not a blanket ban on technology.

Host 1

Right. It's structured as a risk pyramid. So let's start at the baseline. Minimal risk.

Host 2

The vast majority of technology sits here at the very bottom. Think of basic AI spam filters or those smart MPCs and video games.

Host 1

The Act largely leaves those alone, right?

Host 2

Yeah, entirely. Yeah. There are no heavy conformity assessments. The European Union actively wants to foster this kind of low-risk, everyday innovation.

Host 1

But then the law skips straight to the absolute top of the pyramid. The red lines. Article 5 outlines the prohibited practices.

Host 2

Aaron Powell And it is crucial to note that these bans have been live and strictly enforced since February 2nd, 2025.

Host 1

So if you're building these today, you are actively breaking the law.

Host 2

There is no grace period remaining. Dala doesn't care how accurate your code is. If your use case violates fundamental EU rights of human dignity, it's banned.

Host 1

Let's list some of these out. We're talking about subliminal manipulative harm, right?

Host 2

Yes. Exploiting vulnerabilities based on age or disability is another.

Host 1

And social scoring, evaluating people to generate a universal trust score that leads to detrimental treatment.

Host 2

It's the literal prevention of a digital caste system. Also, untargeted facial scraping from the internet or CCTV footage is completely prohibited.

Host 1

So you can't just vacuum up faces to build a database. What about in the workplace? I know there was HR Tech claiming to analyze employee engagement.

Host 2

That is a hard band. Article 5 prohibits emotion inference in workplaces and educational institutions.

Host 1

So you can't use software to analyze microexpressions on a Zoom call to see if someone is paying attention.

Host 2

Exactly. Your internal emotional state is protected privacy. And of course, predictive policing profiling is also banned.

Host 1

Like predicting someone will commit a crime based on personality traits. The justice system can't outsource the presumption of guilt to a statistical model.

Host 2

Now there is a brand new addition to this list of prohibitions, thanks to the July 2026 digital omnibus on AI.

Host 1

Right. Regulation 2026 slash 1744 added a highly critical new prohibition to Article 5. It's now explicitly illegal to generate non-consensual sexually explicit content.

Host 2

Yes. This covers CSAM, deep fake pornography, and nudification apps. And this specific ban enforcement goes live in just a few days.

Host 1

And December 2nd, 2026. The technology was moving so fast they had to implement an immediate block.

Host 2

Exactly. Now, alongside those bans, there's another requirement that went live back in February 2025. Article 4, AI literacy.

Host 1

Okay. So the law states that providers and deployers must take measures to ensure their staff has sufficient AI literacy.

Host 2

Right. Anyone dealing with the operation or use of AI systems on the company's behalf needs training.

Host 1

Wait, does this mean every company needs a certified AI officer like the GDPR requires a data protection officer?

Host 2

No, no. You can breathe a sigh of relief there. There is no mandated governance structure or certification required here.

Host 1

So I don't need to hire an expensive chief AI officer, but what does compliance actually look like? Can I just send a PDF instruction manual on Slack?

Host 2

A PDF link will not survive regulatory audit. You must contextually train staff on the risks of the actual tools they use.

Host 1

What does that curriculum look like practically?

Host 2

Well, if your marketing team uses a large language model to draft copy, their training must focus on hallucinations.

Host 1

Because generative AI doesn't retrieve facts, it predicts words.

Host 2

Exactly. It will confidently invent false claims about your product. They need to be trained to fact check.

Host 1

And if engineers are using AI coding assistance.

Host 2

The literacy must match the operational risk.

Host 1

That makes perfect sense. So bands and literacy are live. Let's shift to the foundational layer. General purpose AI or GPAI.

Host 2

Right. Articles 51 through 55. And these rules have been live for over a year since August 2nd, 2025.

Host 1

When we say GPAI, we're talking about the massive foundation models, right? Like OpenAI or anthropic.

Host 2

Yes. The providers of these models face baseline duties under Article 53. They need extensive technical documentation and strict copyright policies.

Host 1

But the big one is Article 53, paragraph 1. Right? The public training data summary.

Host 2

That requirement completely altered the secrecy of model training. They are legally required to make a detailed summary of the content used for training publicly available.

Host 1

So publishers and news organizations can see if their copyrighted material was swept up?

Host 2

Exactly. Now what about the open source community? The Act does provide a narrow exemption for free and open source models as long as parameters are publicly available.

Host 1

But that exemption vanishes if the model poses a systemic risk, correct?

Host 2

Yes. Systemic risk under Articles 51 and 55. And they chose a strictly mathematical threshold for this.

Host 1

Right. 10 to the power of 25 flops, floating point operations. If you train a model with that massive compute power, you cross the line.

Host 2

Because at that scale of compute, the model exhibits emergent unpredictable capabilities. That unpredictability is the systemic risk.

Host 1

So if a model hits that 10 to the 25 threshold, what are they required to do?

Host 2

They have to conduct mandatory red teaming. They hire independent experts to aggressively try and break the model's safety guardrails before it's deployed.

Host 1

Like trying to get it to write malware or output dangerous instructions.

Host 2

Exactly. They must also report serious incidents directly to the commission and conduct heavy cybersecurity testing.

Host 1

Okay, so that's the foundational layer. Here's where it gets really interesting. What about the everyday digital products? The rules that just went live a few months ago.

Host 2

You're talking about Article 50, transparency risk. These rules went live on August 2nd, 2026.

Host 1

Which is hyper relevant right now. The first part of Article 50 is chatbot disclosure, right? Users must know they are interacting with a machine.

Host 2

Yes. Unless it's blindingly obvious. Your UI must clearly state it's an AI. You cannot trick a vulnerable customer into believing they're getting human empathy.

Host 1

Which might hurt engagement, but the law prioritizes trust. The second part is about synthetic media.

Host 2

If your system generates audio, image, video, or text, the outputs must be marked in a machine readable format.

Host 1

Machine readable watermarking. So I can't just put a visible logo in the corner.

Host 2

No, because anyone can crop that out. It has to be interoperable. The industry standard is C2PA, which uses cryptographic hashing embedded in the file's metadata.

Host 1

And the third part is clear labeling for deep fakes, right? If you generate manipulated content that falsely appears authentic, you have to disclose it.

Host 2

Right, with very narrow exceptions for obvious satire.

Host 1

Let's bring this back to our cold open, the SAS team shipping the gen AI assistant. They fall squarely under Article 50.

Host 2

They absolutely do. Which means they must ship those disclosures right now. If the UI doesn't disclose the AI and the images don't have C2PA marks, they cannot launch.

Host 1

Okay. But what if they do add the disclosures, they train their staff and they avoid the banned practices? Are they in the clear for their other feature?

Host 2

The automated C V ranking tool for recruiters?

Host 1

Not at all. Because if your AI affects human livelihoods, you enter the high-risk zone.

Host 2

Right.

Host 1

And that brings us to the high-risk horizon, governed by Article 6 and its annexes. How do they define high-risk AI?

Host 2

There are two types. The first is Annex I. These are AI systems used as safety components in products that are already heavily regulated.

Host 1

Like medical devices, toys, aviation, machinery.

Host 2

Tangible hardware.

Host 1

Tangible hardware.

Host 2

Right. But the second category, Annex III, is the critical list for software teams. These are standalone domains.

Host 1

Let's list some of these Annex III domains. Biometrics is one, right?

Host 2

Yes. Also, critical infrastructure management. An education like AI used for student grading or determining admissions.

Host 1

Because an algorithmic mistake there alters a student's entire life trajectory. What else?

Host 2

Credit and insurance evaluation. And crucially for our cold open team, employment and CV screening tools.

Host 1

AI used for the recruitment or selection of natural persons. So the C V ranking tool is definitively an Annex 3 high-risk system.

Host 2

Without a doubt.

Host 1

Which is why the team was panicking. This raises an important question though. What did the July 2026 digital omnibus do to the timeline for high-risk systems?

Host 2

It provided massive relief. Regulation 2026-1744 physically shifted the enforcement dates for high-risk systems.

Host 1

What are the exact dates as of today, November 2026?

Host 2

For standalone Annex 3 high-risk systems, the new deadline is December 2nd, 2027.

Host 1

Okay, so they have just over a year. And what about the physical regulated products under Annex 1?

Host 2

Those were pushed back to August 2nd, 2028.

Host 1

Does this delay mean tech teams can just ignore Annex 3 for another year? Can they launch the C V tool today and worry about compliance next Thanksgiving?

Host 2

Absolutely not. The compliance burden is massive and takes months, if not a full year, to build into your engineering architecture.

Host 1

Let's break down that burden. We have provider conformity under Articles 8 through 25 versus deployer duties under Article 26. What does the provider have to do?

Host 2

They need a continuous risk management system. They have to write extensive technical documentation. They must design the software to automatically generate forensic logs.

Host 1

And what about Article 10? Data governance. You can't just train on biased, scraped data, right?

Host 2

No. Article 10 requires ensuring your training data is relevant, representative, and to the best extent possible, free of errors.

Host 1

Meaning you have to mathematically prove you mitigated historical biases in your hiring data set.

Host 2

Exactly. And then you must complete a conformity assessment, register in the EU database, and affix a digital CE marking.

Host 1

That is a staggering amount of work for the provider. But what about the deployer, the HR agency that just buys the software? Are they off the hook?

Host 2

Deployers are governed by Article 26, and they are definitely not off the hook. First, they must control input data and follow the provider's instructions.

Host 1

They also have to notify workers if AI is used in the workplace to evaluate them, right?

Host 2

Yes. Transparency is key. And critically, their IT department must retain the automatically generated logs for at least six months.

Host 1

To allow for forensic investigation if someone claims discrimination. And they also have to ensure, quote, competent human oversight.

Host 2

This is vital. The law explicitly rejects the idea of a rubber stamp.

Host 1

So a recruiter can't just blindly click approve on the AI's C V ranking 500 times a day.

Host 2

No. The human must have the competence, training, and the actual authority to disregard or override the AI. If the UI doesn't force meaningful review, it's noncompliant.

Host 1

Let's talk about the stick. The penalty framework under Article 99. These fines have been live since August 2025.

Host 2

And they are designed to be GDPR level deterrence. For prohibited practice violations, you're looking at up to 35 million euros.

Host 1

Or 7% of global turnover, whichever is higher. That's existential.

Host 2

Yes. For high-risk violations, it's up to 15 million euros, or 3% of global turnover.

Host 1

Though there are proportional caps for SMEs, so it doesn't instantly bankrupt a small startup for a technical error.

Host 2

Right. But it's still a devastating financial blow.

Host 1

Okay, so we've summarized the timeline.

Host 2

But in the real world of SAS APIs and open source, the line between provider and deployer is dangerously blurry.

Host 1

It's extremely blurry.

Host 2

Which brings us to the gray areas and common traps.

Host 1

So what does this all mean for a startup building on, say, an open AI or anthropic API? They take an upstream GPAI API, use it to build a C V screening tool, and put their own brand on it.

Host 2

That is the flipping trap. According to the Commission GPAI guidelines, you are no longer just a deployer.

Host 1

Wait, really? Just because I branded it?

Host 2

Because you placed a high-risk system on the market under your own trademark with the specific intended purpose. You legally flip into a high-risk provider.

Host 1

So I suddenly inherit all those article eight through twenty five conformity burdens. The CE marking, the article ten data governance.

Host 2

You inherit all of it.

Host 1

Okay.

Host 2

Because the API creator didn't build an HR tool. You did. You can't outsource your liability.

Host 1

Oh wow. And what if I substantially fine-tune an open source base model?

Host 2

If you materially change its core capabilities, you might flip into becoming a GPAI provider yourself, triggering the Article 53 documentation rules.

Host 1

That's a massive trap. Now, what about Article 6, paragraph 3? I was reading that and it looks like a narrow exception for systems that fall into high risk categories.

Host 2

It is an exception, but it's very narrow. It says an annex three system isn't high risk if it doesn't pose a significant risk of harm.

Host 1

Like if it just performs a narrow procedural task, like using OCR to convert a messy PDF resume into plain text without ranking it.

Host 2

Exactly. It's preparatory. It does zero evaluating.

Host 1

Is this basically like a hall pass from the compliance principle?

Host 2

Sort of. Or like a zoning variance. But here's the catch. To get the hall pass, you still have to document your assessment that it's not high risk.

Host 1

So there's no silent opt-out.

Host 2

None. You have to actively document it before going to market. And you still have to register the system in the official EU database.

Host 1

And if you're doing any kind of profiling, the exception vanishes entirely.

Host 2

Immediately. You're back to being high risk.

Host 1

Okay. So we know the rules, but who is actually enforcing all this? Who is knocking on doors?

Host 2

It depends. The European AI office handles the giant GPAI model centrally.

Host 1

But what about our everyday SAS deployers?

Host 2

For them, we look at EDPB statement 3/2024. Local national data protection authorities, the DPAs, will enforce the law locally.

Host 1

Wait. The exact same privacy regulators who audit my cookie banners under GDPR?

Host 2

The exact same ones. They already have the investigative infrastructure. It creates a complex web of market surveillance coordination, but your local DPA is your primary auditor.

Host 1

That is crucial to know. All right, let's bring it all together for the listeners. Founders, PMs, engineers, what are the practical product takeaways for today, November 2026?

Host 2

We have three immediate marching orders.

Host 1

Number one, document your article for AI literacy today. Right. Set up internal contextual training for your team on the risks of the tools they use. Number two, ship your article 50 disclosures immediately.

Host 2

If you have a chat bot or AI image generator live, ensure the UI disclosures and machine readable C2PA marks are pushed to production now.

Host 1

Because those rules are live.

Host 2

And number three, screen your roadmap against Annex 3 immediately.

Host 1

Use the official EU compliance checker or look into the Article 57 regulatory sandboxes to see if your 2027 roadmap hits high risk triggers.

Host 2

You need to know your exposure well before the December 2027 deadline hits.

Host 1

Which brings up the big question. When is it time to stop reading blogs and call specialist legal counsel?

Host 2

The moment you realize you are building an Annex 3 tool, like our C V tracker, or heavily fine-tuning an open source model.

Host 1

Because you've crossed a red line.

Host 2

Exactly. The engineering requirements are too complex to guess at.

Host 1

Exactly. Join us for episode six, AI privacy and Consent, where we will collide the AI Act with GDPR Article 22, Automated decision Making.

Host 2

To leave you with a final thought to mull over, the AI Act regulates the intended purpose of your system.

Host 1

Right. The provider defines the purpose.

Host 2

But in a world of general purpose AI, where models are infinitely adaptable by the end user through prompting.

Host 1

How long until a minimal risk text generator is unexpectedly used by a customer in a way that creates high-risk harm.

Host 2

Precisely. The law has drawn a statutory line in the sand based on predictability. But the technology is a rapidly rising tide.

Host 1

It definitely is. Thanks for listening to this deep dive. We'll see you next time.

Save to playlist

Manage playlists →

Share