The EU AI Act grows teeth on August 2nd. Here's what actually changes for a Belgian dev.
On the 2nd of August the EU AI Act stops being a thing that exists on paper and becomes a thing with fines attached. If you read LinkedIn this week you'd think every developer in Europe needs a compliance officer and a €35 million war chest by Sunday. I'm a backend dev in Belgium who ships the occasional LLM feature and uses Claude Code every day, so I actually went and read the regulation instead of the panic. Most of what applies to someone like me is small. Some of it is nothing. Here's the honest split.
First, untangle three dates everyone smushes together
The Act (Regulation (EU) 2024/1689) has been in force since 1 August 2024. That's not the same as being applicable, which isn't the same as being enforceable with penalties. The rollout is staggered on purpose:
- 2 February 2025 — the banned practices (Article 5) and the "AI literacy" duty (Article 4) kicked in.
- 2 August 2025 — obligations for general-purpose AI models, the governance bodies, and the penalty framework — but the fines for GPAI providers (Article 101) were deliberately held back a year.
- 2 August 2026 — the rest applies generally, and this is the date the teeth come out.
So when a headline screams "€35 million fines start August 2nd," it's wrong. That top tier — €35M or 7% of worldwide turnover — is only for the Article 5 prohibited practices (social scoring, manipulative systems, most live facial recognition), and those have been punishable since February 2025. What genuinely switches on this August is the 3% / €15M tier for general obligations, real enforcement power for the AI Office over model providers, and the Article 50 transparency rules becoming binding.
Does using Claude Code to write code make me liable? No.
Let's kill the one that's been going around developer Slacks. Using an AI coding assistant to write your software makes you, at most, a deployer of an AI system — arguably not even that, since it's a professional tool you use internally. The code it produces isn't a regulated output. There is no documentation duty, no fine, nothing, beyond the Article 4 "AI literacy" obligation, which for a small shop means roughly "understand the tools your people use." No penalty article is even attached to Article 4. If someone tells you Copilot creates AI Act paperwork, they're selling you something.
Shipping an LLM feature: you're a provider, but not that kind
Say I add a "summarise this article" button that calls the Anthropic API. Now I'm a provider of an AI system. The instinct is to panic about training-data summaries and model documentation — the scary Article 53 stuff. But those obligations sit with the people who make the model: Anthropic, OpenAI, Google, Mistral. Calling their API doesn't inherit them. That distinction — model provider vs. system provider vs. deployer — is the whole game, and it's written into Article 3.
What actually lands on me is smaller and sensible:
- Article 50 transparency. If a user is talking to a bot, tell them it's a bot. If your system generates synthetic audio, images, video, or text, it has to be marked as machine-generated. Deepfakes and AI-written text published to inform the public need disclosure. This is the obligation most app builders will actually touch, and it becomes binding on 2 August.
- Article 4 AI literacy. The same low bar as above.
The heavy regime — risk management systems, conformity assessments, logging, human oversight — only attaches if your use case is on the Annex III high-risk list: hiring and CV filtering, credit scoring, biometric identification, access to essential services, education scoring. A "summarise this" button is not high-risk. An automated CV-ranker for a jobs board is. Know which one you're building.
The regulation is aimed at the people training frontier models and the people automating decisions about your life. If you're neither, most of it reads like weather in another country.
The SME footnote that matters
Everyone quotes the "7% of global turnover" ceiling. For SMEs and startups, Article 99(6) quietly flips it: the fine is whichever is lower, the fixed amount or the percentage. For a small company with modest turnover, that's the percentage — a very different number from €35M. There's also Article 62, which pushes member states to give SMEs priority access to regulatory sandboxes and cheaper conformity assessments. The Act is genuinely trying not to crush small players. That part rarely makes the headlines.
The Belgian bit (with a caveat)
Enforcement runs through national authorities, and here Belgium is, let's say, a work in progress. The BIPT (the telecoms regulator) has been named as the lead market-surveillance body, with the FPS Economy coordinating, but the full formal designation wasn't buttoned up by the August 2025 deadline and was still being finalised into 2026. So if you're looking for the Belgian rulebook and can't find a tidy one yet — that's not you, it's the timeline. I'd check BIPT's own pages before relying on anything specific.
What I'm actually doing about it
Nothing dramatic, which is the point. My blog's little AI touches don't hit Annex III, so my whole compliance surface is: label anything obviously AI-generated, and don't pretend a bot is a human. If I ever build something that scores or filters people, that's the day I read Chapter III properly. Until then, the most compliant thing I can do is not confuse a regulation about frontier models and automated decisions with a to-do list for someone who calls an API twice a day.