Edge Cases

There is a lot of noise right now about August 2, 2026.

LinkedIn posts shouting fines. Consultants selling €5,000 audits. A real estate agency I follow in Berlin even posted a countdown timer on their homepage. “Only 53 days until the EU AI Act fines you.”

Almost all of it is wrong.

Not slightly wrong. Wrong by years.

I have been reading the actual EU regulation since November. I have built kits for German, French, Italian, Norwegian, and Slovenian small businesses. I spent three months emailing the policy heads of DIHK, Bitkom, CPME, and four other associations to confirm what they tell their members.

Here is what August 2, 2026 actually means. And what it doesn’t.

What changed on May 7, 2026
If you only read one paragraph of this post, read this one.

On May 7, 2026, the European Council and Parliament agreed to the Digital Omnibus. This is an EU technical word for “we simplified the AI Act because too many companies were panicking.”

The Omnibus delayed the high-risk obligations. The ones with the scary fines. The 35 million euro headlines.

Those are no longer August 2026.

They are now:

December 2027 for standalone high-risk systems
August 2028 for embedded high-risk
You can read it yourself in the Consilium press release dated 7 May. It is one Google search away. Most consultants have not updated their slides.

That is your opening.

What August 2, 2026 actually applies to
Two articles. Not the whole AI Act. Two.

Article 4 — AI Literacy
Every business that uses an AI system at work must make sure the people using it have a “sufficient level of AI literacy.”

This is not a certificate. It is not a 40-hour course. It is not Microsoft Copilot training.

In plain words: if your team uses ChatGPT, Claude, Gemini, Midjourney, or any AI tool, you need to be able to show that they understand:

What the tool actually does
What it can get wrong
What data they should not paste into it
Who is responsible when the output goes wrong
For a five-person marketing agency, this can be a one-page policy plus a 30-minute meeting. Documented. Done.

For a 50-person consultancy, it should be a written policy, a short training module, and an annual refresh.

The fine if you ignore it can reach 15 million euros or 3 percent of global turnover, whichever is higher. For an SME the practical risk is lower. But the documentation is cheap, and an irritated competitor or fired employee can file a complaint.

Get the one-pager done.

Article 50 — Transparency
If you provide a chatbot, an AI-generated image, a deepfake, or AI-generated text to the public, you must tell people it is AI.

For chatbots: a small line at the top saying “You are chatting with an AI assistant” is enough.

For images: if the image was generated by AI and could be mistaken for real, you must label it. Marketing material with a synthetic person, AI-generated stock photography on your homepage, anything where someone might think there is a real photographer behind the image.

For audio: same logic. AI voice in a podcast or video advert. Disclose it.

For text: this one is narrower. It applies mostly when AI text is used in public-interest content. A blog post you wrote with AI assistance does not need a label. A press release fully generated by AI about a matter of public interest does.

The wording rules are flexible. The fines are not. Same 15 million / 3 percent ceiling.

What August 2, 2026 does NOT apply to
This is where most of the noise lives. I will keep this short.

It does not apply to:

Your internal AI use of ChatGPT to write emails
Your use of GitHub Copilot to write code
Your use of Midjourney for client work where the client knows
High-risk AI systems (those move to Dec 2027 / Aug 2028)
General-purpose AI model providers (different obligations, different timeline)
Anyone with fewer than ten employees making personal AI experiments
If a consultant tells you that your small accounting firm needs a €5,000 AI Act audit by August because of the fines, they are either misinformed or selling you fear.

Show them the May 7 Digital Omnibus. Watch them backpedal.

What you should actually do before August 2
This is the practical part. If you run a small business in the EU and use AI at work, do these three things.

One. Write a one-page AI Use Policy. List the AI tools you use. List the data your team is not allowed to paste into them (customer data, financial data, legal documents, anything covered by NDA). Sign it as the responsible person.

Two. Hold a 30-minute meeting. Walk the team through the policy. Ask three questions: do you understand what each tool does, do you know what not to paste, do you know who to ask when unsure. Write down the date.

Three. If you have a chatbot or use AI-generated images publicly, add a one-line disclosure. “Chatting with an AI assistant” or “Image generated with AI.” Done.

Total time investment: about three hours.

Total cost: about zero.

The €5,000 audit can wait. By Dec 2027 you will know if your work even falls under high-risk. Until then, the documentation above is enough.

Where this leaves you
The companies who panic now will spend money they did not need to spend. The companies who ignore everything will get caught flat-footed when the news cycle restarts in late July.

The middle path is boring. It is reading the actual text. It is doing the one-page policy. It is putting a disclosure label on a chatbot.

If you want the one-page template, I made one. Free. No email required for the German and French versions. Direct download from my Gumroad.

It is the same template the consultants charge €800 for. I will not. The audience for the paid kits is the businesses that need more than the template, and those buy because the template makes them trust the kit, not the other way around.

That is a story for another post.

Get the free templates:

🇩🇪 DE: Bin ich betroffen? — kostenloser Check
🇫🇷 FR: Suis-je concerné ? — vérification gratuite
Want the full thing?

🇩🇪 KI-Verordnung Kit — €149
🇫🇷 Kit Conformité Loi IA — €149

*Aleš Frelih writes about EU AI Act, forex trading, money, and AI tools from a desk in Norway. This is informational. Talk to a lawyer before relying on it for compliance.*