I thought my AI phone assistant was clean. It wasn't. I only noticed when I went through it line by line, two days after the new transparency obligations had come into force.
Here's the story: for months, an AI assistant has been taking my calls whenever I'm on a building site or in a conversation. It has always dutifully said, “Hey, I'm Kevin's assistant. How can I help you?” And I was convinced that this meant I was doing everything right.
Two things were wrong with it.
First: “Kevin's assistant” — what does a caller think when they hear that? They think of a person in the office. Of a staff member. That is exactly what the new rule is meant to prevent. The word “AI” simply didn't appear in my greeting.
Second, and that was the real shock: there was an AI notice. It hung on a single switch in the configuration. If that wasn't set correctly, the announcement dropped out without replacement off, and nobody would have noticed, because the conversation otherwise carried on completely normally. No error in the log. No warning. Nothing.
So I hadn't forgotten to include a notice. I had included it — as a separate announcement before the greeting, at a point that could be silently switched off. My greetings themselves come in several variants, depending on the time of day and on whether someone has already called before — and the word "AI" did not appear in a single one of them. The disclosure was bolted on, not built in.
If this sounds familiar, keep reading. If not: most people don't know how many entry points their own systems have.
What has been at stake since 2 August
Since 2 August 2026, the transparency obligations of the European AI Regulation have applied. The core is simpler than the fuss suggests: People should be able to tell when they are dealing with an AI system.
Not everywhere. Not on every use. But in four cases — and I deliberately don't write them as legal text, but as they look in day-to-day operations:
1. On your website, someone talks to a chatbot
A visitor types a question into the chat window. If an AI answers there, they need to be able to tell, right from the start. first Contact, not only when they ask.
A practical question for you: does your chat window say "Chat" or "Customer service"? Or does it say that an AI is answering? And: what happens when someone opens the window another way — via a button in the footer, via a link, on mobile? Does the notice appear there too?
2. Your phone accepts calls automatically
That is my case from above. When an AI picks up the phone, the notice belongs at the start of the call — not in the fine print on the website, not only when asked.
And the question that caught me: does your system have several greetings? One for mornings, one for regulars, one as a fallback when something doesn't load? Then the notice has to be in all stand. A caller doesn't land in the variant you tested — they land in whichever one happens to be up next.
3. You use AI voices, AI images or AI videos
This is about systems that generate synthetic content. The legislation requires machine-readable marking – meaning something software can detect, not necessarily a visible label on every image.
If you use a synthetic voice in an advertising video, or show an AI-generated model on your homepage, that's where you should take a closer look. If you had an AI tool touch up a background image: probably not.
4. Emotion recognition or biometric categorisation
Rare with small businesses, but if it happens: affected persons must be informed. Anyone who uses something like this usually knows it.
Important for context: in the first two cases the obligation falls on the provider of the system, and for the last two the Operator. Anyone who builds a third-party chatbot into their website is usually the operator, and so has different obligations than the manufacturer. That is not a detail: it decides who has to do what.
What you can safely ignore
In the weeks before the deadline I read a lot that is simply not true. Let's clear that up before you spend unnecessary effort on it:
- “Every AI text must be labeled.” Wrong. If a human has reviewed a text and someone bears editorial responsibility for it, the disclosure requirement does not apply. This article, for example: I wrote it, I stand behind it, end of story.
- "From now on, strict high-risk requirements apply." No. These deadlines have been postponed: to December 2027 and August 2028 respectively. Anyone selling you high-risk conformity today hasn't read the timeline.
- "Older content will have to be labelled retroactively." No. Anything created before 2 August 2026 stays as it is.
- "If there's AI somewhere in it, I need a notice." No. If the person paying attention would already clearly recognise the use of AI from the circumstances, no additional notice is needed. And tools that only support editing a piece of content do not fall under this in the same way.
There is also a deadline that hardly anyone mentions: for systems that already before were on the market before 2 August, a transition period applies for machine-readable marking of synthetic content until 2 December 2026. For information during direct interaction — i.e. chatbot and phone — this extension applies not.
Five questions you can answer in ten minutes
Take your own company as the example:
- Do people on your website or on the phone speak directly with an AI system?
- If yes, it recognises that in the first sentence, or only when they ask?
- Is there more than one way into this system? And is the notice shown on all of them?
- Do you use synthetic voices, people or videos?
- Could you explain to someone where AI is used in your business, without looking it up?
If question 3 gets you thinking: welcome to the club. That's exactly where I made my mistake.
Why I'm writing about this
Because I've been through it myself, and not just in theory.
Over the last few weeks I've reviewed my own systems: the AI phone assistant, the chatbot on this website, my CRM, and the newsletter pipeline. In most of them I found things I thought were already sorted. The chatbot, luckily, was fine. With the phone assistant, the notice was in the wrong place. In the CRM, there was an explanation that simply claimed the opposite of what the data showed. On the website, an analytics tool had been switched on for weeks yet measured nothing.
These aren't beginner mistakes. They're the mistakes that happen when you've built a system yourself and therefore know exactly how it works should. You then check the path you have in your head — and not the three others.
I am not a lawyer and do not provide legal advice. I am the one who technically takes apart your website, your chatbot and your phone assistant and tells you what he sees.
My offer: 15 minutes, going through your systems together
Together we'll go through what's actually running for you: website, chatbot, phone, content. By the end, you'll know four things:
- What already fits — so you don't touch it.
- What you should change — specific, naming the exact spot where it applies.
- What's not relevant for you at all — that's usually the biggest part, and it saves you the most time.
- What you can implement yourself right away — much of it is one sentence in the right place, not a project.
Takes fifteen minutes of your time. No offer afterwards if you don't want one.
Let's take 15 minutes to look at it
No obligation. Not legal advice — the legal assessment of your individual case belongs with a lawyer. What I give you is a clear technical assessment.
Sources
Everything above comes from official sources, not from summaries of summaries:
- European Commission: timeline of the AI Regulation
- European Commission — Guidelines on transparency obligations (20 July 2026)
- Bundesnetzagentur — transparency obligations under Article 50
Status: 4 August 2026. This article reflects the legal position as it follows from the sources cited, and does not replace legal advice in individual cases.
