Since 2 August, your chatbot may be legally required to admit it is a bot
- Technology
- IT consultancy
- Knutsford
On 2 August a new rule quietly arrived for every chatbot, AI receptionist and automated WhatsApp reply in Europe: the person on the other end has to be told they are talking to a machine. It is Article 50 of the EU AI Act, it applies now, and unlike most of that law it came with no grace period for this particular duty.
If you have added a chat widget to your website this year, or an AI receptionist to your phones, two questions follow. Does an EU law reach a business in Cheshire? And should you care either way?
The answers are "it depends, and anyone who says otherwise is simplifying" and "yes, but not mainly because of the law".
What the rule actually says
The obligation is plainer than most legislation manages. An AI system that interacts directly with people must be built so that those people are informed they are dealing with AI, from the start of the first interaction, clearly, unless it is already obvious.
The European Commission's own guidance sets out who counts: the system has to be a real two-way exchange rather than a form with a submit button, the AI itself has to be doing the talking rather than drafting replies a human sends, and the person on the other end has to be an actual person. A chatbot on a website, an AI answering the phone and an automated WhatsApp assistant all sit squarely inside that description.
Formally, the duty to design the system this way sits with whoever provides it, not with the firm that signs up and pastes the widget onto its site. In practice that distinction will not help you. The disclosure appears in your chat window, under your brand, to your customer, and the setting that controls it is in your configuration screen. Where the legal line between provider and deployer falls in any given arrangement is a question for a lawyer; who the customer blames is not.
Does it reach a firm in Cheshire
The AI Act is an EU regulation, and the UK left. But it reaches businesses outside the EU when the output of the AI system is used inside it, which is less exotic than it sounds: a visitor in Dublin opening the chat widget on your website is arguably enough.
So the honest scope answer splits in two. A trades firm whose chatbot books boiler services for homeowners around Stockport is probably not within reach of an EU market surveillance authority, and nobody has yet demonstrated what enforcement against a UK-only microbusiness would even look like. An e-commerce firm shipping to the EU, or a consultancy with clients in Amsterdam, is a different case, and "probably" is doing real work in both sentences. UK enforcement mechanics are unsettled, and this post is not legal advice; if your exposure matters to you, that is a solicitor conversation.
The fines quoted in the scarier coverage, up to 15 million euros or 3% of worldwide turnover, are real numbers from the regulation. They are also numbers written for the firms that build these systems, and the law explicitly allows proportionality for small businesses. Quote them to a ten-person firm and you are marketing, not informing.
Why you should disclose anyway
Here is the argument that actually decides it, and it has nothing to do with jurisdiction.
A customer who discovers mid-conversation that Sarah was software does not care which side of the Channel the rule applies on. They feel lied to, and they are right.
The disclosure costs one line of text. "You're chatting with our AI assistant. It can book appointments and answer most questions; if you need a person, ask and it will pass you over." Nobody reads that and closes the tab. People deal with machines all day; what they do not forgive is being made to guess.
What loses a customer is not the label. It is a bot that cannot do the job, or one that pretended not to be one. We said in the AI receptionist piece that an answered call is not a handled call; a disclosed bot that hands over cleanly when it is out of its depth beats a disguised one every time it matters.
What a decent disclosure looks like
The pattern is the same everywhere: say it first, say it plainly, and offer the exit.
In a website widget, the opening message carries it. Not the page footer, not the terms link: the first bubble the visitor sees. On WhatsApp, where your automated replies sit inside a personal-feeling thread, the first automated message in any new conversation says it is automated and how to reach a person. On the phone, the AI receptionist's greeting includes it, which the reputable products now do by default: "You've reached Smith's, this is our AI assistant" costs the call nothing.
And the one thing not to do: give the bot a human name, a stock-photograph face and a typing indicator, then admit the truth in a terms page nobody opens. The regulation's test for when disclosure is unnecessary is that the machine-ness is obvious to a reasonably observant person, and the guidance says that exception is to be read narrowly. A designer who has spent effort making the bot feel human has, by their own success, made the disclosure mandatory. If the honest version of your bot would lose the conversation, the problem is not the honesty.
- 01List every place software talks to your customers in your name: website widget, WhatsApp, phone, email autoresponders with AI drafting
- 02Read the first message a customer meets in each one, and check it says what they are talking to
- 03Add the one line where it is missing, and make sure asking for a human actually works
That audit is twenty minutes, and most firms doing it find the gap is not the widget they configured carefully last month but the automation someone switched on in 2024 and forgot.
We build and deploy these systems for clients, so the disclosure line above is not a compliance theory; it is a configuration field we fill in when we set one up. If you are not sure what your own bot says first, or whether the EU question applies to how you sell, ask us and we will look with you.
We work with businesses across Knutsford, Alderley Edge, Wilmslow, Altrincham, Stockport and Warrington, and remotely for clients anywhere in the UK.
If any of this sounds like your business, we will tell you plainly whether we can help.
