Do you have to tell callers they're speaking to an AI?

Ask a supplier whether you have to tell callers an AI is answering and you’ll get a confident answer. Ask them which law says so and the confidence usually goes.

So here is the position, with the primary sources, because this is a question where the secondary sources have quietly diverged from the legislation.

Does UK law require you to disclose an AI on a call?

No. There is no UK statute that requires a business to tell a caller they are speaking to software rather than a person.

That surprises people, largely because the UK has spent two years being described as “about to regulate AI”. What it has actually done is take a sector-by-sector approach rather than pass a single AI act, which leaves phone calls governed by the rules that already applied: data protection, and consumer protection.

The ICO code that keeps getting quoted

You will see it claimed, on vendor sites, in sales decks, and increasingly in RFPs, that the ICO’s AI code of practice requires disclosure, explainability and redress for voice AI, and that it came into force in May 2026.

Two of those three things are wrong.

What came into force on 12 May 2026 is SI 2026/425, which does exactly one thing: it obliges the Information Commissioner to prepare a code of practice on processing personal data in AI and automated decision-making. Its operative provision is a duty on the Commissioner, not on you. It sets no deadline. The code has not been written, so it cannot require anything of anyone yet.

When it lands it will matter, and it will be about personal data rather than disclosure as such.

What about the EU AI Act?

That one is real, and it is now live. The transparency duties in Article 50 became applicable on 2 August 2026. Where they apply, a person interacting directly with an AI system has to be told so, and the penalties are not decorative: up to €15 million or 3% of worldwide annual turnover, whichever is higher.

The question is whether they apply to you. Article 50 bites on AI systems placed on the EU market or whose output is used in the EU. A dental practice in Sheffield answering Sheffield callers is not in scope. A firm taking enquiries from customers in Dublin or Berlin is a different conversation, and one worth having with someone qualified rather than with a blog post.

If you’re weighing this up, the practical test is where your callers are, not where your company is registered.

The rule that already applies to everyone

Set the AI-specific regimes aside and one thing has been true throughout: actively misleading a consumer about who or what they are dealing with is already a problem under UK consumer protection law. Nobody needed an AI act for that.

This is a lower bar than disclosure. Not volunteering that the voice is synthetic is not the same as claiming to be a human called Sarah when asked directly. The first is legal. The second was never a grey area, and it is the one place we do intervene: the platform prompt tells the model never to claim to be a person, and to say so plainly if a caller asks.

So should you disclose anyway?

We think so, and we’ve built accordingly without making the choice for you. The greeting is yours, spoken word for word, and the one we start you with discloses.

The argument isn’t legal, and we should be straight that it isn’t evidential either. We have no data on how UK callers react to being told, and we couldn’t find anyone who does. What follows is reasoning, not a finding.

It rests on the two failures being differently recoverable. Tell a caller in the opening line and the worst case is that some of them would rather have had a person, but they know where they stand before they’ve invested anything, and they can ask for one. Let a caller work it out four minutes in, after explaining the leak and spelling their postcode, and the worst case is someone who feels they were allowed to assume something untrue. Nothing material was misrepresented. It can still land that way.

The first failure you can answer on the call. The second one you hear about afterwards, if at all.

What this means in practice

• UK callers only, and no plans to change: disclosure is your call. We’d still do it, for the reason above, but nothing obliges you.
• Any EU callers: treat Article 50 as applying and get advice. It has been in force since 2 August 2026 and the penalties are structured for large companies, which is not a reason for a small one to test them.
• Either way: don’t let the agent claim to be human when asked. That was already the rule, and it’s the only part of this we enforce for you.
• Watch for the ICO code. When it’s published it will bring real obligations around personal data in AI systems. Anyone telling you today what it requires has not read it, because there is nothing to read.

If you want to hear what a disclosed greeting sounds like before deciding, you can build an agent and test it in your browser without putting it near a phone line. The greeting is the first thing you write, and you’ll know within a call or two whether it costs you anything.