How to Tell Customers You Use AI

There is one situation where how to tell customers you use AI is settled for you by law, and a lot of others where it is a judgment call about trust. Mixing those two up is why this question feels harder than it is. Most small businesses worry about the judgment calls and overlook the legal one, which is the only part with a fine attached.

Start by separating them.

The one place it is a legal duty

If customers interact with a chatbot or similar conversational system, they have to be told they are dealing with AI unless it is obvious from context. In the EU this is Article 50(1) of the AI Act, which has applied since 2 August 2026, and the reference text and practical summary set out the requirement: disclosure in a clear and distinguishable manner, at the latest at the time of the first interaction.

Three details worth knowing, because coverage of this tends to compress them.

It follows your customers, not your address. If people in the EU use your chatbot, the obligation reaches you wherever you are based.

The obvious-from-context exception is read narrowly. You cannot assume a customer should have worked it out. A help desk assistant is treated as non-obvious.

And the duty sits primarily with the provider of the system rather than with every business that embeds one. If you added a vendor's chatbot to your site, you are usually the deployer, and your practical job is to confirm the vendor's disclosure is present and has not been switched off in the settings, which it sometimes is by default for a tidier design. If you built your own assistant on a model API, you are the provider and the duty is yours directly.

Fines under the AI Act's transparency provisions run to substantial sums, which is reason enough to check the toggle rather than reason to panic.

The three places it is a trust decision

Everything else is about whether disclosure helps or hurts you, and in practice it usually helps.

Quotes and estimates. If a price came out of an AI tool, customers care less about the tool than about whether a person stands behind the number. The disclosure that matters here is not "we used AI" but "I have checked this and it is my price". Having a wrong figure discovered later is considerably worse, and what to do when AI quotes a customer the wrong price is the scenario to be ready for.

Customer service replies. Email and messaging replies drafted by AI and sent by a person do not need a label. The line most people find reasonable is whether a human read it before it went out. If nobody read it, you are closer to the chatbot case than to the drafting case, whatever channel it arrived on.

Published content. Website copy, newsletters, and social posts drafted with AI are not something customers expect to be labelled, and labelling them tends to draw attention to the tool rather than the work. What damages trust is not AI involvement but content that is visibly unchecked: wrong opening hours, a service you do not offer, a confident claim about a product you have never stocked.

How to tell customers without inviting an argument

The common mistake is phrasing that sounds apologetic, because an apology implies there is something to apologise for and invites the customer to agree. Three patterns that work:

For a chatbot, plain and first: "Hi, I am an automated assistant. I can help with bookings, opening hours, and order status. Ask for a person any time and I will pass you over."

That last clause is doing real work. The most common complaint about business chatbots is not that they are AI, it is that there is no visible exit. Offering the handover immediately removes the thing people actually dislike, and what to do when a customer refuses to deal with AI covers the rest of that conversation.

For a client asking directly: "Yes, I use AI tools in parts of my process. I review everything before it reaches you and I am responsible for the result either way."

Short, not defensive, and it moves the subject to accountability, which is what they were asking about.

For a policy page, if you want one: name what you use AI for, say what a person always checks, and say what you do with conversation data. Keep it to a paragraph. Nobody reads a page, but having one to point at ends a certain kind of email quickly.

The thing to do this week

Two concrete checks, both quick.

1. Open your own chatbot as a customer. Is there a disclosure in the first message, and is there a visible way to reach a person? If either is missing, fix it today.

2. Decide what happens to the transcripts. Customers increasingly ask, and the answer should not be "I do not know". Retention is a decision rather than a default, and how long to keep AI chat logs walks through choosing a period.

The broader point is that disclosure is cheap and discovery is expensive. Customers who are told upfront that they are talking to an assistant overwhelmingly carry on talking to it. Customers who work it out for themselves halfway through a complaint tend to tell other people about it. For where AI fits into a small operation generally, see our guide to AI for small business.

Frequently asked questions

Do I legally have to tell customers my chatbot is AI?

In the EU, yes. Article 50(1) of the AI Act has required disclosure at the first interaction since 2 August 2026 unless it is obvious from context, and the obligation applies if your users are in the EU regardless of where you are based.

Do I need to disclose AI-drafted emails?

Generally no, provided a person reads and sends them. The expectation most customers hold is that a human is accountable for what reaches them, not that no software was involved.

Should I label AI-written website content?

Most businesses do not, and customers do not expect it. What matters more is accuracy, since unchecked content with wrong details does the reputational damage people fear from the label.

What if I use a third-party chatbot on my site?

The legal duty usually sits with the system's provider, but you should verify the disclosure is enabled and visible in your configuration. Some integrations allow it to be turned off for appearance, which leaves you deploying a system that is not disclosing.