Imagine opening a website chat with a question about an order. The assistant has a first name, a friendly photo and a message saying it's checking with the warehouse. Several replies later, you realise nobody has checked anything. The conversation has been generated by AI.
A useful chatbot can answer straightforward questions without pretending to be a member of staff. Customers need to understand what they're using and how to reach someone when the answer needs a person.
If you're adding a chatbot to a small business website, work through the customer journey before choosing its avatar or greeting.
Work out which role your business has
The EU AI Act's direct-interaction rule is a provider obligation. Under Article 50(1), providers must design relevant systems so people know they're interacting with AI, unless that is obvious in context. The information must be clear, distinguishable and accessible, and provided by the first interaction at the latest.
A business using a supplier's existing system may be a deployer. Developing a system, or having one developed and supplying it under your own name, can make you a provider. The Article 3 definitions are the starting point. Branding and contractual arrangements need to be considered together.
Ask your supplier who is responsible for the notice and how your website settings affect it. Make sure an innocent-looking theme change can't remove essential information.
Also check whether the Act applies to your particular UK business activity. The duties aren't determined simply by where the website owner is registered. Even where the EU rule doesn't apply, an honest description helps customers decide how much to rely on the answer.
Say what it is before asking for information
Here is example wording you could adapt:
You're chatting with our AI assistant. It can help with delivery information and finding products. To contact our team, choose “Message a person”.
Use that final sentence only if the option exists. If your alternative is an email form, say so. If staff will reply during office hours, explain that instead of promising an immediate handover.
For a straightforward shop, you might also add a brief reminder not to enter payment card details. The exact wording should fit the information your chatbot actually needs and the safeguards you have in place.
Put the notice where the conversation begins. Test it with a fresh browser session and on a narrow phone screen. Ask someone who wasn't involved in the setup what they think they are talking to. Their answer will tell you more than a design review among people who already know it's AI.
Don't confuse a visible notice with technical marking
The greeting customers read and the machine-readable marks covered by Article 50(2) do different jobs. A visible AI notice doesn't by itself meet a provider's applicable content-marking duty. Technical marks likewise don't explain the interaction to a customer who can't see them.
The Commission's transparency questions and answers confirm that Article 50 generally applies from 2 August 2026. The limited transition to 2 December 2026 concerns Article 50(2) marking for qualifying systems placed on the market before August. It isn't a general extension for chatbot notices.
Get the supplier to explain its approach to each relevant obligation. Keep the answer with your purchase and configuration records. If it says the whole feature is exempt, ask which provision and circumstances it relies on. Read that alongside the Commission's transparency guidance.
Be precise about what the assistant can do
A customer shouldn't need to guess whether a reply has changed their order. Use different wording for different outcomes.
“Here are the steps to request a return” means advice has been provided. “Your return request has been submitted” means an action should really have happened. Test both paths with a fictional order and check the corresponding record in your system.
Our suggested checks are:
- Can it invent a discount, delivery date or refund promise?
- Can it distinguish a general policy from the status of a particular order?
- Does it say when it lacks access to the answer?
- Does the contact route work after a failed reply?
- Can a customer leave the conversation without losing access to support?
Don't write a confident greeting to compensate for a limited system. Narrow its stated purpose to the work it handles reliably.
Keep privacy information and AI disclosure distinct
“This is AI” doesn't explain what happens to a conversation. Customers may enter names, order numbers or details of a problem without thinking of the chat as a form.
Provide a clear route to the relevant privacy information. Check with your data protection lead or adviser that it reflects what the chatbot and its suppliers actually do. The ICO's AI and data protection resources are a starting point for that separate assessment.
For your implementation checklist, record who can read transcripts, what connected records the assistant can access and how long each category of information is retained. Decide who will handle requests and complaints about the conversation. Avoid collecting information just because the supplier's default form includes a field for it.
Check the finished experience after updates
Save the wording, a screenshot and the date you tested it. Repeat the short customer journey after changing the chat provider, widget theme, languages or handover settings.
Keep a few awkward questions in the test, including “Are you a person?” and “Has anyone actually checked my order?”. A helpful answer should accurately explain the service and offer the next available step. That's the standard customers can use, whatever technology powers the chat.