Now you have to say it's an AI. Say what it can do, too.

On August 2 the transparency rules in the EU AI Act started applying. Article 50 says a chatbot has to tell people they are talking to an AI, at the latest at the very start of the conversation, unless it is already obvious. The Commission’s own summary is short and blunt about it. There is no transitional relief for systems that shipped years ago, and fines run to 15 million euros or 3% of worldwide turnover.

A news video making the rounds this month closed on the right question: do disclosure labels actually buy public trust, or do they just produce a wave of compliance paperwork while the harder problems wait?

Here is what the evidence says. A bare label costs you something real. In a field experiment with more than 6,200 customers, published in Marketing Science, undisclosed chatbots sold about as well as proficient human reps. Telling customers up front that they were talking to a bot cut purchase rates by more than 79.7% and made the calls shorter. The authors traced the drop to what customers assumed the moment they heard the word: less knowledgeable, less able to actually help.

That assumption is rational. “You are talking to an AI” tells a customer who is on the other end. It says nothing about what that thing is allowed to do about their broken order. So they price in the worst version they have met, which is a bot that walks them through three menus and then puts them in a queue.

The label answers who. The customer is asking what.

Compare two openers.

“Hi, you’re chatting with our AI assistant.”

“Hi, you’re chatting with Celeste, our AI assistant. It can look up your order, change a shipping address, resend a receipt, and start a return inside the return window. For a damaged item, or a refund outside that window, it will bring in Dana from our team.”

Same disclosure. Same compliance. The second one answers the question the customer was actually trying to get answered, and it sets a floor they can hold you to.

Most teams cannot write the second version honestly

Not for lack of knowing their own product. It is that the scope of their bot lives in a paragraph of a prompt. Prompts describe intent. They do not enforce it. If the boundary is instructions, then “it can do these four things and nothing else” is a hope, and the first oddly worded email finds the edge of it.

The sentence gets honest when the boundary is decided outside the model. The model proposes an action. Code checks that action against what this agent is permitted to do, on this topic, for this customer, at this value, and either lets it run or holds it for a person. That check cannot be talked out of anything, because it is not reading the conversation. It is reading the config.

Scope you can read off the settings page

That is how we build. Celeste learns from a business’s own support history and then practices against it privately, drafting real answers to real past tickets where the owner can read them and nobody outside sees a word. Topics go live one at a time. Order status graduates before returns. Returns graduate before refunds. Refunds may never graduate, and that is a legitimate end state, not a failure. You can see the whole shape of it on how it works.

Every write action the agent can take carries a risk classification the owner sets, and the decision to run it or hold it belongs to hard code, not the model. Which produces a small, useful side effect: the scope sentence in your chat window is a readout of your configuration. Add a topic and the sentence gets longer. Leave it alone and the sentence stays short and true.

Disclosure is the floor

Article 50 makes you state one fact. State two. The first is now required and, on the numbers above, quietly expensive. The second is voluntary, hands some of that trust back, and only works if you can stand behind it.

The businesses that come out of this rule looking good are the ones that could already say in one sentence what their agent handles and exactly where it stops. If you cannot write that sentence today, the regulation did not create that problem. It just made it visible.