Foundation · 6 min
What you will learn
- The test for whether a disclosure is needed at all
- What a useful disclosure says, with wording that works and wording that does not
- Why a label is not a defense for a wrong answer
- Which situations call for which level of disclosure
Disclosure is for the customer, not for the badge
The test is whether knowing would change something. Would the customer ask differently, trust the answer differently, check elsewhere, or want to speak to someone? If yes, tell them. If the answer is no, a label is decoration that uses up attention you need elsewhere.
“AI-powered” answers no question anyone has. It does not say whether a person is involved, what the system can see, whether it can act, or what to do when it is wrong. As a badge it is marketing. As a disclosure it is close to useless.
Discovery is worse than disclosure. A customer who works out mid-conversation that they have been talking to a system, after explaining a problem twice, is more annoyed than one who was told at the start. The cost of saying nothing lands later and larger.
What a useful disclosure says
For a customer-facing assistant, four things carry nearly all the value. That it is automated. What information it can see, such as your order history. Whether it can do anything or only advise. And how to reach a person.
Weak: “This chat is AI-powered.” A customer learns nothing they can act on.
Better: “You are chatting with an automated assistant. It can see your recent orders and answer questions about products and delivery. It cannot issue refunds. Type ‘agent’ at any point to reach a person.”
The second version is longer and does more work. It sets expectations, prevents the two most common dead ends, and gives an exit. Plain language beats careful language here, because a disclosure nobody understands has not disclosed anything.
Put it where the decision happens
A disclosure in a policy page nobody opens is a record, not a disclosure. Place it where the customer is deciding something: at the start of a conversation, not after several exchanges; in the message that delivers an automated outcome, not in a help article about how outcomes are made.
Then check that it worked. Ask a few people what they thought they were talking to and what they thought it could do. That takes an afternoon and tends to be more informative than the drafting.
A label is not a defense
Saying content was AI-generated does not make an inaccurate product claim acceptable, and saying a decision was automated does not make an unfair one defensible. Accuracy, consumer protection, privacy, and accessibility obligations apply exactly as they did before.
That matters because disclosure is sometimes reached for as a substitute for fixing the underlying system. If an assistant is wrong often enough that a warning feels necessary, the warning is not the fix.
How much to say, by situation
| What the AI does | Disclosure | Why |
|---|---|---|
| Drafts copy a person edits and approves | Usually none needed | A person is the author. Follow any platform or market rule |
| Answers shoppers in a chat | Say it is automated, up front | The shopper is deciding how much to trust the answer |
| Shows different prices or offers to different people | Say personalization is in use | People notice differences and draw their own conclusions |
| Declines, restricts or decides | Explain enough to challenge it | An outcome with no reason given cannot be disputed |
The last row is the one that carries real weight. An outcome delivered with no reason cannot be disputed, so a customer who believes it is wrong has nowhere to go except a complaint or a chargeback.
Key takeaways
- Disclose when knowing would change what a customer does, expects or can challenge.
- Say what matters: automated, what it can see, whether it can act, how to reach a person.
- “AI-powered” is marketing. It answers no question a customer actually has.
- Place it where the decision happens, and test that people understood it.
- A label never substitutes for accuracy or for fixing the system behind it.