Should you disclose to customers when they are talking to AI?

Disclosing that the customer is talking to AI is becoming required in regulated jurisdictions and a customer-trust necessity everywhere else. The plain pattern is to disclose up front, in language the customer will read, before they form opinions about the conversation. Honest disclosure costs nothing once you have it; the cost of skipping it lands later, when customers feel deceived.

A customer starts a chat. The AI replies with "hi, how can I help?" The customer assumes a person, talks for ten minutes, then asks "are you a person?" The AI confirms it is not. The customer is annoyed. The same conversation, with "you're talking to our AI assistant" up front, has none of that residue.

What people in the field are saying

The Telus accent case (covered in a union backlash story in Global News) showed how customers and labour groups react when AI work is hidden. EU AI Act Article 14 requires human oversight for high-risk AI; disclosure is the customer-facing piece of that frame.

What does the law actually require?

It varies. The EU AI Act and similar frameworks require disclosure when the customer is interacting with an AI that affects their rights or significant decisions. California, Colorado, and several other US states have specific bot-disclosure rules for commercial interactions. Some jurisdictions require it for voice channels in particular. The cleanest internal posture is to disclose by default and treat the legal question as the floor, not the ceiling.

What does honest disclosure look like?

One sentence up front. "You're chatting with our AI assistant. Ask any time to reach a person." Plain language. No "virtual concierge" or "intelligent helper" euphemisms. Voice channels should say it once at the start of the call. The customer should not have to ask.

Where does this go wrong?

Soft disclosure: the AI says "I'm here to help" without naming itself. The customer has to detect it from the rhythm of the conversation. Late disclosure: the AI confirms its nature only when the customer asks, several minutes in. Misleading naming: the bot is called "Sarah" or has a stock photo, which crosses from neutral to deceptive. Each is a real pattern in deployments today, and each one carries risk the savings from skipping clear disclosure do not justify.

What if customers prefer to think they're talking to a person?

Some do. Some are also disappointed when they find out later. The empirical pattern is that customers prefer honest framing they can act on (escalate, slow down, repeat themselves) to an unclear interaction they have to figure out. The short-term satisfaction of "they didn't notice" tends to convert to longer-term distrust on a delay.

What is the practical pattern?

Disclose at the start. Make the human escalation visible from sentence one ("ask any time to reach a person"). Log the disclosure in the audit trail. Review the disclosure copy with compliance and treat the result as the default, not the upper bound. None of this is expensive. The cost of getting it wrong is.