Should you tell customers they are talking to an AI?

Our position is simple: tell people up front that they are speaking to an AI, in the first sentence of the call, the first WhatsApp message and the footer of every email it writes. This article makes the case with sources we opened, sets out what South African law says as far as we could confirm it, and gives you wording to copy.

One caution first. This is general information, not legal advice. Where the law is unsettled we say so, and where we could not open a primary source we say that too.

Does South Africa have a law that says you must disclose AI?

We found no specific rule. The closest thing is a policy that is not in force. On 26 April 2026 the Minister of Communications and Digital Technologies announced the withdrawal of the Draft National AI Policy after its reference list was found to contain fictitious sources, according to the government news service SAnews. The withdrawal was published in Government Notice 3978 of 2026, dated 4 June 2026, and Cabinet approved it on 5 June, according to SAnews.

On 26 May 2026 ITWeb reported that the department aims to take a revised policy to Cabinet by November and release it for public comment by January 2027. We found no later official update, so we cannot tell you where it stands today. Check the department’s website before you rely on any of this. A policy is a statement of direction, not an Act of Parliament.

That does not mean you are free to let an AI pretend to be a person. Two existing laws are worth knowing, and neither was written with chatbots in mind.

What does the Consumer Protection Act say about an AI pretending to be a person?

Section 41 of the Consumer Protection Act covers false, misleading or deceptive representations. Section 41(1), in the text published in the Government Gazette in 2009, says that in relation to the marketing of any goods or services, the supplier must not, by words or conduct, directly or indirectly express or imply a false, misleading or deceptive representation concerning a material fact to a consumer. It also says a supplier must not use exaggeration, innuendo or ambiguity as to a material fact, or fail to disclose a material fact if that failure amounts to a deception, or fail to correct an apparent misapprehension on the part of a consumer that amounts to a false, misleading or deceptive representation. The Act may have been amended since, so check the current text.

Section 41(2) applies the same limits to a person acting on the supplier’s behalf. We read that as covering whatever you put on the phone or WhatsApp for you, but that is our reading.

Here is how we read it for an AI. If an AI is given a human name, a human voice script and a line like "I’m Sarah from the office", a customer may form a false belief about who they are dealing with. If the customer then asks "are you a real person?" and the answer dodges, that looks like a failure to correct an apparent misapprehension. Whether being talked to by an AI is a "material fact", and whether a particular conversation counts as marketing, are questions we could not find a court or the National Consumer Commission answering. Treat this as a risk, not a ruling.

Does POPIA require you to tell people they are talking to an AI?

Not as far as we can read. Section 71 of POPIA is about decisions. It says that, subject to subsection (2), a data subject may not be subject to a decision which results in legal consequences for them, or affects them to a substantial degree, which is based solely on the automated processing of personal information intended to provide a profile of that person.

Subsection (2) sets exceptions, including a decision taken in connection with a contract where the person’s request has been met or appropriate measures protect their interests. Those measures must give the person a chance to make representations and enough information about the logic of the processing. In our reading, an AI that answers the phone and takes a message is not making that kind of decision, while an AI that on its own refused a customer credit on the basis of a profile could be.

So section 71 matters if your agent decides things about people. It does not, on its text, say anything about announcing that a caller is speaking to a machine. For the wider POPIA picture, including what to tell customers about how their information is handled, see POPIA and AI agents. If your AI also records calls, read May a business record phone calls in South Africa?, because the recording notice is a separate duty from the AI notice.

What do other countries require? (context only)

This section is about the European Union, not South Africa, and it does not bind a South African business that serves South African customers. It is useful because it shows where regulators are heading.

Article 50 of the EU AI Act deals with transparency. In the text published by the AI Act site artificialintelligenceact.eu, Article 50(1) requires providers to design AI systems intended to interact directly with people so that those people are informed they are dealing with an AI system, unless that is obvious to a reasonably well-informed person from the context. A law firm’s note from Houthoff says these transparency duties have applied since 2 August 2026, with a transition period to 2 December 2026 for one part, the machine-readable labelling of generated content, for systems already on the market. EUR-Lex, the official text, would not load for us, so we have not checked the article or its dates there. Check it if you serve EU customers.

Do customers react badly when you say it is an AI?

The honest answer is that studies disagree, and the details matter. We opened two.

• In a 2019 field experiment published in Marketing Science by Xueming Luo and colleagues, more than 6,200 customers were randomly given structured outbound sales calls from chatbots or human workers. Disclosing the chatbot’s identity before the conversation reduced purchase rates by more than 79.7%, and the authors say a late disclosure and prior customer experience with AI can mitigate the effect. The abstract we read does not name the country of the sample.
• In a 2024 study in AI & Society by M. J. van der Goot and colleagues, an experiment with 194 participants and interviews with 8 people compared disclosed and undisclosed customer service chatbots. The experiment found no significant effect of disclosure, and interviewees said willingness to help and a friendly tone mattered whether or not the representative was artificial. The abstract we read does not state the country.

We do not read those two as a contradiction. The first is cold outbound selling to people who did not ask to be called, where a caller announcing "I am a bot" gives them an easy reason to hang up. The second is customer service, where the person has a problem and wants it solved. Our agents answer people who contacted the business, which looks more like the second setting. That is our judgement, and neither study measured a South African small business.

What should you actually say? Wording you can copy

The examples below use a fictional business, Example Plumbing. Change the name, keep the structure: say it is an AI, say what it can do, and say how to reach a person.

The email footer only says "checked by the team" if that is true. If the AI sends on its own, do not claim a check that does not happen.

What if the customer asks "am I talking to a robot?"

Answer yes, once, in one sentence, and then keep helping. Do not deflect, joke around it, or change the subject. Section 41(1)(c) is about failing to correct a misapprehension, and a direct question is the clearest case of a customer showing you one.

1. Say yes plainly"Yes, I am the business’s AI assistant, not a person."
2. Say what happens next"I can carry on and take your details, or I can ask someone from the team to call you back."
3. Hand over on requestIf they say they want a person, stop trying to handle the request and pass it on, with the details so far, so they do not have to repeat themselves.

When is it not enough to disclose?

Disclosure does not fix a bad answer. If the AI invents a price or promises a time you cannot keep, telling the customer it was an AI does not undo the misleading statement. That is why anything about money or a booking should wait for the owner’s yes. We describe that as draft mode: the agent drafts and the owner approves before it talks to customers on its own.

When we build an AI receptionist or a support chatbot, the agent says up front that it is an AI. For background on what these agents can and cannot do, see what an AI receptionist is.