Should your AI agent say it is an AI?

A D2C brand in Surat tests an AI agent on order-confirmation calls. A teammate asks whether the agent should say what it is. Someone says callers will hang up. Someone else says it feels wrong not to. Both have a point, and the useful question is how to say it, not whether to hide it.

This post is general information, not legal advice.

What the research shows, and its limits

The best-known study is a field experiment by Luo, Tong, Fang and Qu in Marketing Science (2019). It involved more than 6,200 customers called by chatbots or human workers. Undisclosed chatbots matched proficient human workers. When customers were told beforehand that the caller was a bot, purchase rates fell by more than 79.7%. The authors found that customers who knew they spoke to a machine were curt and saw it as less knowledgeable and empathetic, and that disclosing later in the conversation, or customers' prior AI experience, reduced the penalty (IDEAS/RePEc listing; Temple University).

Three cautions. The study is from 2019, before voice assistants and chatbots became familiar. It tested outbound sales of a financial product, not support or reminders. And we read it through its abstract and secondary summaries, not the full paper.

A newer working paper, "Selling with AI Voice Agents: Does Disclosure Still Matter?" by Habel, Zawadzki and Freise, is listed on SSRN with a randomised field experiment that reports disclosure no longer reduces conversion. We could see only a search summary of it, not the paper, so treat it as a signal that the penalty may have shrunk, not as settled (SSRN).

Honest reading: the effect is unsettled, and probably depends on the call type, the wording and when you say it. A reminder or support call has far less to lose than a cold sales pitch.

What the law says

India

We found no Indian rule that requires an AI caller to announce itself. TRAI's framework governs commercial calls whether placed by a human or a machine (Bar & Bench), and autodialers and robocalls must be declared to the originating access provider (Mondaq). The DPDP Rules, whose notice duties start in May 2027 on the notified schedule, will require clear notice about personal data use (JSA). Check the details in the India rules post.

United States

In February 2024 the FCC ruled that AI-generated voices count as "artificial or prerecorded voice" under the Telephone Consumer Protection Act, so such calls generally need the recipient's prior express consent. The FCC described this as confirming existing law, and it was not a ban (Orrick; Hunton). That is a consent rule, not a disclosure rule.

European Union

Article 50(1) of the EU AI Act requires that people be told they are interacting with an AI system unless that is obvious. Law-firm analyses say it applies from 2 August 2026 and that the 2026 omnibus changes did not delay it (Goodwin; Houthoff). If you call people in the EU, this applies to you. We read these through summaries and did not check the Official Journal.

Never deny being an AI

Whatever you decide about announcing, one rule has no exceptions: if a caller sincerely asks, "Am I talking to a person?", the agent says it is an AI. Denying it is deception, and it is the moment trust breaks. Write this into the agent's instructions. Cora Call agents are instructed never to deny being an AI, and you should test it by asking.

How to word a disclosure

• Say it early, briefly and plainly: "I'm an AI assistant calling for Surat Silks."
• Pair it with a reason the caller cares about: "I can confirm your order in under a minute."
• Offer a person: "If you'd rather speak to someone, I can arrange that."
• Avoid apology and avoid hiding it in the middle of a long greeting.

You can have a recorded notice play before the agent speaks, as Cora Call offers as an option, or you can have the agent say it in its own voice. Test both. The research above suggests timing and framing matter, so measure on your own calls.

Disclosure is a trust decision, not only a legal one

Even where no law requires an announcement, a caller who later discovers they were talking to a machine may feel misled, and a complaint, a bad review or an opt-out follows. Compare the two risks honestly: a lower answer rate on some calls today, against lost trust that is hard to measure and harder to win back.

Calls to existing customers about their own orders or appointments carry the least risk from disclosure, because the relationship already exists and the purpose is useful to them. Cold calls carry the most, which is also where the research shows the largest penalty. If you are unsure, start with the warm calls, learn how your callers react, and widen from there.

Disclosure for different kinds of call

One wording does not fit every call. A few patterns work well, and each is an example to adapt, not a tested script.

• Reminders and confirmations. The caller already knows the business. A single clause is enough: "This is the AI assistant from Surat Silks, calling about your order."
• Inbound reception. The caller chose to ring, so open with the business name and the assistant's role: "Thank you for calling Sai Dental Care, you are speaking to the clinic's AI assistant."
• Outbound enquiries and sales. Here the research penalty is largest, so put the reason for the call into the same sentence as the disclosure and make leaving the call easy.
• Collections-style reminders. Be especially plain. A caller who feels cornered by a machine will complain, so state the purpose and offer a person at once.

Whatever the pattern, keep the agent's name and voice consistent with what it says it is. An agent that calls itself "Asha from the accounts team" and then admits to being an AI two minutes later has made the disclosure look like a confession.

Recording notices

Disclosing the AI and disclosing recording are different statements. If you record, say so, give the reason, and say how to object: "This call is recorded to check quality." Keep the two sentences short and separate. Where recording is optional, treat it as a workspace setting you decide on purpose, not a default.

Where to start

1. Choose a default: announce in the first sentence.
2. Write the "are you a person?" answer into the instructions and test it ten ways, including in Hindi.
3. Split-test two wordings on a small sample, tracking completed calls and complaints, not only sales.
4. Add a recording line only if you record.
5. Review transcripts weekly for callers who sounded confused or upset.

Sources

1. IDEAS/RePEc: Luo, Tong, Fang, Qu, Machines vs. Humans, Marketing Science 38(6), accessed October 2026
2. Temple University Fox School: Chatbots can help improve sales, provided you don't reveal it's a chatbot, accessed October 2026
3. SSRN: Habel, Zawadzki, Freise, Selling with AI Voice Agents: Does Disclosure Still Matter?, accessed October 2026 (full text not opened)
4. Orrick: Did the FCC really ban AI-generated voice calls?, accessed October 2026
5. Hunton Andrews Kurth: FCC declaratory ruling on AI-generated voice calls, accessed October 2026
6. Goodwin: EU AI Act transparency obligations are now in force, accessed October 2026
7. Houthoff: AI Act transparency obligations and the AI Omnibus Regulation, accessed October 2026
8. Bar & Bench: TRAI's crackdown on spam calls and AI-driven telemarketing, accessed October 2026
9. Mondaq: TRAI second amendment to the TCCCPR 2018, accessed October 2026
10. JSA: Prism Infotech, November 2025 (DPDP Rules), accessed October 2026