The published material on Indian telemarketing compliance is dense, confident and substantially wrong. Two vendor pages will tell you an AI must identify itself within 15 seconds and within 30 seconds respectively; neither cites anything, because there is nothing to cite.
This page is written from the gazette text — TCCCPR 2018, the Second Amendment of February 2025, the Third Amendment draft, and TRAI's own press releases. Where something is genuinely uncertain, it says so.
Tone (usetone.ai) runs these checks on every call it places, so we had to resolve each of these questions to build the product. What follows is what we found.
The four obligations that actually bind
1. You must be a registered sender
Every commercial caller registers on the DLT platform before placing any outbound call. This is not a formality you can defer — the regulation defines unsolicited commercial communication such that an unregistered sender is in breach by default, regardless of consent, content or recipient.
2. You must call from a designated number series
This is the constraint most likely to stop a specific business, and the one most vendor material glosses. TRAI narrowed it in Press Release 91/2026:
| Series | Purpose | Who may use it |
|---|---|---|
| 140xx | Promotional calls | Entities of any sector. Must register with the telecom service providers. Subject to DND; customers may block by sector. |
| 1600xx | Service and transactional calls | Only BFSI entities regulated by RBI, SEBI, IRDAI or PFRDA, calling their existing customers — plus government, for citizen services. |
Regulation 3(1) closes the back door: commercial communication may take place only using registered headers or numbers from the special series assigned for commercial communication.
The practical consequence: if you are a D2C brand, a SaaS company, a clinic or an edtech, you cannot get a 1600 number. Any pitch built on "your agent calls from a trusted 1600 number" applies to a regulated-BFSI subset and not to you. Ask any vendor making that claim which regulator licenses you.
Regulation 34A, new in 2025, also bars call-management applications such as Truecaller from blanket-tagging or filtering the designated series — which is part of why access to 1600 is guarded.
3. You must pre-declare that you are using an auto-dialer
This one surprises people, because it is usually written about as a future obligation. It is already law. Regulation 4, as substituted by the Second Amendment of 12 February 2025 and in force since roughly April 2025:
Every Sender shall notify the Originating Access Provider, in advance, about the use of Auto Dialer or Robo-Calls as well as the intended objective of such calls in writing.
The Third Amendment draft does not create this duty. It renames it to "A2P calls", broadens the definition to cover artificial voice explicitly, and adds the consequence that an undeclared call is unsolicited commercial communication per se.
Note what that means for scope: this duty attaches to the technology, not to the purpose. A transactional call placed by an auto-dialer is covered. A promotional call placed by a human is not.
4. You must respect the recipient's registered preferences
Both the do-not-call registration and the preference dimensions in Schedule II — sector, day, and time band.
The five things that are wrong
"An AI must disclose itself within 15 seconds." No such requirement exists in TCCCPR 2018, the Second Amendment, or the Third Amendment draft. The widely repeated claim that TRAI has had an AI consultation running for over a year is also fiction — there is no TRAI consultation paper on AI or synthetic voice at all. The only genuine disclosure hook is the synthetically-generated-information duty in the IT Amendment Rules, which is MeitY rather than TRAI, a different instrument, and of contested scope.
To be clear about our own position: Tone plays a spoken AI disclosure on every call anyway. We think it is correct, and it will very likely become a requirement. But it is a product decision, not a legal one, and a vendor who tells you otherwise has not read the instrument.
"Penalties up to ₹10 lakh for AI calling violations." Those figures are financial disincentives on access providers for misreporting, under regulations 27–28, capped at ₹50 lakh per month per licensed service area. They are not sender-side fines for making a call. What a sender actually faces is resource disconnection, plus a ₹5,000-per-resource restoration charge. That is usually worse in practice — losing your numbers stops your business the same afternoon — but it is a different thing, and planning against the wrong consequence leads to the wrong controls.
"The 160 series." It is 1600. This one is just a typo that propagated, but it is a useful tell: a page that says 160 has been copied rather than read.
"All commercial calls are prohibited before 9 AM and after 9 PM." The time band is a customer preference dimension, not a blanket prohibition. Where a customer has registered no preference, the industry Code of Practice default is 08:00–21:00 on all days. So the real obligation is per-recipient, not global. (The Code of Practice is internally inconsistent here — one section says 08:00–21:00 and another says 09:00–21:00. Treat 08:00–21:00 as the outer default and the recipient's own registration as authoritative.)
"AI calling is legal if you have consent, DND and disclosure." Incomplete rather than false. It omits the number-series restriction — the constraint most likely to block a given customer — the regulation 4 pre-declaration, and the cap on how many telemarketers a sender may register.
What this means for how you build
The regulation is structured so that the absence of a check is itself the violation. Unregistered is guilty; undeclared auto-dialing is UCC per se. That makes "we forgot to run that check" indistinguishable from "we chose not to", which is why a compliance layer has to fail closed rather than log a warning and proceed.
Concretely, before a call connects you need to know: the sender is registered and the header is valid; the auto-dialer declaration is on file; the recipient is not on the do-not-call register; the recipient is inside their registered time band; the carrier's own preference register agrees; and, where you are relying on it, that consent exists and has not been revoked.
You also need to be able to prove all six ran, months later, for a call somebody is now disputing. A log line is not proof — it is editable. This is why Tone seals a per-attempt evidence record as a hash chain, where each record commits to the one before it: a partial or altered history stops verifying, which is a property a log file cannot have.
Read the source
Every load-bearing fact on this page comes from a primary source. When a compliance question matters to a decision you are making, read the gazette text — including where it disagrees with this page.
- TRAI, Press Release 91/2026 on the number series
- TCCCPR 2018 and the Second Amendment (12 February 2025), via TRAI's regulations index
- The industry Code of Practice on preferences
This page is a description of regulation as we read it, for engineers deciding how to build. It is not legal advice, and it is not a substitute for counsel who will put their name to an opinion.