Is AI Calling Legal in India? TRAI Rules for Developers

The presales review is going well until the compliance lead asks: is this even allowed? Yes. Automated commercial calling is legal in India. The harder question is what compliance consists of and what to make a vendor prove before you sign, because the CLI header, the brand and the RERA exposure on every call are yours, not the vendor's.

Is AI calling legal for real estate businesses in India?

Yes. No rule in India prohibits a business from using an automated or AI-generated voice to call customers. What is regulated is commercial communication: who may be called, on what basis, at what hours, from which numbers, and with what records. An AI agent must satisfy the same obligations a human telecaller does.

The framework is TRAI's Telecom Commercial Communications Customer Preference Regulations, the TCCCPR: issued 19 July 2018, amended 21 December 2018, and amended again on 12 February 2025, the regime in force today. A draft Third Amendment, published 13 March 2026, remains a draft; more on that below. A presales call is commercial communication, and the developer is the sender the obligations attach to.

One disclosure: IndiaCalling.ai sells AI voice agents to developers, so this checklist is aimed at us too, and is written so a buyer can hold us to it. It is not legal advice.

Does an AI voice agent change what compliance requires?

No. The obligations are the same whether a human or an AI agent makes the call: consent or relationship, preference scrubbing, permitted hours, registered numbers, disclosure and records. What changes is scale. A configuration error a telecaller would make on one call, an AI agent makes on every call, and the network notices.

That is why TRAI is moving, citing major access providers' AI-based detection of UCC as a reason the changes were felt necessary. On why it wants automated calling declared up front, TRAI writes that "it may be technically challenging to differentiate such calling from P2P calling as telecom network may not be able to differentiate between the voice of an actual human interacting on call and a pre-recorded or artificially generated human voice." So the proposed mechanism is declaration, not detection.

What should a developer make an AI calling vendor prove before signing?

Seven things, in writing: advance declaration of automated calling, the provenance of every consent, preference scrubbing before every batch, calling windows enforced by the system, the number series calls originate from, an automated-caller disclosure in the opening line, and who holds the recordings. A vendor that hesitates on any of them is telling you something.

1. Advance declaration to the originating access provider

The draft's proposed rule is that "every Sender shall declare to the Originating Access Provider, in advance, about the use of Application-to-Person (A2P) calls", and a call made without that declaration "shall be treated as unsolicited commercial communication". A good answer names the operator and attaches the declaration. A bad answer is "we handle that on our side" with nothing attached.

2. Consent provenance and storage

Ask where each lead's consent comes from, what form it took, when it was captured and where the record lives. A good answer is exportable fields on the lead record: source portal, enquiry timestamp, consent type, expiry. A bad answer is a spreadsheet with no timestamps and no way to show an operator the basis for one specific call.

3. Preference scrubbing before every campaign, not at import

Preferences change. A buyer who was callable at import may have sent BLOCK 2 to 1909 since, the one-SMS block on the whole real-estate category. A good answer is a scrub against the current preference register before each day's batch, with a log of what was suppressed and why. A bad answer is a scrub run once at onboarding.

4. Calling windows enforced by the system, not a rota

Promotional calls have permitted hours. A human desk enforces them with a roster; an AI agent will dial at midnight unless the system forbids it. A good answer is a hard window in the platform, retries that queue rather than fire, and a dial log showing nothing outside it. A bad answer is "our team makes sure of that."

5. Which number series the calls originate from, and why

The draft refers to the 140xx and 1600xx series as number resources designated by the Authority for commercial communications. Make the vendor say which series your calls originate from, why it is correct for your call type, and show the allocation. A bad answer is a rotating pool of ordinary mobile numbers, which is what dodging blocks looks like.

6. Automated-caller disclosure in the opening line

The buyer should know from the first sentence they are speaking to an automated agent calling on behalf of a named developer. A good answer is the opening script, with a recording that proves it. A bad answer is an agent built to pass as human; when a buyer works it out mid-call, trust in the project goes with it.

7. Recording and transcript custody, retention and deletion

The item vendors, us included, like least. Ask who owns the recording, where and for how long it is stored, who can retrieve it against a lead ID, and what happens at contract end. A good answer: you own it, retrievable by lead within minutes, retention you set, deletion on exit certified. If a vendor cannot hand you your own recordings, the evidence that protects you belongs to someone else.

What happens if a developer's outbound calling gets flagged?

Complaints reach the sender's access provider, and enough of them in a short window trigger action against the sender's telecom resources, not just the number used. Under the current regime the trigger is five complaints from unique recipients in ten days. The draft proposes a stricter trigger, tied to the operators' AI detection.

The draft's explanatory note adds the condition: the trigger would fall from five complaints to three where any CLI allotted to the sender has been flagged as a suspected UCC CLI by the access providers' AI system during those ten days.

The proposed consequences: on a first violation, outgoing services on "all telecom resources including PRI/SIP trunks, SIMs etc." would be barred by all access providers for fifteen days, "irrespective of whether those telecom resources were actually used or not in making such communications". The exposure is the whole telecom estate, including the trunks your site offices run on. A second violation would mean disconnection of all telecom resources for one year and blacklisting.

One sentence on RERA: a misleading claim on a sales call is the developer's exposure, not the vendor's, which argues for scripted, logged, reviewable calls rather than against automation.

Would the proposed 2026 changes affect calling a portal lead?

Potentially, yes. TRAI proposes removing 'inquiry' from the definition of 'Relationship', one of the bases on which a sender may contact someone without explicit consent. A portal enquiry is an inquiry. If notified as drafted, calling a portal lead would rest on explicit consent or an application, not on the enquiry alone.

Today, an inquiry can arguably form a relationship with the sender, the basis most desks assume for portal enquiries. The draft would remove 'inquiry', in TRAI's words "to avoid the nuisance of UCC for the recipients in the name of 'inquiry'", leaving the application limb: products or services applied for by the recipient within the preceding three months.

Portal leads would not become uncallable, but an enquiry alone might no longer carry the call; an application, or explicit consent captured at the portal handoff, would. That is a procurement question: can the vendor's lead record tell the three apart, does consent carry an expiry rather than a one-time checkbox, and can the calling rule change when the regulation does?

Now the status. TRAI issued the draft on 13 March 2026. As of 13 September 2026, exactly six months later, nothing has been notified. TRAI's Regulations index is demonstrably current: its footer reads "Last Updated: 10 September, 2026", and its newest entry is a repealing regulation on television advertising duration released the same day. The latest TCCCPR entry on that index is still the Second Amendment of 12 February 2025. A contract signed today sits under the Second Amendment and should say what happens when the queued changes land.

The network cannot tell your AI agent from your telecaller. The regulator has said so. That is why the proposed rule is declaration, not detection, and why undeclared automated calling would be UCC by default.

How should a developer's presales desk actually run compliant AI calling?

As a configuration, not a policy document. Consent state on every lead record, preference scrubbing before each batch, calling windows the system will not cross, an opening line that discloses the caller, and every recording retrievable against the lead ID. If it is not in the system, it is not enforced.

A representative configuration, illustrative rather than a measured client result: a Hyderabad developer puts an AI voice agent on its daily portal-lead queue, roughly 40 fresh enquiries a day across three portals. Each enquiry carries its source, timestamp and consent state from the portal handoff. Before dialling, the list is scrubbed against the current preference register and suppressed numbers are logged. A 9.40 pm enquiry queues for the next permitted slot. The agent opens every call by saying it is an automated assistant calling for the developer about the buyer's enquiry, and offers a site visit or a human callback. Every call is recorded and indexed against the lead ID.

Then the payoff. Four months later, a buyer says the caller promised a corner unit at the launch price. The presales head pulls the lead, pulls the recording, and listens to the relevant minute before the buyer has finished their coffee. Either the agent said it, and the developer knows what to honour and what to fix, or it did not, and the conversation ends. A desk with no recording has neither option.

Legality was never the question. The question is whether, when an operator or a buyer asks how a call was made, the developer can produce the consent, the scrub log, the number series, the opening line and the recording. Make any vendor, us included, show you that before signing. Book a compliance walkthrough with IndiaCalling.ai and we will run it against your own lead queue.