TRAI's A2P call rules for AI voice agents: declare the numbers, fix the behaviour

TRAI's Third Amendment to its commercial-communications rules, notified on 18 September 2026, does not ban AI calls in India. By our reading of the notified text, it does not make an agent announce that it is one either. It makes every outbound automated call declarable: you tell your telecom operator, in advance, which number ranges carry them. Separately, operators' machine-learning systems now watch how calling numbers behave.

So the unit of compliance has moved from the call to the number. For teams running AI voice agents, the work is number inventory and campaign discipline, not a new line in the script.

What follows is an operator's reading of the notified text, not legal advice. If voice agents are new to you, start with the complete guide to AI voice agents.

What TRAI notified

The instrument is the Telecom Commercial Communications Customer Preference (Third Amendment) Regulations, 2026. TRAI's press note summarises it, and the full text is on its site. Five parts matter for voice AI.

1. A definition that covers AI agents. An Application-to-Person (A2P) call is a voice call "initiated by an application, software system, or automated platform without direct human dialling", including autodialling, robo-calls and pre-recorded or artificial voice. An outbound AI voice agent fits that squarely.
2. A duty to declare. Every Sender must tell its originating operator in advance that it uses A2P calls, with the calling line identification (CLI) ranges it will use. The operator records this on the DLT platform. An A2P call made without that declaration is treated as unsolicited commercial communication.
3. Detection by the receiving network. Terminating operators must flag CLIs with a high probability of sending spam, based on "behavioural parameters", and share them with the originating operator within two hours. The originating operator then notifies the Sender by SMS or email.
4. An escalation ladder counted per sender. Five or more of one sender's CLIs flagged within ten days triggers know-your-customer (KYC) re-verification. A second instance brings physical verification and, if misuse is confirmed, a fifteen-day bar on outgoing service across all of that sender's telecom resources. After that the text points to one-year disconnection and blacklisting.
5. A termination charge with an exemption. The receiving operator may bill the originating operator a small, capped charge on A2P calls, metered on call duration. Calls from the number series designated for commercial communication are exempt.

TRAI's explanatory memorandum gives the reason in one figure. In April 2026, 15,736 of 20,591 disconnections for spam violations were SIP or PRI connections used for A2P calls, about 75 percent.

What it does not say

Three things are missing from the text, and some coverage has blurred them.

• No ban. Automated and AI calls remain legal. The rule adds a registration step and a monitoring regime.
• No in-call AI disclosure. By our reading, the declaration goes to the operator, not to the person you call. Nothing in the notified text makes the agent say it is an AI. We'd still have it say so, as a matter of trust.
• No published detection criteria. The text says "behavioural parameters" and leaves TRAI free to specify them later. Anyone who claims to know the exact thresholds is guessing.

Declaration is also not consent. A declared call to someone who has registered a preference against it, with no consent or relationship on record, is still unsolicited.

Three dates, and the order matters

The commencement clause staggers the amendment by days after Gazette publication. Counting from 18 September:

• After 30 days (about 18 October): most provisions, including detection, flag-sharing, the escalation ladder and the seven-day inquiry definition.
• After 60 days (about 17 November): the declaration duty and the termination charge.
• After 90 days (about 17 December): a lower complaint threshold. For unregistered senders, three complaints from unique recipients in ten days is enough to suspend the numbers used, when one of the sender's CLIs has also been flagged by detection. Without a flag it stays at five.

Look at the order. Your numbers start being profiled a month before you're required to declare them.

Press coverage has mapped the phases differently, so confirm the dates with your operator.

Why the number is now the thing you manage

Until now, outbound compliance in India lived in two places: the list (consent, preference scrubbing) and the series you dialled from. Under the amendment each CLI also carries a declared purpose and a behaviour record, kept by networks you have no contract with.

That changes one common habit. Diallers have long treated numbers as disposable: when answer rates on one drop, rotate to a fresh one. The escalation ladder counts flagged CLIs per sender, so spreading the same traffic over more numbers raises your exposure. It doesn't dilute it. Five flagged numbers is a worse position than one.

The economics point the same way. Commercial calls were already meant to run on designated series: 140 for promotional calls, 1600 for service and transactional calls from banks, insurers and government, and since August 2026 1601 for other sectors, starting with utilities, courier and logistics.

The amendment exempts those series from the termination charge and bars call-management apps from blanket blocking or spam-tagging them.

For your cost model, the termination charge is a new metered component on the telephony leg, next to the speech, model and voice components of an AI call. It is billed between operators, so expect it to reach you through your carrier contract.

Your exposure is the A2P minutes you place outside the designated series, multiplied by the capped rate in the regulation. On the designated series it is zero. The platform side is on current pricing.

Who is the Sender when a platform places the call?

The duty sits on the Sender: in practice, the business whose communication it is and whose telecom resources carry it. A voice AI platform is the "application" in the definition. It isn't an access provider and usually isn't the number holder.

That is how it works on Callaro's bring-your-own-telephony setup. Phone numbers sit on telephony connections the customer brings, across several supported carriers, and the carrier bills the customer directly. Your carrier relationship is where the declaration lives.

We can show you exactly which numbers carry automated calls. We can't file the declaration for you, and no vendor should imply it can.

If a vendor supplies your numbers as part of a bundle, ask three questions before mid-November:

1. Who is the registered Sender for these numbers, you or us?
2. Have the CLI ranges been declared as A2P, and can we see the record?
3. If a number is flagged, who receives the notice, and how fast do we hear?

What to change in the campaign itself

Seven changes, in the order we'd make them.

1. Make the pool the declaration

List every CLI that carries automated outbound calls and what each is for. In Callaro a campaign dials only from the numbers attached to it: a primary number plus any you add to its pool, all from your own telephony connections. That pool list is your declared range. Don't add an undeclared number mid-campaign because capacity is short.

2. Keep pools small and stable

Callaro picks a number for each attempt with a weighted random draw, and each number's weight follows its live answer rate in that campaign. Between healthy numbers that is a fair optimisation. It must not become a way to outrun a bad list. If a number's answer rate collapses, the list or the calling hour is the problem, and a fresh number inherits it.

3. Let people recognise and return the call

Once a number reaches a contact, Callaro keeps that contact on the same number for later attempts, so retries look consistent on their phone. When the person rings back, inbound routing recognises them as a campaign contact, routes them to the right agent and cancels the queued retry. A number nobody can call back is how a legitimate campaign earns complaints.

4. Cap attempts per person

Retry schedules in Callaro carry a maximum number of attempts per contact and a cap on attempts per day, with gaps that widen and stay inside the calling window. Set both low. The complaint threshold counts unique recipients over ten days, and repeated calls are how a neutral contact becomes a complainant. Our outbound campaigns playbook covers pacing in more depth.

5. Turn a spoken opt-out into a suppression before the next dial

When a campaign call ends with a do-not-call-requested or wrong-number outcome, Callaro adds the number to your Do-Not-Call list automatically and cancels that person's scheduled follow-ups. The list is checked at campaign launch, again at dial time for every contact, and on single calls triggered through the API.

One limit. This is your own suppression list. It doesn't replace the preference and consent scrubbing your operator and telemarketer run on the DLT platform.

6. Put an expiry on inquiry follow-ups

Under the amended definition, a relationship based on an inquiry lasts seven days, and the inquiry must exist in writing or digital form that you can produce. For form-fill leads, call fast and stop on time.

Follow-ups created through Callaro's API can carry an expiry, and the scheduler drops an expired follow-up instead of dialling it late. Keep the form submission itself in your CRM. That record is yours to hold.

7. Watch what your traffic looks like from outside

Nobody outside the operators knows the parameters, but a spam run has a recognisable shape: many very short calls, few answers, the same people dialled again and again. Campaign analytics in Callaro show a call-duration distribution, outcomes by attempt number and a reachability map by hour. Watch the share of calls in the shortest bucket.

How your agent handles voicemail and call screening decides much of that bucket. Make it a deliberate setting for each campaign.

Worked example: the first two hours after a flag

A lender runs payment reminders from eight declared numbers. Its operator sends an SMS: one CLI is suspected of sending unsolicited communication. One flag is a notice, and this is what we'd do with it.

1. Make sure the notice reached a person. The operator writes to the contact on your account. If that is whoever signed the contract two years ago, fix it today.
2. Pull the number. Removing it from the campaign pool stops new contacts landing on it. Contacts it has already reached stay pinned to it for their retries, so if you want it silent, mark the number itself inactive.
3. Read the number's health. The phone-number page shows attempts, answered calls and answer rate for each campaign that used it. Compare it with its siblings. If it looks like the rest, the problem is the campaign, and the other seven are next.
4. Check the last ten days. Count attempts per contact, the share of very short calls, and opt-outs that arrived after a second or third attempt.
5. Fix the cause before the count reaches five. Tighten the attempt caps, narrow the calling window to the hours when people answer, and scrub the list again.
6. Write back to the operator. Say what the traffic was, the consent or relationship basis, and what you changed.

For reminder and recovery programmes, see the collections playbook.

Where this doesn't apply, and where software doesn't help

• Inbound. The definition covers calls an application initiates. An agent answering your support line isn't placing A2P calls, by our reading. The callbacks it schedules are.
• Outside India. Other regulators chose different gates. In the US, the FCC ruled in 2024 that AI-generated voices count as "artificial" under the TCPA, which brings them under its prior-consent rules. Don't carry this playbook across borders unchanged.
• Human click-to-dial. Whether a call a person launches through a dialler counts as "without direct human dialling" is a question for your operator.
• A weak list. With no consent, no inquiry trail and no existing relationship, promotional calls to people who registered a preference against them are not fixed by a better agent or a cleaner number. Callaro doesn't change that, and we'd rather lose that campaign than run it.

What this means if you're buying

The cheapest way to run outbound AI in India has been many ordinary numbers, hard dialling and rotation. That model now has a penalty path attached: a fifteen-day bar on every line you hold, then a year's disconnection.

When you evaluate a platform, ask four things. Who owns the numbers? Can you see each number's answer rate and history? How fast does an opt-out become a suppression? Are attempt caps enforced by the system or left to the prompt?

On Callaro's bring-your-own-telephony setup the answers are: you do, yes, before the next dial, and by the system. The orchestration page has the detail.

Start this week with the inventory. The declaration takes one conversation with your operator. The behaviour record takes months to build, and the networks are already keeping it.