Is AI Calling Legal in India? Yes — TRAI + DLT

No vendor can. Every duty in the TCCCPR framework lands on the Sender — the business whose calls they are. Regulation 4 says "Every Sender shall notify the Originating Access Provider, in advance". Substituted clause (bw) turns an unregistered Sender's commercial communication into an Unsolicited Commercial Communication. The consequences in para 71 attach to the Sender's own telecom resources, not to the platform's. (All three from the TCCCPR (Second Amendment) Regulations 2025, checked 20 Jul 2026.) A vendor can make DLT registration, template hygiene and script control easier, and a good one will. None of that moves the duty.

So the useful question is not "are you TRAI compliant" — it is "under which regulation number", followed by what you can actually see about a vendor without booking a sales call. Here is what the platforms in this market publish about themselves.

Every figure above was read on the vendor's own published pages on 21 Jul 2026 and is quoted from their own wording rather than estimated by us. Scalify Labs is the exception on sourcing: it has no live pricing page, so its figures come from its own services page and its own pricing blog post. Three of them need their caveats carried with them or the row is misleading:

• Agni advertises "all-in from ₹2/min", but its own rate card shows "upto ₹2/min" only on the quote-only Enterprise tier at 10,000+ minutes. The buyable entry is ₹2,999/mo with ₹8/min overage.
• Scalify Labs heads its page with ₹0.40/min, inside a table its own page labels "Rates are estimates based on publicly available information and direct conversations. Always get a formal quote" — and voice minutes are billed on top of the monthly plan, against a stated minimum engagement of 10,000 minutes a month and a 3-month initial commitment.
• Our own row needs one too. There is no subscription and no minimum volume, but the minimum first credit load is ₹10,000, the phone number recharge is ₹2,000 per 30 days debited from credits, and for founding clients the Care Plan runs at ₹5,000/month for the first three months (list ₹12,000) in return for the ₹40,000 setup being free. That is ₹7,000 of month-one fixed cost before a single minute. All prices exclude GST; 18% applies when you load credits, so a ₹10,000 load is billed ₹11,800. Full breakdown on /pricing.

Rates in this market moved within weeks during collection, so treat 21 Jul 2026 as a hard check date and re-verify before you rely on any of it. The maintained board lives at /research/india-ai-voice-agent-pricing-index.

None of this is a compliance answer. It is a due-diligence one: the DLT registration, the Regulation 4 intimation and the consent record are yours whichever platform you pick, so the thing worth comparing is how much a vendor will put in writing before you commit — the price included.

If you read nothing else, read this table.

Everything below is the working out.

TRAI issued a clarification on the 140 and 1600 number framework on 10 Jul 2026 (SCC Online Blog, "TRAI clarifies 1600 and 140 series number framework", checked 20 Jul 2026). Two things in it matter enormously to a small business, and neither appears in the vendor articles currently ranking for this query.

First, the 1600 series has a stated eligibility limit. Per that clarification, the 1600 series is for service and transactional calls made by regulated entities in the BFSI sector, and for government-to-citizen communications by government entities — the BFSI regulators named being RBI, SEBI, IRDAI and PFRDA. If you run a dental practice, a car showroom, a gym or a coaching centre, you are not a BFSI regulated entity and you are not a government body. The advice "just get a 1600 number for your service calls" was reasonable in 2025 and is not obviously available to you now.

Second, caller-ID treatment is now spelled out. Per the same clarification, calls originating from the 1600 series cannot be tagged, blocked or filtered. Calls on 140 cannot be tagged or filtered either, except where blocking is carried out in accordance with DND preferences registered by customers. That is a meaningful commercial fact: a compliant 140 call will not be labelled "Spam" by a caller-ID app, which is the exact fear most owners have about switching away from a normal mobile number.

A correction while we are here. One of the highest-ranking pages on this topic, caller.digital's "AI Cold Calling in India 2026" (checked 20 Jul 2026), uses the phrase "the 140/160-series" in a heading and in the body. TRAI's own text and the Jul 2026 clarification both say 1600, including the form "1600xx". It is a one-character difference that will send you to your operator asking for a series that does not exist.

The honest gap this leaves

So where does a non-BFSI SME send a service or transactional auto-dialer call? TRAI's 2025 text says such calls "should be permitted through 1600 or any other series allotted for the purpose" and that "service or transactional robocalls should not be subject to any restrictions" (para 18). Read alongside the Jul 2026 eligibility limit, there is a real gap: no series is clearly allotted to a general SME for automated service calls. I have not found a primary source resolving it. If you are about to spend money on an outbound build, that is a question for your access provider in writing, not for a blog.

Almost every article on this topic paraphrases a paraphrase. Here is the source, so you can go and read it.

The governing instrument is the Telecom Commercial Communications Customer Preference (Second Amendment) Regulations, 2025 (1 of 2025), No. RG-25/(25)/2023-QoS, notified at New Delhi and dated 12 Feb 2025. It amends the Telecom Commercial Communications Customer Preference Regulations, 2018 (6 of 2018), and is made under section 36 read with sections 11(1)(b)(v) and 11(1)(c) of the TRAI Act 1997 (TRAI Gazette notification PDF, checked 20 Jul 2026).

That matters for three reasons.

1. It tells you the framework is old, not new. The 2018 regulations already governed commercial calling. The 2025 amendment tightened definitions, added the auto-dialer intimation duty, and hardened the enforcement ladder. Nobody invented a new AI rulebook.
2. It tells you what TRAI actually thinks about AI. The only references to AI and machine learning in the amendment are obligations on access providers to deploy AI/ML solutions to *detect* spam, robocalls, auto-dialer calls, pre-recorded announcements and SIM-box usage (regulation 22(1)(h)-(i)). There is no provision in the amendment regulating AI voice agents as a category, and none requiring a caller to disclose to a human that they are speaking to a machine.
3. It tells you where the money actually lands. More on that in the enforcement section, because it is the single most misreported part of this topic.

The practical upshot: when a vendor tells you their platform is "TRAI compliant", the correct follow-up is "under which regulation number". Compliance here is not a property of software. It is a property of your registration, your number series, your consent records and your call content. A platform can make those easier. It cannot hold them for you.

Every obligation in the TCCCPR framework hangs off which of three buckets your call falls into. Get the bucket wrong and every downstream control is wrong too.

All quoted definitions are from the TCCCPR (Second Amendment) Regulations 2025, substituted clauses 2(f), 2(g) and (av), checked 20 Jul 2026.

The contamination rule that catches everybody

This is the sentence to tape to your wall: "if promotional content is mixed with any type of commercial Voice Call, such voice call shall be treated as a Promotional Voice Call" (substituted clause (av)).

One sales sentence inside an otherwise legitimate service call reclassifies the entire call. It now needs promotional consent, DND scrubbing and a 140-series origin. This is precisely where a badly-written AI agent will hurt you: an agent written to be maximally helpful can drift into an offer unless the script forbids it — which is why we freeze promotional lines out of service flows.

Worked example, with an invented business. Meghana Dental, Khammam (invented, not a client) sets up an agent to call patients whose six-month cleaning is due. "Your six-monthly cleaning is due, shall I book you in" is arguably a service call to an existing customer. Add "and we are running twenty percent off whitening this month" and, on the plain text of clause (av), the whole call is promotional. Two clauses in a prompt file, two completely different compliance regimes.

The registration line

The amendment draws the line bluntly: "any commercial communication made by a Sender which is not registered with any Access Provider for the purpose of sending commercial communication shall be treated as an Unsolicited Commercial Communication" (substituted clause (bw)). UCC expressly excludes transactional messages and calls, service messages and calls, communications made on Central or State Government direction in public interest, and communications directed by the Authority. Registration, in other words, is not optional paperwork. It is the thing that stops your legitimate marketing from being classified as spam by definition.

Every widely-read page on this topic is written for outbound cold calling. That is not what most Indian small businesses want. What they want is a phone that gets answered at 9pm and on Sunday.

So: does an AI receptionist answering your inbound calls need DLT registration, a 140 number, and DND scrubbing?

Read the definitions again. Every one of them describes a call made by a Sender to a recipient. "Promotional Voice Call", "Service Voice Call", "Transactional Voice Call", "Unsolicited Commercial Communication" — the whole architecture is about outgoing commercial communication and the recipient's registered preference against receiving it. A customer who dials your published business number and hears an agent has not received an unsolicited commercial communication. They initiated it.

I want to be careful here, because this is exactly the kind of place where a vendor overstates. I have not found a TRAI provision that says in terms "inbound answering is out of scope". What I can say is that the entire preference-register mechanism, the number-series mechanism and the UCC definition are constructed around outbound calls, and none of them has an obvious operative effect on a call you did not place. That is a reading of the primary text, not a ruling.

Where inbound crosses into outbound

The practical risk is that inbound work quietly becomes outbound work. Watch for these:

• Callbacks. A missed call that your system rings back is an outgoing call. Whether it is a service call depends on what it says.
• Reminders. An appointment reminder is an outgoing call to your own customer. Squarely service territory under limb (i) — informational, to your customer, not promotional — provided nobody bolts an offer onto the end.
• Speed-to-lead. Calling a web form submission within sixty seconds is an outgoing call to someone who just asked to be contacted. This is the most defensible outbound motion there is, but it is still outbound, and it is still your job to classify it and to keep the script clean. We wrote about the mechanics separately in speed-to-lead AI calling in India.
• Campaigns. Ringing last year's enquiry list about a new scheme is promotional. Different regime, different number series, different consent.

If you are trying to work out which side of this line your use case sits on, the split is laid out in more detail in inbound vs outbound AI calling, and the product-level version is on /features/ai-voice and /features/outbound-calls.

India's Do Not Disturb system is not one switch. It is eight sectoral blocks, each with an SMS or USSD opt-out code, plus two blanket settings — BLOCK PROMO (*1909*50#) and FULLY BLOCK.

Source: TCCCPR (Second Amendment) Regulations 2025, Schedule preference categories, checked 20 Jul 2026.

The two blanket settings behave differently, and the difference is worth understanding before you assume a number is unreachable:

• FULLY BLOCK blocks service-type commercial communications that require explicit consent, and all promotional types.
• BLOCK PROMO blocks only promotional types, leaving service, transactional and government communication deliverable.

(Notes 3 and 4 to the preference categories, same instrument.)

Consent beats preference

The most commercially important rule in the whole framework: where a Sender has acquired explicit digital consent from the intended recipient, "such Promotional calls with Explicit Consent of the Recipient shall be delivered to the Recipients irrespective of their preferences registered ... in the Preference Register" (Explanation to clause (av) and para 18). Consent is acquired through the access providers' Digital Consent Acquisition platform.

What that means in practice: a properly captured consent is not a nice-to-have that reduces complaints. It is the mechanism that makes a DND-registered number legally callable for promotional purposes. It is the single highest-value compliance asset you can build, and it is built at the point where the customer fills your form or walks into your shop — not later, and not by your calling platform.

And an obligation with a deadline

Registered Senders and registered telemarketers must self-certify annually their registration details and all registered headers, content templates and consent templates. "Any failure ... shall lead to automatic suspension of such Senders and RTMs or their registered headers, content templates and Consent Templates" (regulation 22(1)(c)). Automatic. Put it in a calendar the day you register.

If you take one action item from this page, take this one.

The 2025 amendment substituted Regulation 4 with the following: "Intimation regarding use of Auto Dialer or Robo-Calls.— 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."

Read it slowly. The duty is:

• On the Sender — that is you, the business, not your platform vendor and not your telemarketer.
• In advance — before you start, not after the first complaint.
• In writing — an email you can produce later, not a phone conversation.
• Including the intended objective — so you have to state what the calls are for, which forces you to classify them as promotional, service or transactional before you go live.

An AI voice agent placing outbound calls is, on any ordinary reading, an auto-dialer or robocall arrangement. I have not seen a single competitor page surface this clause, and it is the cheapest compliance step available to you. A written intimation costs nothing and creates a paper trail showing good faith at exactly the moment you would most want one.

Here is the sequence I would follow before any outbound AI calling goes live in India. Timelines are deliberately absent because I could not verify registration turnaround times or fees from any primary source, and neither, as far as I can tell, can the vendors who publish them.

1. Decide the bucket. Promotional, service or transactional — per call type, not per campaign. Write it down.
2. Register as a Sender with an access provider on the DLT platform for the purpose of sending commercial communication.
3. Register headers, content templates and consent templates for what you actually intend to say.
4. Send the Regulation 4 intimation in writing to your originating access provider, stating the use of an auto-dialer and the objective.
5. Get the right number resource. Promotional auto-dialer traffic on 140. For service or transactional, ask your provider in writing which series you are eligible for given the Jul 2026 clarification.
6. Wire consent capture into the form or counter where the customer actually is, via the DCA platform.
7. Freeze the script so promotional lines cannot leak into service calls. This is a prompt-engineering control with legal consequences.
8. Diarise annual self-certification.

If a vendor cannot walk you through steps 1 to 5 in their own words, they are selling you software, not a compliant motion.

This is the section where the popular articles are most confidently wrong, and correcting it should genuinely change how you think about risk.

The 2025 amendment does contain a table of rupee financial disincentives — one thousand rupees per instance, five thousand rupees per count of registration found non-compliant, ten thousand rupees per instance, two lakh rupees for misreporting, five lakh rupees for a second consecutive misreporting, ten lakh rupees for each subsequent consecutive misreporting (regulations 27 and 28). New regulation 28A caps the total at fifty lakh rupees per calendar month per Licensed Service Area.

Those amounts are levied on access providers — the telecom operators — not on the calling business. If you have read that a business faces a fixed rupee penalty per violating call, that figure is a blog-level paraphrase that has conflated carrier-side disincentives with business-side liability. I looked for a per-call fine on senders in the primary text and there is not one.

What the instrument actually says about businesses is that TRAI may "at its discretion, impose financial disincentives on registered Senders and Telemarketers or forfeit their security deposit, and also suspend or blacklist them, in case violation of the regulations can be attributed to the failure of such entities" (clause (g)). Discretionary. No tariff.

The teeth are elsewhere, and they are much sharper than a fine.

All from the TCCCPR (Second Amendment) Regulations 2025, para 71 and regulations 22 and 29, checked 20 Jul 2026.

Sit with the second row. Not the marketing line — *all* telecom resources, across *all* operators, for a year. For a business that runs on the phone, that is not a cost line. That is the end of the business. And the DID-counted restoration charge means a firm running a trunk with fifty numbers is looking at ₹2,50,000, not ₹5,000 — and a hundred DIDs hits the ₹5 lakh cap.

You are not defenceless. A Sender may contest: TRAI may act "on receipt of a complaint from the Sender, within sixty days of action taken against it by the Access Provider", and may direct restoration of telecom resources and deletion from the blacklist if the investigation lacked adequate evidence (substituted regulation 29). Sixty days is the window. Keep your consent records and your call logs in a form you could hand to a regulator, because that appeal is an evidence contest and you will be the one who has to produce evidence.

TRAI governs the call. The Digital Personal Data Protection Act 2023 governs the data the call produces — the phone number, the recording, the transcript, whatever the customer told your agent about their health or their finances.

On 14 Nov 2025, MeitY published Gazette notifications bringing into force the DPDP Act 2023 and the DPDP Rules 2025, both on staggered timelines, and establishing the Data Protection Board in the NCR with four members (Shardul Amarchand Mangaldas & Co, "Enforcement of the DPDP Act and notification of the DPDP Rules", checked 20 Jul 2026). Three enforcement dates were notified.

The practical upshot, and I think this is the fair framing: the DPDP obligations that actually bite an AI calling deployment — notice, consent, security, breach reporting — are not enforceable until 14 May 2027. But the Act and Rules are already notified. The obligations are known today. Anyone building a system now that will still be running in mid-2027 should build to them rather than retrofit.

The Schedule

The maxima are large enough to be worth reading in the Act's own words.

Source: dpdpa.com, The Schedule (DPDP Act 2023), checked 20 Jul 2026. A sourcing caveat I owe you: MeitY's own PDF and indiacode.nic.in both returned HTTP 403 when I tried to fetch them on 20 Jul 2026, so these figures rest on a secondary compiler rather than on the Gazette. Verify against the Gazette before you rely on them for anything that matters.

What this means for call recordings

A recording of an identifiable person talking to your business is personal data on any ordinary reading. That points at data minimisation, a retention decision you can defend, security on wherever the audio sits, and the ability to answer a customer who asks what you hold.

What I will not tell you is that you must play a recording disclosure within the first fifteen seconds, or cite specific DPDP rule numbers for consent logging. Both claims circulate on vendor blogs. I could not verify either in the Rules text, so I am not repeating them. Whether and how you must announce recording on a voice call in India is a question for a lawyer, and the honest answer from me is that I do not know.

A compliance page that projects total certainty is telling you something about the author, not about the law. Here are the places I could not get a clean answer, stated plainly.

Does the AI have to say it is an AI?

Unsettled in India. I found no provision in the TCCCPR (Second Amendment) Regulations 2025 requiring a caller to disclose to the human recipient that the caller is automated. The only advance-declaration duty the amendment creates runs in the opposite direction — to the Originating Access Provider, in writing, in advance, under Regulation 4. Claims that Indian law mandates recipient-facing AI disclosure appear in search snippets and vendor posts; I could not trace them to a primary source.

What we do anyway: every agent we build identifies the business at the start of the call, and if a caller asks whether they are speaking to a person, the agent says it is an automated assistant and takes a message or a callback request. Not because a regulation compels it, but because the alternative gets you complaints, and complaints — as the enforcement section showed — are the actual trigger mechanism.

What are the legal calling hours?

The "call only between 10am and 7pm" line appears widely. I could not find a blanket window in the primary text. The 2025 amendment treats time band as a dimension of recipient preference, which is not the same thing as a statutory window applying to everyone. So: respect registered time-band preferences, do not call at antisocial hours because it is obviously bad practice, and do not treat any specific clock range you read on a blog — including this one — as the law.

Does TCCCPR apply the same way to B2B calling?

I did not find a clean primary answer on whether calls to numbers registered to businesses rather than consumers are treated differently. Several pages assert a B2B carve-out. I am not going to assert one. Assume the framework applies until someone with a bar licence tells you otherwise.

What does DLT registration cost and how long does it take?

I could not verify fees or turnaround times from any primary source, and we do not publish a figure for them. Competitors publish numbers; those numbers may be accurate or may be guesses, and there is no way for you to tell from the page. Your access provider is the authority on this, and the answer will differ by operator. The absence of a published figure is itself the finding.

Who is liable if the vendor's platform breaks the rules?

The TCCCPR obligations that matter land on the Sender. Regulation 4 says "Every Sender shall notify". Clause (bw) turns an unregistered Sender's communication into UCC. Para 71's consequences attach to the Sender's telecom resources. A vendor can hand you a platform that makes compliance easy, and a good one will. None of that transfers the duty. If someone tells you their platform makes you compliant, ask them to put the indemnity in the contract and watch what happens.

Now the part where I apply the same scrutiny to us.

What we actually do

Dvaarik is a studio, not a self-serve product. I build a custom AI voice agent for each business — I am Rohith Sriramula, a laid-off engineer working from home in Paloncha, Bhadradri Kothagudem, Telangana. There is no team and no office. Seven agents: Receptionist, Leads, Sales, WhatsApp, Instagram, Booking and Payments. Twenty-two scheduled Indian languages plus code-mixed Hinglish, Tenglish and Tanglish, and most world languages.

On compliance specifically, here is the honest division of labour.

The arithmetic, including where it does not work

Voice is ₹2 per minute, whole-minute billing, and that is the permanent base rate, not a founding discount. Chat is ₹2 per conversation. Setup is ₹40,000 one-time, currently free for our ten founding clients — the trade stated plainly: a free build in return for the Care Plan at the founding rate for the first three months, and an honest testimonial plus a short case study after go-live. Care Plan lists at ₹12,000/month, founding rate ₹5,000/month locked for six months, required for the first three months and optional from month four. If we supply the phone number, recharge is ₹2,000 per 30 days, auto-debited from credits. Minimum first credit load is ₹10,000, top-ups from ₹500, credits never expire, and a zero balance pauses gracefully rather than breaking. WhatsApp message costs are billed by Meta to your own account with zero markup from us. Full breakdown on /pricing.

Now the uncomfortable maths. Month one debits ₹2,000 for the number and ₹5,000 for the Care Plan — ₹7,000 of fixed cost before a single minute. A clinic taking 40 calls a month at three minutes each burns 120 minutes, or ₹240. Total ₹7,240 for 120 minutes of talking. That is roughly ₹181 per answered call, or about ₹60 per minute of talk time.

At that volume this is expensive. If those 40 calls are appointment confirmations you were already handling fine, do not buy this. The model only makes sense if the calls you are currently *missing* are worth something. Put your own value on a booking — if it is ₹500, you need about fourteen saved bookings a month to cover the ₹7,000 fixed line. If you cannot look at your missed-call log and believe that, the answer is no. You can do that sum with your own numbers using the missed-call calculator.

When you should pick someone else

• You need legal certainty, not a vendor's reading. On a regulatory question, a telecom or data-protection law firm beats a vendor every time — Shardul Amarchand Mangaldas, whose DPDP enforcement note I cite above, is the kind of source I mean. I have quoted the Gazette accurately and I still cannot give you an opinion you can rely on. If money turns on the answer, pay a lawyer.
• You want to run high-volume outbound cold campaigns. We do not do cold purchased lists, full stop. If that is your motion, a dedicated outbound platform will serve you better and you should read their operational playbooks — while noting that the campaign performance figures on those pages are vendor-asserted and unsourced.
• You want a self-serve product with a free trial. There is no signup here. Every agent is built by hand, which is slower and does not suit someone who wants to try something at midnight.
• You need enterprise procurement, an SLA and a vendor security questionnaire answered by a compliance team. One person at home cannot credibly staff that. /enterprise is honest about what we can and cannot commit to.
• Your volumes are tiny and your phone is already answered. See the arithmetic above. I would rather tell you no than take ₹10,000.