AI Voice Agents in India: Hindi, Regional Languages, TRAI DLT and DPDP Compliance

India is the largest phone-first market in the world and one of the hardest to automate well. The language layer is a real engineering problem. The compliance layer is a real legal one. Businesses that have deployed voice agents successfully here have treated both as design constraints from the start rather than as launch-week discoveries.

Language: build for Hinglish, not for Hindi

The mistake is to think of India as a Hindi market with an English fallback. The reality on the phone:

• Hinglish is the default register for a large share of urban callers: Hindi grammar with English nouns, or the reverse, switching several times per sentence.
• Regional languages dominate outside the Hindi belt. Tamil, Telugu, Kannada, Malayalam, Marathi, Bengali, Gujarati and Punjabi callers expect to be understood, and many will hang up on a Hindi-only line.
• English varies from fluent to functional, with Indian English pronunciation and vocabulary that trips up models tuned on American speech.
• Audio quality varies enormously: mobile lines, background noise, speakerphone, and the caller walking through traffic.

A working agent detects language per utterance, handles code-switching without stalling, and confirms critical fields by reading them back, which also helps with the audio problem. Numbers deserve special attention: amounts in lakhs and crores, dates spoken in mixed formats, and phone numbers read in Indian groupings.

Voice selection matters too. A neutral Indian English or Hindi voice built for the market outperforms an imported voice with an Indian language model behind it, both in comprehension and in how callers respond.

Outbound: the TRAI framework

Any outbound commercial voice call in India runs inside TRAI's Telecom Commercial Communications Customer Preference Regulations and the Distributed Ledger Technology (DLT) system that operators run under them. The practical requirements:

• Register as a Principal Entity on a DLT platform run by a telecom operator, with the business's PAN, GST and authorisation documents, then register headers and templates for the campaigns you will run.
• Scrub every list against DND preferences before dialling, and refresh the scrub regularly. Calling a DND-registered number is the fastest way to get a header blocked.
• Use the correct number series. Promotional calls use the 140 series; transactional and service calls use the designated series for that category. Calling from ordinary numbers for commercial purposes is non-compliant.
• Respect the calling window. Promotional calls are limited to 9 am to 9 pm.
• Honour opt-outs immediately and keep records.

Whether a call is transactional or promotional decides which rules apply. "Your order has shipped" is transactional. "Would you like to upgrade" is promotional. An agent that starts transactional and drifts into a pitch has changed category mid-call, and the compliance obligations change with it.

Inbound calls to your own number from your own customers are not subject to the outbound telemarketing rules, which is one reason inbound is the right place to start.

Data: the DPDP Act applies to every call

The Digital Personal Data Protection Act 2023 governs the personal data a voice agent inevitably handles: phone numbers, names, transcripts, recordings. The obligations include a lawful basis (usually consent) with clear notice of purpose, purpose limitation, data minimisation, the ability to honour correction and erasure requests, grievance redressal, and breach notification. Penalties for serious violations reach ₹250 crore.

For voice agents, the design implications are concrete: disclose at the start of the call that the caller is speaking to an automated assistant and that the call is recorded and processed, collect only what the workflow requires, define retention for recordings and transcripts, and keep the consent record alongside the call record. Consent captured under TRAI rules and consent required under DPDP should be reconciled into one record rather than treated as two separate boxes.

Data residency is a decision to make deliberately. Many enterprise and BFSI buyers require that recordings and transcripts stay in India, and the infrastructure should be chosen to make that possible.

Lenders: RBI rules on top

If you are a bank, NBFC, or fintech operating under the RBI's digital lending guidelines, or a vendor serving one, the RBI's Fair Practices Code and recovery agent norms govern collection and servicing calls. In practice:

• Calling hours for recovery contact are 8 am to 7 pm, and this applies to automated calls, retries and callbacks.
• The agent must identify the lender and itself, state the purpose, and provide grievance redressal information.
• No harassment. Threatening or abusive language, excessive frequency and disclosure of the debt to third parties are prohibited, and the agent's scripts must be built to make them impossible rather than merely discouraged.
• A human escalation path must exist.
• The lender carries liability for the conduct of outsourced agents, human or automated. Vendor contracts should make the compliance split explicit.

We cover the collections use case in depth, including the US rules, in our guide to AI voice agents for loan collections.

Where voice agents are working in India

• Customer support for consumer apps and marketplaces: order status, returns, delivery rescheduling, KYC follow-ups, in the caller's language, integrated with the order and ticketing systems.
• Appointment and lead management for clinics, diagnostic chains, education providers and real estate developers. Our piece on AI in education and edtech touches the admissions and retention side.
• Payment reminders and collections for lenders, within the RBI framework, with UPI payment links sent by SMS or WhatsApp during the call.
• Insurance servicing for renewals, claim status and policy queries, with IRDAI disclosure requirements built into scripts.
• Logistics and delivery coordination, where the same agent confirms addresses and reschedules in Hindi, English or the regional language of the delivery zone.

A realistic path

Start inbound. Pick one workflow with a clear outcome and a real volume of calls. Record two weeks of calls, with consent, so the discovery sprint can measure the true language distribution and audio conditions. Build multilingual from day one with read-back confirmation on every critical field. Run in shadow mode, then after hours, then on overflow. Instrument by language and by outcome. Only then add outbound, with DLT registration, DND scrubbing and consent records in place before the first dial.

Claudeter has a team in India and builds these agents with the language stack, consent architecture and, for lenders, the RBI conduct controls designed into the workflow from the discovery sprint.

Frequently asked questions

Is AI voice calling legal in India?

Yes. There is no law banning automated voice calls, but outbound commercial calls must comply with TRAI's regulations: DLT registration as a Principal Entity, header and template registration, DND scrubbing, approved number series and the 9 am to 9 pm window for promotional calls. The DPDP Act governs the personal data involved, and sector regulators such as the RBI and IRDAI add rules for their industries.

Can an AI voice agent handle Hinglish and regional Indian languages?

It can, when it detects language per utterance rather than per call, handles code-switching, and confirms critical fields by reading them back. Regional language coverage should be chosen based on the actual caller distribution measured from real recordings.

What does the DPDP Act require for AI call recordings?

A lawful basis with clear notice of purpose, purpose limitation and minimisation, processes to honour correction and erasure requests, grievance redressal and breach notification. Recordings and transcripts are personal data and should have defined retention and a consent record attached.

What are the RBI rules for AI collection calls?

Recovery contact only between 8 am and 7 pm including retries, identification of the lender and the caller, no threatening or abusive language, no disclosure to third parties, frequency limits, a human escalation path, and grievance redressal information. The lender remains liable for the conduct of any outsourced agent, human or automated.

Sources: Oli AI, what RBI, TRAI and DLT rules say about AI collection calls; SquadStack, TRAI compliance for AI voice calling.