Voice AI Agents in Indian Languages: Where They Work for Customer Calls

A customer rings a shop in Madurai and asks, in Tamil, where their order is. A voice agent answers, understands, looks up the order and replies in Tamil. Five years ago this was a demo. Today parts of it work. This post looks at what speech technology exists for Indian languages, which calls suit a machine, and the legal checks that come before you start dialling.

What the technology can do now

A voice agent has three parts: speech-to-text to hear the caller, a language model to decide the reply, and text-to-speech to speak it. Indian-language support has grown in each part.

• Bhashini, the Government's language technology programme, is described in an October 2025 official document as offering real-time translation for the 22 scheduled languages and tribal languages, with speech recognition among its capabilities. It has been used for live multilingual announcements at the Maha Kumbh 2025.
• Sarvam AI, an Indian company, lists on its developer page a speech-to-text model, Saaras v4, covering 23 languages (the 22 Indian languages plus English), and a text-to-speech model, Bulbul v4 Flash, covering 10 languages including Hindi, Tamil, Telugu, Kannada, Bengali, Marathi, Gujarati and Punjabi.
• Zoho announced in July 2025 two speech-to-text models of its own, for English and Hindi, alongside its agent platform.

Language lists show coverage, not call quality. Accuracy on a noisy phone line, with regional accents, mixed Tamil-English speech and shop-floor background noise, is what decides whether it works, and no list can tell you that. Test with recordings of your own customers.

Good uses: narrow, repeated, low stakes

• Order or delivery status. The caller wants one fact and the system has it.
• Appointment booking and reminders. A clinic, salon or service centre with fixed slots is a natural fit. See our guide to online booking systems.
• Opening hours, address, fees and documents needed. The same ten questions asked all day.
• Call routing and message-taking after hours, so a human receives a clean summary in the morning.

Bad uses: emotional, complex or risky

• Angry complaints, bereavement, medical or safety emergencies.
• Payment disputes and debt collection, where tone and fairness matter.
• Anything where a mis-heard number causes harm, such as an address, an amount or a medicine dose.
• Elderly or first-time callers who are put off by a machine. Always offer a quick way to reach a person.

The rule that works: the agent handles the simple call completely or hands over early with the details already captured. It should never trap the caller.

Rules to check before you start

This section is general information, not legal advice, and telecom rules change.

• Telemarketing and DLT. TRAI's commercial communications regulation was amended in February 2025. Promotional calls use the 140 number series, which has been monitored on a distributed ledger (DLT) platform since October 2024, and service or transactional calls use the 160 series. A voice agent that makes promotional calls must follow these rules like any other caller.
• Consent. Under the framework as described by one law-firm review in March 2026, inferred consent lasts only while the contractual relationship does, and explicit consent for a commercial transaction is valid for 7 days. Do not call people who have not agreed to hear from you.
• AI-specific rules are not settled. That same review said TRAI was still considering an amendment on AI-based telemarketing, disclosure and consent verification, and that the existing regulation did not yet treat AI as a separate risk category. Check the current position before launch.
• Penalties exist. Senders with repeated violations can have telecom resources suspended and can be blacklisted for up to two years.
• Recording and personal data. A recorded call is personal data. Say at the start that the call is with an automated assistant and may be recorded, keep recordings only as long as needed, and review them under the DPDP Act with your adviser.

A sensible way to start

1. Pick one inbound call type, such as order status, and count how many calls it is today.
2. Record fifty real calls with permission and test speech recognition on them.
3. Start with inbound calls before any outbound dialling.
4. Announce clearly that callers are speaking to an automated assistant.
5. Set a handover trigger: two failed attempts or the word "agent" passes the call to a person.
6. Review a weekly sample of calls and fix the misses.

Many small firms get more value from a WhatsApp flow first, since text is cheaper and easier to check; see WhatsApp for business in India. VS Tech Builders builds the booking, order and customer systems that a voice agent must read from, with a clean API so voice and chat can share one source of truth. Contact us to scope a pilot, and read how AI agents in business software work for the wider picture.

Sources

Checked on 3 November 2026. Technology changes quickly; always confirm details with the official source before you act on them.

• 22 Languages, Digitally Reimagined — Press Information Bureau, Government of India
• Models — Sarvam AI Documentation
• Zoho Launches Zia LLM, Introducing Prebuilt Agents, Agent Builder, MCP and Marketplace — BigDATAwire
• TRAI's crackdown on spam calls and AI-driven telemarketing — Bar and Bench

Frequently asked questions

Can AI voice agents speak Tamil and Hindi?

Yes, the building blocks exist. Sarvam AI lists speech recognition for 22 Indian languages plus English and text-to-speech for ten languages including Hindi and Tamil, and Zoho announced Hindi and English speech models. Quality on real phone lines varies, so test recordings of your own customers before launching.

What calls should a voice AI agent handle?

Narrow, repeated, low-risk calls: order status, appointment booking and reminders, opening hours, fees and document lists. Keep complaints, disputes, emergencies and any call where a misheard number could cause harm with human staff, and always give callers an easy way to reach a person.

Do TRAI and DLT rules apply to AI calling?

Promotional calls must follow TRAI's telemarketing framework, including the 140 series and DLT monitoring, whether a human or a machine dials. A March 2026 legal review said AI-specific provisions were still under consideration. Check TRAI's current position and take legal advice before outbound campaigns.

Do I need consent to record calls handled by an AI agent?

Recorded voice is personal data, so tell callers at the start that the call is automated and may be recorded, keep recordings only as long as needed, and handle them in line with the DPDP Act. Exact requirements depend on your case, so confirm with a lawyer.