AI-powered agents for contact center in India: what they can and cannot do

An AI agent in a contact center conducts a customer conversation end to end, deciding what to say, calling systems to retrieve or change data, and handing it to a person when it should. It differs from agent assist, which stays alongside a human and suggests rather than acts.

Indian deployments are well past pilot. Meesho runs a voice bot handling 60,000 calls a day in Hindi and English, with a reported 75% cut in support costs and average call handling time halved. Bajaj Finance told its Q1 FY27 earnings call that AI voice and text bots now handle 71% of its do-it-yourself customer service volumes, at roughly one third the cost of human labor.

How does an AI voice agent work?

Four stages run in sequence: speech recognition, a language model, speech synthesis and telephony. The whole loop has to finish before the silence reads as a fault, which makes architecture decide more than model choice.

Each stage adds delay. Speech recognition turns audio into text. A language model decides the response, often calling an external system to look up an order or a balance. Text-to-speech renders the reply, and telephony carries it back. The whole loop has to complete before the pause feels wrong, which is why a system that demos well can still fail on a real Indian mobile network.

The published figure for Twilio Conversation Relay is a median under 0.5 seconds and under 0.725 seconds at the 95th percentile, measured mouth to ear on Twilio internal benchmarks. 

How accurate is Indian language speech recognition?

No engine is reliable across every Indian language, so engine choice is an architecture decision rather than a one-time procurement one.

The Voice of India benchmark from AI4Bharat at IIT Madras, built from unscripted telephone calls rather than clean read speech, covers 15 languages across 306,230 utterances.

No system held a word error rate under 20% across all 15 languages. Twilio research across 15 countries found the consequence at the customer end: 38% of consumers name accent misunderstanding among their AI failures and 37% name language barriers.

So do not bet a deployment on a single engine. Twilio Conversation Relay lets you select the speech recognition provider between Deepgram and Google, with per-language coverage to confirm first. This way, if one engine underperforms for a particular language, switching providers is a simple configuration change rather than a migration. For text-to-speech, you can choose from Google, Amazon Polly, or ElevenLabs.

Where you need an engine outside that set, and the Apache 2.0 release of Sarvam’s models in 2026 makes that live for Indic languages, Media Streams delivers raw call audio over a WebSocket to a model you run yourself.

The reasoning layer stays yours either way, so a mis-transcription only becomes a mis-action if the model is prompted to assume rather than confirm.

What does RBI expect when customers talk to AI?

The Reserve Bank of India (RBI) recommends telling customers they are talking to AI, human handling of complaints about AI decisions, human final decisions, and a route to a person. They are recommendations rather than binding directions.

The RBI FREE-AI committee reported on 13 August 2025 with 26 recommendations under a stated principle of innovation over restraint. Four bear directly on a contact center.

Customers should be clearly informed when they are interacting with an AI model. Complaints about AI decisions should be handled by human representatives. Final decision-making should vest with humans rather than AI models. And anyone raising a grievance should have the option of shifting to a human.

These are recommendations, not binding directions, and no RBI master direction has given them force. Build for them anyway. The same committee reported that 20.8% of surveyed regulated entities are deploying AI, while 67% expressed interest in exploring AI use cases, so the supervisory questions are coming to an industry that has not answered them yet.

Where should an AI agent hand off?

On low confidence, on repetition, and on any category carrying a financial consequence. Handoff quality is what decides whether automation reduces call volume or raises it.

Three triggers are worth designing before deployment rather than after a complaint.

Confidence. When recognition or intent confidence drops below a threshold, transfer rather than guess. In a market with word error rates above 20% on several languages, this fires more often than teams plan for.

Repetition. A customer repeating themselves is the earliest reliable signal of a failing conversation, and it arrives before audible frustration.

Category. Complaints, disputes and anything with a financial consequence go to a person, because RBI recommends it and because the alternative is defending a decision you cannot reconstruct.

Handoff quality is what separates a working deployment from one that raises call volume. Twilio research found only 15% of consumers experience a smooth handoff from AI to a person, while 78% say being able to switch matters. Transferring without the context the customer already gave is worse than not automating.

Which platforms run AI voice agents in India?

Four groups compete, separated by how much of the stack you keep and whether the model behind the agent is yours or the platform’s.

The four differ most in how much of the stack comes with the platform and how much stays yours.

Start with what the agent runs on. Twilio carries 27.9 billion voice calls a year across 4,800 carrier connections in more than 180 countries, and the model is yours rather than the platform’s. There is no allow-list at the platform layer, so an agent built on OpenAI, Anthropic, Azure or AWS Bedrock runs over Twilio channels without rebuilding it.

The supply side shifted in 2026. Sarvam AI, selected under the IndiaAI Mission, open-sourced models under Apache 2.0, so Indic voice AI no longer requires a US vendor. That is a reason to keep the model layer swappable rather than to sign a five-year contract on today’s best one.

What does the Twilio platform bring to India?

A model you own, a carrier layer underneath it, and the memory and knowledge components that stop a customer repeating themselves on the second call.

The pieces that make an agent work over time went generally available in May 2026, and each targets a documented failure mode.

Agent Connect runs agents you built on OpenAI, Anthropic, Azure or AWS Bedrock across voice, SMS, WhatsApp, RCS and chat, with memory context injected automatically. Twilio Conversation Memory carries facts and prior conversations forward, which matters because 54% of consumers say AI rarely or never has previous context about them. Enterprise Knowledge grounds answers in your own policies rather than the model’s guesses.

Twilio Conversation Relay handles speech recognition and synthesis while your model does the thinking, with Deepgram Flux turn detection added in May 2026 to cut the pauses that make a bot sound wrong. It is PCI compliant and HIPAA eligible, and priced at $0.07 per minute on top of voice.

Genspark, running a consumer voice agent on this stack across more than 40 countries, reports a 94.3% call success rate, 3.2 times higher retention among users of the feature, and sub-300 millisecond latency, having scaled from 10 calls a day to more than 800.

Twilio provides toll-free Indian numbers and domestic outbound voice runs through your carrier, so the working architecture is your Indian carrier connected over Twilio Elastic SIP Trunking, the Twilio service that connects an existing carrier to the platform, with the agent above it.

Twilio Conversation Relay handles Hindi, Tamil, Kannada, Malayalam, Marathi, Telugu and Indian English for live conversation, while Conversational Intelligence transcription covers English, European and Latin American locales only, so post-call analytics on a vernacular call needs its own component.

Will AI agents replace Indian contact center jobs?

The Indian record shows real reductions and real expansion, sometimes at the same company, which makes the transition plan the variable rather than the technology.

The Indian record is more specific than the argument usually allows, and it cuts both ways.

Eternal’s Nugget resolves 85% of roughly 15 million monthly conversations across Zomato, Blinkit and Hyperpure, taking annual support costs from $20 million to $9 million and lifting human agent productivity 40%.

It has also coincided with real reductions: around 600 support roles in April 2025, and the closure of a Hyderabad support center in August 2026 affecting 250 to 300 people, all employed through third-party providers rather than on direct payroll.

Bajaj Finance points the other way on skills. Its AI unit is expanding from 230 to 400 people while bots handle 71% of self-service volume. The pattern across both is a shift in what the workforce does rather than a clean subtraction, and it favors the operators who plan the transition rather than discover it.

What if you run contact centers for clients?

The agent has to be per-client rather than per-platform: isolated credentials, per-client reporting, and a carrier arrangement each client can keep.

An outsourcer needs the agent to be per-client, not per-platform. Sub-accounts give each client its own credential isolation, API keys and usage record, so agent performance, cost and audit trail can be reported client by client rather than reconciled from a shared pool.

Messaging Services sit below that as the per-client boundary for sender IDs and opt-out handling. Bring-your-own-carrier answers the client whose carrier contract cannot move, which is the most common blocker on an Indian transition.

Two constraints to design around. No Twilio Interconnect point of presence sits in India, with Singapore and Tokyo nearest, so plan media routing from a Bangalore or Hyderabad floor accordingly. And Elastic SIP Trunking termination defaults to one call per second, raisable to five, which matters if a client expects predictive dialling.

Billing by the communication event rather than the agent seat suits a book of business with high seat churn and client contracts priced per interaction.

Frequently asked questions

What is an AI voice agent in a contact center? It is software that conducts a customer call end to end, using speech recognition, a language model and text-to-speech, and calling business systems to act on what it is told. Agent assist is different: it stays alongside a human and suggests rather than acts.

How accurate is Hindi and Indian language voice AI? On the Voice of India benchmark of unscripted telephone speech, the best system ranged from 5.0% word error rate on Hindi to 24.8% on Maithili, and no system held under 20% across all 15 languages. Test on your own recordings, per language, on telephone-quality audio.

Do Indian rules require telling a customer they are talking to AI? Not as law. The RBI FREE-AI report of 13 August 2025 recommends it for banks, alongside human handling of AI complaints, human final decisions and a route to a person. Those are recommendations. MeitY’s February 2026 rules cover synthetic media labeling and do not require chatbot disclosure.

Which voice AI platforms support real Indian phone numbers? Carrier connectivity separates a platform from a software-only voice tool. Twilio provides toll-free Indian numbers and connects an existing Indian carrier over Twilio Elastic SIP Trunking or bring-your-own-carrier, with the domestic leg carried by that carrier. Ask any vendor which Indian number types it provisions and who carries the call.

How fast does an AI voice agent need to be? Fast enough that the pause does not read as a fault. The published figure for Twilio Conversation Relay is a median under 0.5 seconds and under 0.725 seconds at the 95th percentile on Twilio internal benchmarks. Very few platforms publish a latency figure with a stated method.

Can an AI agent handle Hindi and English in the same call? Code-switching is where systems degrade most, and no public benchmark isolates it. Run your own recordings through any shortlisted engine before committing, and design the confidence-based handoff to catch what the model misses.

What does an AI voice agent cost to run in India? Platform vendors bill the communication event rather than the seat. Twilio Conversation Relay is $0.07 per minute on top of voice, Twilio Flex starts at $35 per monthly active user or $1 per active user hour, and Amazon Connect Customer publishes $0.038 per voice minute with no seat fee. Model the bill against your own call volumes.

Read the Twilio Conversation Relay overview for the architecture.

Sources

• Voice of India: A Large-Scale Benchmark for Real-World Speech Recognition in India, AI4Bharat at IIT Madras with Josh Talks, 2026

• Svarah: Evaluating English ASR Systems on Indian Accents, AI4Bharat, IIT Madras and Microsoft India

• RBI FREE-AI committee report, 13 August 2025, and the RBI survey of regulated entities reported within it

• TechCrunch on Meesho, November 2024

• Medianama on Bajaj Finance Q1 FY27, 7 August 2026

• MongoDB on Eternal’s Nugget, March 2026, and Business Standard on the Hyderabad center closure, 31 August 2026

• Twilio, Inside the Conversational AI Revolution, November 2025, 4,800 consumers and 457 business leaders across 15 countries

• Twilio SIGNAL 2026 product announcements, May 2026

• Genspark customer story and Twilio global infrastructure