Introduction
For a long time, “serious” conversational AI meant something built in San Francisco or Beijing. That assumption is quietly breaking. Sarvam AI, India’s full-stack AI platform, showed up at the India AI Impact Summit 2026 with a portfolio of foundation models and an end-to-end agent stack that doesn’t feel like a second-tier copy of an American product. It feels like the real thing.
This article walks through a hands-on test of Sarvam’s conversational voice agent — a customer-support persona called Rupa built on top of a fictional e-commerce platform — and uses that demo as a lens to talk about what actually matters in production voice agents: empathy modeling, escalation, refund logic, latency, and when a polite scripted reply does more harm than good. By the end you’ll have a clear picture of where India’s AI ecosystem sits in 2026, and a checklist for evaluating any conversational agent platform — not just Sarvam’s.
📚 Table of contents
- What Sarvam AI actually is
- Why India needed a full-stack AI platform
- Inside the demo: Rupa, the QuickCart support agent
- What the agent did well
- Where it cracked — and what that teaches us
- Anatomy of a production-grade voice agent
- Use cases beyond e-commerce
- How to evaluate any conversational AI platform
- Best practices for shipping voice agents
- Common mistakes to avoid
- India’s AI moment — what the Impact Summit signals
- Try Sarvam AI yourself — dashboard, docs, and product video
- Conclusion
- Frequently asked questions
🇮🇳 What Sarvam AI actually is
Sarvam positions itself as a full-stack AI platform built in India, for Indian and global use cases. Full-stack here is doing a lot of work — it means the company isn’t just shipping one foundation model and calling it a day. It owns the layers most builders care about:
🧠 Models
- Multilingual LLMs tuned for Indic languages and English
- Automatic speech recognition (ASR) for accented and code-mixed speech
- Text-to-speech (TTS) with natural Indian voices
- Translation and transliteration models that handle Hinglish-style input
🛠️ Agent stack
- A no-code console for designing conversational personas
- Tool calling, knowledge base attachment, and structured action outputs
- Voice-first runtime — ASR → LLM → TTS stitched end to end
- APIs for embedding agents into existing apps and contact-center stacks
The shorthand to keep in your head: Sarvam is to Indian conversational AI what ElevenLabs + Anthropic + Voiceflow are, bundled together and tuned for accents the rest of the market keeps butchering.
📡 Why India needed a full-stack AI platform
On paper, you can already build voice agents with off-the-shelf parts — OpenAI for the brain, Deepgram for ASR, ElevenLabs for the voice, a glue layer in Python. That works for English-speaking, Silicon-Valley-shaped problems. It falls apart for the Indian market in three predictable ways.
🗣️ Accent & code-mix
A real Indian user says “mera order kab aayega yaar?” not “could you tell me when my order will arrive?” Generic ASR models trained on American English transcripts trip on that the moment a sentence switches language mid-clause.
💸 Cost per minute
Indian use cases are volume-heavy and margin-thin. A bank that runs a million collection calls a month can’t pay $0.10 per minute for a foreign vendor. Locally hosted models with INR pricing are the only path to viable unit economics.
🔒 Data residency
Banking, healthcare, and government workloads can’t legally ship customer audio to a foreign cloud. A homegrown platform that runs inside Indian data centers solves a compliance problem most international stacks pretend doesn’t exist.
Those three problems are what the “full-stack” framing is really for. You can’t solve any of them without owning the model, the inference layer, and the deployment story together. That’s the bet Sarvam is making.
🎬 Inside the demo: Rupa, the QuickCart support agent
The agent was configured with a system prompt that read, roughly: “You are Rupa, a customer support agent at QuickCart, an e-commerce platform for electronics and home appliances. You handle order tracking, returns, refunds, and product inquiries.” Voice-only conversation, no chat fallback. Click Start Speaking, and the agent answers in real time.
🧩 The test scenario
- User has ordered an air purifier.
- The promised delivery date passed — it hasn’t arrived.
- It’s been pushed by a day, multiple times in a row.
- The user is travelling tomorrow and is past their patience threshold.
- They want either a real explanation, a refund, or compensation — ideally all three.
This is a deliberately spicy scenario. It’s the exact kind of call that breaks bad chatbots and burns out human agents. Empathy, accountability, refund authority, and the ability to not over-promise all matter inside the same two-minute window.
✅ What the agent did well
The demo wasn’t a brochure-friendly happy path. The user pushed, interrupted, and changed their mind mid-sentence. Rupa held up well in a few specific ways.
🤝 Real-feeling empathy
When the user said they were out of patience, the agent didn’t bulldoze ahead with policy. It acknowledged the frustration, apologized specifically (“repeated delays”, not “inconvenience caused”), and stayed in the user’s emotional frame.
⏱️ Low latency
Round-trip from user-stops-talking to agent-starts-talking was tight — sub-second. That’s the floor for voice conversation to feel human. Above ~1.2 seconds, callers start assuming the line is dead and talk over the agent.
🎯 Bounded actions
When the user asked for a refund, the agent didn’t promise the impossible. It quoted a realistic 5–7 business-day window for the refund to reflect. That’s the agent respecting a real operational constraint instead of telling the user what they want to hear.
🔁 Conversational repair
The user reversed course mid-call (“refund me” then “actually let me wait”). Rupa tracked the state change cleanly instead of stubbornly continuing the refund flow. That’s a sign of decent dialogue-state management, not just a stateless prompt.
🩹 Where it cracked — and what that teaches us
No demo is flawless and the interesting bugs are the ones that show up in every voice agent on the market right now, not just Sarvam’s.
⚠️ Patterns to watch for in any voice agent
- Hallucinated root causes. When pushed for why the delivery was late, the agent invented “unexpected supply chain issues.” The agent had no way to know that — it pattern-matched on a plausible-sounding excuse. In production, the right behaviour is “I don’t have visibility into that, let me escalate” or a real tool call to a logistics system.
- Soft-deflecting concrete asks. When the user asked for a coupon, the agent said it would “see what it can do” later. That language is the bot version of a sales rep buying time. A production agent should either have explicit coupon-issue authority via a tool, or refuse cleanly.
- Repetition of empathy phrases. “I completely understand your frustration” appeared three times in two minutes. After the second instance, the line stops carrying weight and starts feeling robotic. Empathy needs to be rationed.
- No real escalation handoff. Saying “I’ll escalate this” without actually triggering a ticket, a callback, or a human transfer is worse than not promising anything. It teaches users not to trust the agent.
None of these are damning. They’re the standard set of issues every voice-agent team is working through in 2026, and they’re mostly fixable in the system prompt and tool definitions rather than in the model itself.
🧬 Anatomy of a production-grade voice agent
A platform demo is one thing. Running a voice agent that handles a real complaint queue is another. Whether you build on Sarvam, Vapi, Retell, or stitch your own from APIs, the shape of a good agent looks roughly the same.
🎙️ Voice pipeline
- ASR with streaming partials, not just final transcripts
- Voice activity detection — know when the user has stopped
- Interruption / barge-in support
- TTS with prosody control, not flat read-aloud
- Sub-second total round-trip is the floor
🧠 Brain layer
- System prompt with persona, scope, and refusal rules
- Tool calls for any data the agent shouldn’t invent
- RAG over policy / product docs
- Memory layer for repeat callers
- Temperature low enough to stay on policy, high enough to sound human
🛡️ Guardrails
- Output filters for PII, slurs, and competitor names
- Hard caps on what the agent can promise (refund amount, discount %)
- Forced escalation triggers (legal threats, self-harm, fraud)
- Confidence thresholds for handoff to a human
📊 Observability
- Full transcript + audio recording per call
- Latency breakdown: ASR / LLM / TTS / network
- Sentiment trajectory over the conversation
- Tool-call success rates and failure reasons
- CSAT or post-call survey signal feeding back into evals
🧭 Use cases beyond e-commerce
Customer support for online shopping is the canonical demo, but it’s the least interesting application. Conversational agents start mattering when they replace high-volume, low-margin call work that humans hate doing anyway.
- Banking & collections. Reminder calls for missed EMIs, KYC follow-ups, balance enquiries in seven languages. The bot version is faster, polite by default, and doesn’t escalate the way some human collectors do.
- Healthcare scheduling. Appointment booking and reminders in regional languages, with the ability to read out fasting instructions or pre-procedure prep.
- Government service helplines. Pension status, ration card issues, subsidy eligibility — long-tail enquiries that don’t scale with human agents.
- Logistics last-mile. Delivery confirmation, rescheduling, address clarification. Voice beats SMS in rural geographies where literacy in the local language is higher than in English.
- Field-force enablement. An agronomy bot that walks a farmer through pesticide ratios in Marathi while they’re standing in a field.
- Education & tutoring. Spoken English practice for learners who can’t afford a human tutor — conversational, not flashcards.
📌 The pattern across all of these: a conversation that has to happen, in a language other than English, where a human can’t scale to demand. That’s the wedge.
🔍 How to evaluate any conversational AI platform
Don’t take vendor demos at face value — including this one. Here’s a practical rubric you can apply to Sarvam, or any of its competitors, before you commit a workload.
🧪 The eval checklist
- Adversarial inputs. Throw heavy code-mixing, accents, background noise, and partial sentences at the ASR. See how it degrades.
- Refusal tests. Ask for things outside scope — legal advice, competitor pricing, personal data. A platform that hallucinates here will hallucinate in production.
- Latency under load. Single-user demos always feel snappy. Test concurrency and measure p99, not p50.
- Tool-call reliability. Build a fake order-lookup tool and intentionally make it fail or return malformed data. Watch how the agent recovers.
- Long-conversation drift. Twenty-minute call. Does the agent forget the user’s name? Repeat itself? Drop the system-prompt persona?
- Compliance posture. Where is the audio stored, for how long, who has access, what gets sent to the model provider? Get answers in writing, not in a sales call.
- Total cost of ownership. Per-minute model cost is the small number. ASR, TTS, telephony, dev time, and observability are the rest of it.
🏗️ Best practices for shipping voice agents
✅ Do this
- Write the persona as if it were an actual job description — role, scope, what it can and can’t do
- Define an explicit escalation policy — what triggers a human transfer, and how that handoff happens
- Cap utterance length so the agent doesn’t monologue
- Log everything: audio, transcript, tool calls, latency per stage
- Run weekly evals on a frozen set of hard calls — track regressions like you would a unit test suite
- Ship a human-in-the-loop fallback before you ship full automation
❌ Avoid this
- Letting the agent invent reasons it doesn’t actually know
- Saying “I’ll escalate” without an actual escalation tool wired up
- Hard-coding empathy phrases that repeat verbatim every call
- Skipping a fallback when the model returns low-confidence output
- Over-promising refunds, discounts, or timelines the business can’t honour
- Shipping in only English when 70% of your callers speak Hindi or a regional language
🚫 Common mistakes to avoid
- Treating the bot as a brand mascot, not an employee. A bot that sounds fun but can’t actually act on a refund is a worse customer experience than a competent human.
- Skipping the “I don’t know” pathway. Production-grade agents need a clean way to say “I don’t have that information” and route the caller. Otherwise they hallucinate.
- Building English-first, translating later. Indic languages have their own pragmatics. Translating an English persona word-for-word produces culturally tone-deaf replies.
- Ignoring telephony quality. The fanciest agent on Earth sounds terrible on an 8 kHz PSTN call. ASR accuracy plummets. Test on real phone audio, not studio mics.
- No human review of failure cases. If your support team isn’t listening to a sample of bot calls every week, you have no feedback loop.
🌅 India’s AI moment — what the Impact Summit signals
The bigger story isn’t one demo. The India AI Impact Summit 2026 was the first time the Indian AI conversation moved past “wrapper startups on OpenAI” and into foundation-model territory. Sarvam announced new models. Other Indian labs followed. Government funding lined up behind sovereign compute. The pattern looks a lot like what happened in China between 2022 and 2024 — an ecosystem catching up by building its own primitives instead of renting them.
🧭 What this means for builders
- You now have a credible Indian alternative when compliance or pricing rules out a US vendor
- Indic-language workloads are first-class, not an afterthought
- Hiring for AI roles inside India will get easier as more local options compete for engineers
- Expect more vertical Indian platforms — legal AI, agri AI, BFSI AI — built on top of Sarvam-class models
None of this means Indian labs have caught up with frontier US models on raw reasoning. They haven’t, yet. But for the specific shape of conversational and multilingual workloads that dominate Indian enterprise demand, a homegrown stack now exists. That’s a real shift.
🔗 Try Sarvam AI yourself
Reading about a voice agent only gets you so far. The fastest way to form an opinion is to log into the console, build a small agent, and throw your own hard scenarios at it.
🌐 Sarvam AI website
Company site — product overview, model releases, and case studies.
sarvam.ai
🎛️ Agents dashboard
Build, configure, and test your own conversational agents in the Sarvam console.
dashboard.sarvam.ai/agents
📘 API documentation
Getting-started guide and full API reference for embedding Sarvam into your stack.
docs.sarvam.ai
▶️ Product walkthrough (YouTube)
Official product video covering the agent builder and live voice demos.
youtube.com/watch?v=5gtZkg9xYAc
Conclusion
A two-minute demo can’t prove a platform is production-ready. What this demo does prove is that a competent, low-latency, emotionally aware voice agent can be built end to end on an Indian stack today — without renting brains from the other side of the world. The cracks in the demo are the same cracks every voice-agent vendor is working through, and most of them are solvable at the application layer.
If you’re building a voice product for an Indian audience, Sarvam is now on the short list of platforms worth piloting. Treat any vendor demo as a starting point, run it through the eval checklist above, and design the agent like a real employee — with scope, authority, and an escalation path — not a chatbot wearing a name tag.
Explore More on DevShelf
-
Claude AI — Tool Profile
Anthropic's Claude powers many Western voice agent stacks — see how it compares to Sarvam's integrated approach.
-
ElevenLabs — Tool Profile
The leading Western TTS platform often paired with frontier LLMs for English voice agents.