The Complete Guide to Voice AI for Indian Ecommerce (2026)

This is the guide we wish had existed when we started building voice agents for Indian D2C brands. It covers how the technology actually works, where it earns money and where it doesn't, what it costs, why Hinglish breaks most systems, what the compliance picture looks like, and how to roll one out in thirty days without irritating your customers.

It's long. It's meant to be the page you come back to rather than read once. If you're only here for one thing, skip to the 30-day rollout plan — everything before it is the reasoning that makes the plan make sense.

Why voice, and why now

Indian ecommerce has a structural problem that no other large market has at quite the same scale: a very large share of orders are paid in cash on delivery, and a very large share of those never complete. Return-to-origin rates in the 20–30% range are normal. Each one costs forward freight, reverse freight, handling, and three weeks of inventory locked in transit. For most D2C brands, RTO is the second-largest line item after ad spend, and unlike ad spend it produces nothing.

The fix has never been a mystery. Talk to the customer. Confirm the order before you ship it. Call them when delivery fails. Fix the address before the courier gets lost. Every retailer who has ever run a shop understands this instinctively.

What blocked it was arithmetic. Doing this properly means calling every COD order within hours, in the customer's language, plus every failed delivery, plus every high-value abandoned cart — spiky, unpredictable volume, most of it needing to happen in a narrow window. Staff for the peak and you're paying idle salaries; staff for the average and you miss the days that matter. Below a few thousand orders a month, the maths simply doesn't work, which is why the brands that solved RTO were the ones large enough to run a call centre.

That's the constraint that broke, and it broke fairly recently. Three things had to land at once: speech recognition good enough for Indian-accented, code-mixed speech; language models fast enough to respond in a conversational beat rather than an awkward pause; and text-to-speech that doesn't sound like a railway announcement. By 2026 all three are commodity infrastructure. The result is that a brand doing 800 orders a month can run the same operational playbook as one doing 80,000.

How a voice AI agent actually works

Worth understanding the pipeline, because nearly every failure mode you'll encounter maps to one of these stages, and knowing which one is broken is the difference between fixing it and switching vendors pointlessly.

The telephony layer

Something has to actually place the call over the Indian phone network. This means a telephony provider, a number (or a pool of them), and SIP trunking to carry the audio. The practical concerns here are unglamorous but decisive: does your caller ID show consistently, what's your connect rate on different circles, and — the one that bites everyone — do you have enough balance and enough concurrent channels for a morning burst of four hundred calls.

Speech to text

The customer's audio is transcribed in real time. This is the single most important component for Indian use cases and the one where general-purpose models disappoint most. More on that in the Hinglish section, because it deserves its own treatment.

The reasoning layer

A language model reads the transcript plus context — the order, the customer's name, the product, the conversation so far — and decides what to say and what to do. "What to do" is the part people underrate: the agent should be able to call functions. Mark the order confirmed. Tag it cancelled. Record a delivery window. If the agent can talk but can't write anything back to your systems, you've bought a very expensive answering machine.

Text to speech

The response is spoken. Quality matters more than you'd expect — not for vanity, but because an obviously synthetic voice gets hung up on. Indian-accented voices meaningfully outperform generic ones on Indian numbers, and correct pronunciation of Indian names and place names is a real differentiator that generic vendors quietly fail at.

The orchestration around all of it

The unsexy layer that decides everything. When does the agent decide the customer has finished speaking? What happens when they interrupt mid-sentence? How long is too long a silence? What happens on voicemail, on a busy tone, on a wrong number? How many retries, at what times of day? Two systems with identical models can perform completely differently based on this layer alone. It's where most of the real engineering lives.

Why this isn't an IVR (and why that matters)

Most Indian consumers' experience of "automated call" is an IVR: press 1 for this, press 2 for that. They have learned that this experience is bad, and they hang up on it fast. If you introduce voice AI to your team as "like an IVR but smarter," you'll get a lot of resistance from people who have already watched IVR fail.

The difference is categorical, not incremental:

The practical upshot: a customer who would abandon an IVR will finish a 25-second conversation with a competent agent, because it doesn't ask them to do work. They just answer a question, the way they would with a person.

Where it pays: the use cases ranked by ROI

We've covered this in depth in 7 ecommerce calls every AI voice agent should automate. The short version, ranked honestly by return rather than by how interesting they are:

1. COD order confirmation — start here, always

Call before dispatch, confirm the customer wants the order, tag the result in your store. This is the highest-ROI automation available to an Indian D2C brand, and it isn't close. It intercepts the 10–20% of COD orders that were never going to be accepted, before you pay a rupee of freight on them. Everything else on this list is funded by the savings from this one. The full setup is in the Shopify RTO guide.

2. NDR recovery — highest value per save

When a delivery attempt fails, call within hours, get a real delivery window, push it back to the 3PL. These are orders already 90% converted; losing them at the last mile is the most expensive failure in the funnel. The full sequence is in what happens after a customer doesn't answer your delivery call.

3. Pre-dispatch address and scheduling checks

Confirm the address is complete, capture a landmark, ask when someone will be home. Shrinks NDR upstream rather than recovering from it. Often folded into the confirmation call — one call, two jobs.

4. High-value abandoned carts

Only above a value threshold, and never with a reflexive discount. Recovers at full price and tells you why carts are dying — information you cannot get from analytics.

5. Returns and exchange coordination

Collapses a week of WhatsApp back-and-forth into one call, and converts a chunk of refunds into exchanges. Matters most for apparel and footwear.

6. Post-delivery check-ins and replenishment

Genuine upside, easy to overdo. Sample rather than blanket-call, and only for brands where the repeat cycle is predictable.

The Hinglish problem

This is where most voice AI deployments in India quietly underperform, and where vendor demos are least representative of reality.

Indian customers don't speak Hindi. They don't speak English either. They speak both, in the same sentence, often in the same clause: "Haan bhai order confirm hai, but delivery Saturday ko karwa dena, main office mein rahunga." Hindi grammar, English content words, and a code switch that happens mid-utterance without warning.

Speech recognition systems are typically trained and tuned per language. Set the model to Hindi and the English words come back as garbage; set it to English and the Hindi structure collapses. Either way, the transcript reaching your language model is wrong, and no amount of clever prompting downstream recovers a mangled transcript. Garbage in, confidently wrong out.

What actually helps, roughly in order of impact:

• Code-mixed recognition. Not "Hindi mode," not "English mode" — a mode that expects both within one utterance. This single setting is usually the largest single quality jump available.
• A domain vocabulary. Feed the recogniser your actual product names, brand name, courier names, and the vocabulary of the specific conversation. "Kurti," "Delhivery," "COD," "RTO," and your SKU names are not in a general model's expectations, and priming it with them measurably reduces errors.
• Regional language coverage that's real. Marathi, Tamil, Telugu, Bengali, Kannada, Gujarati. Ask a vendor to demo the ones you actually ship to — the gap between "supported" on a website and usable on a call is frequently enormous.
• Handling accented English. A Chennai customer speaking English is not a US customer speaking English, and models tuned on American speech drop accuracy sharply.

When you evaluate, do not judge from a scripted demo. Get recordings of real calls from real customers in your categories and your regions, and read the transcripts. The failure is always visible in the transcript.

The economics: what it costs and what it saves

Voice AI pricing in India is generally per minute or per connected call. The number that matters isn't either of those, though — it's cost per prevented loss. A confirmation call costs a few rupees. An RTO costs ₹180–350 in two-way freight before you count handling and the inventory sitting in a truck for three weeks. The ratio is not subtle.

Here's how to model it properly for your own brand rather than trusting anyone's marketing arithmetic:

The two inputs people get wrong: attempts per order (you won't reach everyone first try — budget for two or three) and cost per RTO (almost always higher than the freight line alone, once you count warehouse handling and the working capital tied up).

Against a human team, the comparison isn't purely per-call cost either — it's the fully loaded picture including hiring, training, attrition, supervision, and the fact that a team sized for average volume cannot absorb a festive-season spike. We ran that comparison properly in AI vs. human teams vs. SMS. The honest conclusion there: human teams still win on complex, high-empathy conversations, and lose badly on high-volume, repetitive, time-sensitive ones. Confirmation and NDR calls are firmly in the second category.

Build vs. buy

The components are all available as APIs, so building is genuinely possible. Whether it's sensible depends less on engineering capacity than on where your problems will actually come from.

Building gets you full control of the script, the logic and the data, no per-call vendor margin, and the ability to do things nobody offers off the shelf. What it costs you is rarely the happy path — a prototype that makes a call and sounds decent is a weekend. The expense is everything after: telephony reliability, retry logic, concurrency during a morning burst, voicemail detection, interruption handling, per-circle connect rates, and the long tail of Indian-language quality work. That's months, and it's ongoing.

Buying gets you the orchestration layer someone else has already debugged against millions of Indian calls, plus platform integrations that already write back to Shopify. What you give up is per-call margin and some control.

Our honest read: build if voice is going to be your product, or if you have volume large enough that per-call margin dominates and an engineer permanently assigned to it is justified. Buy if voice is an operational tool for your ecommerce business — which for almost every D2C brand, it is.

How to evaluate a vendor

Demos are optimised. Here's what to actually ask, and what a weak answer looks like.

Ask for real call recordings, not a demo

Specifically: recordings with real customers, in your language mix, in your category, including calls that went badly. A vendor confident in their system will share the messy ones. Listen for the pause before the agent replies, what happens when the customer interrupts, and whether it copes with "kaun bol raha hai?"

Test the failure modes deliberately

Anyone can handle "haan, confirm hai." Ask what happens when:

• The customer answers in a language the agent wasn't expecting
• Someone else picks up — a family member who knows nothing about the order
• The customer asks something off-script ("what's the return policy?")
• There's a TV on in the background, or a bus going past
• The customer says "call me later"
• The call goes to voicemail, or the number is switched off

Check what it writes back

A call that produces no structured outcome is worthless. Confirm that the agent writes tags or notes back to your store, that the mapping is configurable, and that outcomes are distinguishable — "confirmed," "cancelled," "reschedule requested," "no answer," "wrong number" are five very different things and must not collapse into one.

Ask about concurrency and bursts

Your orders arrive in waves. Can the system make 500 calls in the first hour of the morning? What happens during a festive spike at 5× normal volume? This is a boring question that decides whether the system works on the days that matter most.

Ask who owns the recordings and the data

Where are recordings stored, for how long, who can access them, and can you export them. This matters legally and it matters commercially when you want to leave.

Insist on a real pilot

Two weeks on real orders, with agreed success metrics defined before you start. If a vendor won't run a paid-outcome pilot on live traffic, that tells you something.

Compliance, consent and etiquette

Not legal advice — talk to your own counsel — but here's the practical landscape as it stands.

Transactional vs. promotional is the key distinction. A call about an order the customer has just placed with you is transactional: it concerns an existing transaction they initiated. Promotional calls — offers, campaigns, "we miss you" — sit under TRAI's DLT and DND regime and require registration and consent. Do not let the two blur together. A confirmation call that turns into an upsell pitch changes category, and your customer will feel it before any regulator does.

Identify yourself immediately. The first sentence should name your brand and the reason for the call. This is both good practice and good conversion — an unexplained call gets cut.

Respect the hours. Nothing before 9am or after 8pm. Late-evening calls to confirm a COD order are a good way to get cancelled.

Honour opt-outs immediately and permanently. If someone says stop calling, that must persist across campaigns, not just this one.

Disclosure. There's no blanket Indian requirement to announce that a caller is an AI, but norms are shifting and many brands choose to be upfront. Our view: don't pretend to be human if asked directly. A customer who asks "are you a person?" and gets a straight answer usually carries on with the conversation. One who is deflected and works it out anyway does not.

Recordings and data. Call recordings are personal data. Know where they live, how long you keep them, and who has access. India's data protection framework has been tightening, and "the vendor handles it" is not an answer that will age well.

The 30-day rollout plan

The plan below assumes you're starting with COD confirmation, because you should be. It's deliberately unambitious in week one — every failed rollout we've seen started too broad.

Days 1–3: measure your baseline

Before changing anything, pull 90 days of history and write down: RTO rate overall, RTO rate by courier, RTO rate by region, RTO rate by product category, and average cost per RTO. If you skip this you will never be able to prove the system worked, and someone will ask.

Days 4–7: write the script and the outcomes

Decide exactly what the call is trying to achieve and what outcomes you'll record. Keep it to one goal. For confirmation, the shape is:

Note what it doesn't do: no marketing, no upsell, no survey. One question, one confirmation of address, done in under thirty seconds. The longer script variants and the reasoning behind each line are in the Hindi confirmation script post.

Days 8–10: integrate and define the tags

Wire the agent to your store and decide the tag taxonomy before any calls go out. Minimum viable set: confirmed, cancelled, no-answer, callback-requested, wrong-number. Then decide what your fulfilment team does with each — an untagged outcome nobody acts on is worse than useless, because it creates the appearance of a process.

Days 11–14: pilot on 10% of orders

Start small and deliberately. One category or one region, 10% of COD orders. Then — and this is the part people skip — listen to every single call in this phase. All of them. You will find things no dashboard would have told you: a product name being mangled, a question customers keep asking that the script doesn't handle, a pause that's slightly too long.

Days 15–18: fix what you heard

Adjust the script, add the domain vocabulary, tune the timing. This iteration loop is where the quality actually comes from. A system that's 70% right on day 11 and tuned twice is better than one that was 80% right and left alone.

Days 19–25: scale to 50% and hold a control group

Keep a genuine holdout — the same 20–30% of orders left uncalled. This is the only way to attribute the change honestly, because RTO rates move seasonally for reasons that have nothing to do with you, and without a control you will either take credit for a good month or get blamed for a bad one.

Days 26–30: measure, decide, expand

Compare called vs. control on RTO rate, cancellation rate, and delivery rate. If the numbers hold, go to 100% and start the next use case — NDR recovery, which is where the second-largest win lives.

What to measure

Voice AI generates a lot of numbers, most of which don't matter. These do:

And one qualitative practice that no dashboard replaces: listen to ten random calls every week, forever. Not the flagged ones — random ones. Every serious operator we know does this, and every one of them has found something in month four that the metrics never surfaced.

Common mistakes

• Calling too late. A confirmation call two days after the order gets a confused customer who's forgotten ordering. Hours, not days.
• Making the call do too much. Confirmation plus feedback plus upsell in one call converts worse on all three. One goal per call.
• No holdout group. Then you can't prove anything, and the first seasonal wobble becomes an argument.
• Ignoring the write-back. Calls that produce outcomes nobody acts on are pure cost. If "cancelled" doesn't stop the shipment, why are you calling?
• Treating no-answer as a final state. It's the start of a retry sequence, at a different time of day.
• Evaluating on a scripted demo. Judge on real recordings from your regions and your categories, or you're judging the vendor's sales team.
• Deploying and walking away. Product names change, festive season shifts behaviour, courier mix changes. Voice systems need periodic attention, not a one-time setup.
• Calling everyone about everything. The fastest route to customers who dread seeing your number. Calls are for decisions and saves — everything else goes to WhatsApp.

FAQ

Will customers be annoyed by an AI calling them?

Far less than people expect, when the call has a real reason to exist. Customers dislike calls that waste their time, not calls that are automated. A 25-second call confirming an order they just placed is useful to them — it's their money on delivery too. A promotional call dressed up as a service call is the thing that annoys people, and that's a script decision, not a technology one.

What happens if the AI can't handle the conversation?

It should recognise that and exit gracefully — take a message, promise a callback, or transfer to a human if you have one. The failure mode to avoid is a system that keeps trying to steer an unhappy customer back to the script. Ask any vendor exactly what their escalation path is.

Do I need to change my Shopify setup?

Typically no beyond installing an app and mapping tags. The important prep isn't technical: it's deciding what your team does with each outcome tag once it appears.

How many languages do I actually need?

Start with Hindi and English handled as one code-mixed capability, since that covers most of the country. Add regional languages based on your actual order distribution — pull the last 90 days by state and let that decide, rather than adding languages because a vendor lists them.

Can it work for prepaid orders too?

Confirmation calls are mostly a COD phenomenon — a prepaid customer has already demonstrated intent with their money. But NDR recovery, address verification, delivery scheduling and return coordination apply equally to prepaid, and are commonly overlooked because the whole category gets mentally filed under "COD stuff."

What's the minimum volume where this makes sense?

Lower than most people assume — the affordability threshold is what changed. The more useful gate isn't volume but whether you have a measurable RTO problem and someone who will act on the outcomes. A brand with 400 monthly COD orders and a 30% RTO rate has a clearer case than one with 5,000 orders and 8% RTO.

Where to start

If you take one thing from all of the above: the post-checkout funnel is the least contested source of margin in Indian D2C, and almost all of it comes down to asking customers questions at the right moment. Voice AI is simply the tool that made that affordable at small scale.

Start with COD confirmation. Measure it against a control group. Then add NDR recovery. That sequence has paid for itself for every brand we've watched run it properly, and it's the one we'd give a friend starting tomorrow.