A bakery owner in Indore picks up the call with dough on his hands. “Haan bolo. Whipping cream do carton, cocoa chhe packet, aur woh naya wala chocolate compound bhi dekh lena.” A pressure cooker whistles behind him. In twelve words he has used three languages, two units, an English product name and a vague request. A human rep handles this without thinking. Many voice agents do not.
An AI voice agent in Hindi is judged less by how Hindi it sounds and more by whether it hears mixed speech correctly and gets numbers, units and prices exactly right.
We have spent a lot of time on calls like this one. Below is what we have found works, what tends to fail and how to test any agent, ours or anyone else’s, before it talks to your customers.
Indian sales calls are mixed by default, and the agent must expect it
Very few business calls in north and west India stay in one language. Product names are English. Quantities are often Hindi. Pleasantries may be Marathi or Gujarati. A voice agent built for pure Hindi, or for pure English with a Hindi option, will stumble on the most ordinary sentence.
What works is speech recognition trained on Indian speech as it is actually spoken, including code-switching mid-sentence. What fails is a system that detects “Hindi” once at the start and then forces every word through a Hindi model, or the reverse.
The agent should mirror the caller, not flip on a single word
A common failure is over-eager switching. The caller says one Hindi place name in an otherwise English sentence and the agent abruptly replies in Hindi. Or the caller tries a few English words to be polite and the agent switches to formal English for the rest of the call. Both feel robotic.
What works is mirroring: answer in the language the caller is mainly using, follow them if they genuinely switch and do not react to a single proper noun. What also fails is shuddh, textbook Hindi. A shopkeeper says “stock”, “scheme” and “delivery”, not their Sanskritised equivalents. An agent that “translates” these sounds like a government announcement.
Numbers and units are where Hindi calls go wrong
Hindi numbers are a known trap. “Dedh” is one and a half, “dhai” is two and a half, “saadhe teen sau” is three hundred and fifty. Callers mix them with English digits in the same breath. Units add a second layer: carton, packet, box, kilo, piece.
Here is an illustrative example, not a customer figure. Say a carton of cream holds 12 packs and costs ₹1,200, while a single pack costs ₹110. The baker orders “do carton”. The correct value is 2 × ₹1,200 = ₹2,400. If the agent hears “do” correctly but attaches the pack price, it quotes 2 × ₹110 = ₹220. If it gets the unit right but mishears the count as “dus”, it books 10 × ₹1,200 = ₹12,000. Either error is invisible on the call and expensive at delivery.
A warm Hindi voice that quotes a pack price for a carton does more damage than a flat voice that gets the number right.
What works is an agent that answers only from your price list, knows the price by packing and zone, checks the minimum order and reads the order back before ending the call. Our guide on AI order-taking calls to retailers walks through that flow.
Real phone lines are noisy, and silence is a signal too
Calls to shops and kitchens happen over traffic, fans, customers and televisions. An agent that treats a cough as a turn will interrupt itself. An agent that waits forever after silence leaves dead air, and the caller hangs up.
What works is ignoring background noise, checking in gently after a silence and having a graceful line ready if the network drops. What fails is any agent tested only in a quiet office on good Wi-Fi.
Trust on a Hindi call depends on what the agent will not do
Callers who are comfortable in Hindi are often quick to ask direct questions: “Aap insaan ho?” or “Discount kitna doge?” The agent’s answer to both decides whether the customer trusts the next call. It should never claim to be human when seriously asked. It should never invent a discount, price or promise it was not given. Unknowns should go to a callback or a person.
Ask too how “mujhe call mat karna” is handled. An unhappy request not to be called again should go on a do-not-call list before the call ends, whatever language it was said in.
The case for keeping Hindi calls human
A fair objection: Hindi and Hinglish calls carry warmth, humour and local reference that a machine will not fully match. A long-standing relationship between a rep and a shop owner is worth a lot, and some customers will prefer a person. Our comparison of AI calling and human telecalling takes this seriously.
We agree for relationship calls, complaints, credit discussions and negotiation. Where an agent earns its place is in routine order-taking when the rep is away, after-hours enquiries that would otherwise ring out and the hundreds of calls a team cannot make at once. In those cases the choice is not agent versus rep. It is agent versus nobody.
How our agent handles Hindi calls
AI Calling speaks English, Hindi, Hinglish and Marathi, mirrors the caller mid-call and does not switch just because a city name was Hindi. Speech recognition is tuned for mixed Indian speech, and the model answers only from your price list, minimums and schemes. Every call ends with a transcript, a summary and a score for price accuracy and script. The AI Calling guides cover setup and scripts.
What to do this week
- Record ten real order calls from your team, with permission, and note every language switch, number and unit.
- Write a test script full of “dedh”, “dhai”, cartons, packets and English product names.
- Run that script against any agent you are evaluating, from a noisy room on a mobile network.
- Ask each vendor to show the transcript and the order value for every test call, not just the audio.
A Hindi voice agent is ready when a busy shopkeeper forgets to notice it, and the order on your screen matches what he said.