Build a Hindi-First Sales Bot for a Tier-2 Auto Dealership

A car dealership in a Tier-2 town in Madhya Pradesh got 60 WhatsApp enquiries a day, most in Hindi or Hinglish, most after the showroom closed. Their two salespeople answered the morning's backlog and lost the rest. We built a Hindi-first WhatsApp bot over a weekend that books test drives into the showroom calendar, answers EMI and on-road-price questions, and hands genuinely hot leads to a salesperson with full context. In the first month it captured 71% of after-hours enquiries that used to go cold. This is the build — NLU, the booking flow, the EMI logic, and the human handoff. ## The answer in 55 words A WhatsApp Business bot that understands Hindi and Hinglish, runs four intents — test-drive booking, EMI and price questions, model comparison, and a generic enquiry — and routes hot leads to a salesperson. Hindi understanding comes from Sarvam's Indian-language speech-to-text plus Claude for the reply. Bookings write to the showroom calendar; every lead lands in the CRM. Build time: one weekend. Cost: about ₹0.20 per conversation. ## Why Hindi-first, not English with Hindi bolted on Most "multilingual" bots are English bots that translate. They stumble on Hinglish — "bhaiya on-road price kya hai Creta ka, EMI kitni banegi 5 saal me" — because translation flattens the code-mix and loses intent. A Hindi-first design keeps the original message intact and uses a model trained on Indian languages and accents. Sarvam's Saaras V3 covers 22 Indian languages plus English, was trained on over a million hours of Indian audio, and handles code-mix natively — and it processes and stores data in India, which matters for a dealership nervous about customer data. We tune the lexicon for car terms the same way we did for our in-house English app TalkDrill's Indic voice pipeline. ## Why this matters now (June 2025) Two things changed the math. First, the WhatsApp Business Cloud API made it cheap for a small dealership to run a bot on its own number without a heavy BSP contract. Second, Indian-language understanding got genuinely good: Sarvam's Saaras V3 reports a word error rate around 19% on the IndicVoices benchmark's top-10 languages, low enough that a code-mixed voice note about an EMI gets understood. A weekend bot now does what needed a call-centre two years ago. ## The four intents the bot actually handles ## The DIY walkthrough — build it in a weekend We run this on the WhatsApp Business Cloud API, Sarvam for voice-note transcription, Claude Haiku 4.5 for intent and reply, and a small Postgres for state and the lead log. As Hrishikesh, our CTO, says about Indic bots: get the intent router and the handoff right first; the polish comes after. ### Step 1 — receive the message and handle voice notes WhatsApp webhooks deliver text and voice. For a voice note, fetch the audio and send it to Sarvam's speech-to-text. We ask for verbatim transcription so a mixed Hindi-English message comes back faithfully rather than translated. ### Step 2 — classify intent with a strict-JSON prompt Send the transcript (or the typed text) to Claude with a prompt that returns only JSON. We keep the original Hindi or Hinglish in the message field so nothing is lost in translation. ### Step 3 — the EMI calculation (do the math in code, not the model) Never let the model compute the EMI — it will occasionally get arithmetic wrong, and a wrong EMI is a trust-killer. Compute it in code from the standard formula and let the model only phrase the answer. ### Step 4 — book the test drive into the showroom calendar When intent is test-drive and you have model, day, and slot, check the calendar for a free slot and write the booking. Confirm in the customer's language and store the lead. ### Step 5 — score the lead and write it to the CRM Every conversation creates or updates a CRM lead with the transcript, the detected model, and a score. Hot leads (score above 70, or any explicit ask for a quote or visit) trigger a salesperson alert. ### Step 6 — the human handoff When needs_human is true, the bot tells the customer a salesperson will call, collects a good callback time, and pings the on-duty salesperson with the full transcript and the lead score. The salesperson can reply in the same WhatsApp thread — the bot steps back once a human joins. ## Common mistakes — when a Hindi bot goes wrong Mistake 1 — letting the model quote firm prices. On-road price varies by RTO, insurance, and offers. The bot should give the dealership-supplied figure and call anything it computes indicative. We hard-rule the prompt against inventing numbers and pull prices from a sheet the dealership owns. Mistake 2 — translating instead of staying in-language. Replying in English to a Hindi message feels cold and loses customers. The prompt forces same-language replies. For voice notes, transcribe verbatim rather than auto-translating to English first. Mistake 3 — no human handoff path. A bot that can't hand off frustrates a ready-to-buy customer. The single highest-value rule is: any explicit ask for a quote, finance, or a visit goes straight to a salesperson with context. The bot's job is to qualify and route, not to close. Mistake 4 — ignoring the dealership's actual vocabulary. Customers say "kitna padega," "on-road," "exchange me kitna," "down payment." Seed the prompt with the real phrases and the model maps them to intents far more reliably. We sampled a week of real chats before writing the prompt. ## Real example — the metric that surprised the owner The dealership took about 60 WhatsApp enquiries a day, two salespeople, most messages after hours. Before the bot, after-hours enquiries mostly went cold. After the weekend build: 71% of after-hours enquiries got a useful reply and a logged lead, and confirmed test-drive bookings ran about 3.4x the previous month. The owner expected the booking lift. The surprise was the EMI intent — nearly 40% of conversations were really finance questions in disguise, "kitni EMI banegi." Once the bot answered those instantly and indicatively, then offered a callback, the finance-desk's qualified-lead count jumped. The owner moved a salesperson partly onto finance follow-ups because the bot revealed where the real demand was. This is the same conversational-AI discipline our AI and automation team brings to every bot build. For the English-stack version of a Hindi bot, see our 2025 tutorial on a Hindi-first chatbot with OpenAI and Sarvam ASR, and for a voice-bot in a different vertical, a Hindi voice bot for a Tier-2 insurance seller. The proof-of-work is the live voice pipeline behind TalkDrill.

• You take ≥30 WhatsApp enquiries a day, mostly Hindi or Hinglish
• A real share of enquiries arrive after the showroom closes
• You run on the WhatsApp Business Cloud API on your own number
• You use an Indian-language model (Sarvam) for voice and code-mix
• You compute EMI in code and label it indicative, never model-generated
• You pull prices and specs from a sheet the business controls
• You force same-language replies, no silent translation to English
• You hand off any quote, finance, or visit intent to a salesperson with context
• You capture consent, keep an opt-out, and store leads in your own CRM

## FAQ ### Will the bot understand thick regional Hindi and code-mixing? Largely, yes. Sarvam's Saaras V3 is trained on over a million hours of Indian audio across accents and handles code-mix natively, reporting a word error rate near 19% on the IndicVoices top-10 languages. It won't be perfect on heavy dialect, which is exactly why a human handoff path matters for the messages it can't parse confidently. ### Why use Sarvam instead of a global speech-to-text service? Two reasons. Sarvam is purpose-built for Indian languages and Hinglish, so it transcribes code-mixed car-buying chat more faithfully than a general model. And it processes and stores data in India, which is an easier conversation with a dealership cautious about customer data than exporting it abroad. ### How do you stop the bot from quoting a wrong EMI or price? Prices come from a dealership-controlled sheet, and the prompt forbids inventing numbers. The EMI is computed in code from the standard formula, not by the model, and every figure is labelled indicative with an offer of a firm callback. The model only phrases the answer; the math lives in code. ### Can a salesperson take over the same WhatsApp chat? Yes. When the bot hands off, it pings the on-duty salesperson with the transcript and lead score, and the salesperson replies in the same thread. The bot detects a human has joined and steps back so the customer isn't talking to two voices at once. ### What does it cost to run per month? Per handled conversation it is roughly ₹0.20 in model and transcription cost. WhatsApp conversation charges and a small server add to that. For a dealership at 60 enquiries a day, total run cost lands in the low thousands of rupees a month — well under the value of the leads it stops from going cold. ### Can we add more languages like Marathi or Tamil later? Yes. Sarvam covers 22 Indian languages, so adding Marathi or Tamil is mostly a matter of widening the language detection and translating the dealership data sheet. The intent router, EMI logic, and handoff stay the same. Add a language only once you see real demand for it in your chat logs.