01. Why a Voice Agent, and Why Gujarati
Our audience is small business owners across Gujarat. Many of them prefer a phone call to typing, and most prefer to speak Gujarati. A lead that waits a day for a callback has usually moved on. The goal was simple: call every serious lead within minutes, in their language, and hand the real conversation to a person.
The agent never sells, quotes custom prices or sends payment links. Its only job is to understand the request and book time with a human.
02. The Call Pipeline
The agent runs on a conversational voice platform connected to an Indian telephony provider. The same number is also our WhatsApp business number, so the lead sees one identity across chat and call.
Lead form completed
When a lead finishes the WhatsApp form, a WhatsApp message tells them who will call and offers a "WhatsApp only" button if they would rather not get a call.
Lead context passed in
Form answers, a summary of the chat and any previous call outcome are handed to the agent before it dials, so it never has to look anything up mid-call.
Conversation in the caller's language
Gujarati-first speech with switching to Hindi or English, a pronunciation dictionary for our brand and place names, and a short, focused script.
Demo slot and confirmation
The agent books a free slot on the founder's calendar, sends a WhatsApp confirmation, and a post-call webhook records the outcome against the lead.
03. Making Gujarati Sound Right
- Model choice: Only some text-to-speech models support Gujarati at all. We picked the fast model that does, and rejected faster ones that could not speak the language.
- Pronunciation dictionary: Brand and place names were read wrongly until we added a dictionary that tells the voice exactly how to say them in Gujarati script.
- Language switching: A short Hindi phrase like "हाँ, बोलो ना" flipped an entire call to Hindi. Switching now needs two full sentences in the new language.
- Hearing is the weak link: Real-time Gujarati speech recognition sometimes mishears or cuts a sentence. Instead of guessing, the agent politely asks the caller to repeat.
04. Chasing Latency on a Phone Line
A pause of four seconds feels broken on a phone. Our first test calls had exactly that. Measuring each stage showed the model itself was fast; the delay came from what we had put around it.
| Change | Before | After |
|---|---|---|
| Safety checks moved out of the blocking path | 4 to 5 s per reply | about 2 s |
| Knowledge search switched off for short calls | 3.4 s median turn | 1.7 to 2.5 s |
| Lead details passed in before dialling | 7 s first reply | near instant |
In voice AI, every extra step between hearing and speaking is felt by the caller. Do the lookups before the call, not during it.
05. Staying Inside Indian Calling Rules
India regulates commercial calls closely, and an automated caller has to be stricter than a person. We designed the agent around the rules from day one rather than fixing it later.
- Customer-triggered only: The agent calls people who just asked for information, quickly, and only about what they asked.
- Clear disclosure: The first sentence says it is an automated call and that it is recorded.
- Easy opt-out: One tap or one sentence stops all future calls to that number.
- Sensible limits: Daytime hours only and a daily cap on the number of calls.
06. The Small Manners of a Phone Call
Early calls ended with the agent saying goodbye twice and then waiting for the caller to hang up. Another call promised a WhatsApp message that was never sent. Both are small, and both make an AI feel careless. Now a goodbye, the outcome and the hang-up happen together in one turn, and the agent is never allowed to promise a message without sending it.
The result is a caller that is quick, polite and honest about being an AI, and that hands every real conversation to a person. We are now tuning it on live calls, one transcript at a time.