AI Voice Agent

The opportunity

This was late 2022, right as large language models were breaking into the mainstream. Nobody asked for this one. I kept seeing Indian SMBs lose leads to missed calls and slow replies, and I wanted to know if an AI voice agent could answer the first call for them: handle inbound inquiries, schedule appointments, qualify leads, and answer FAQs from the business's own knowledge base.

It was an in house bet at Mighty Minds, built to test one real question. Would small businesses actually pay for this?

What I built

I scoped the product, ran competitive research, and built a working prototype with a technical partner.

It used ElevenLabs for natural voice synthesis and an LLM for the conversation, so the agent could hold a real back and forth, answer from a knowledge base, and capture a lead at the end. Not a slide deck, a thing you could call and talk to.

The validation

Then I did the part most prototypes skip: I took it to the market before scaling it. I reached out to 15 to 20 potential customers and ran demand interviews.

The signal was consistent and useful: interesting, but would not pay right now. Three reasons stood out.

Trust was low. The top objection was, what if the AI says something wrong?

The price did not clear the bar. At Rs 15 to 25K a month it cost more than the receptionist it would replace, around Rs 8 to 10K.

Wrong segment. The SMBs I targeted did not have the call volume to justify automating it.

The decision

So I stopped it. Not because the build failed, it worked, but because the market was not ready and the economics did not close.

In late 2022 most small businesses had not yet seen a tool like this, so trust had to be earned before money would follow. The sales cycle would have been three to six months of education per customer, with no runway to fund that, and the time was better spent on proven revenue.

Killing a product you built is hard. Doing it on evidence, early, before sinking months into a market that is not ready, is the job.

What I would do differently

The idea was not wrong, the targeting was. With another run I would target high volume businesses where the ROI is obvious: large clinics, call centers, e-commerce support.

I would validate willingness to pay before building, not after. And I would start simpler, an AI receptionist that just takes messages, and earn trust before automating decisions.

Built at the end of 2022, this was early rather than wrong. AI voice agents are a real category today. The discipline back then was seeing that being right about where things were heading did not make the unit economics work yet.