We Built an AI That Calls Your Leads in 5 Seconds. Here's What Happened.

Last week I shipped a feature I've been thinking about for two years: a form on a landing page that, the moment you submit it, dials your phone within 5 seconds.

Not "we'll call you in 24 hours." Not "we'll send you an email." Not "a sales rep will reach out shortly." Five. Seconds. While you're still looking at the screen.

The bot's name is Angelina. She speaks English by default and switches to Hindi or Hinglish if you do. She introduces herself, mentions she's an AI (TRAI-compliant), confirms the call is being recorded, and asks if you have a quick minute. If you say yes, she runs a real discovery conversation. If you say "abhi busy hoon," she politely offers to follow up by email.

This is the wedge product behind QuotaHit, the AI sales department I'm building for Indian SMBs.

I want to tell you what's broken about the way most Indian businesses handle inbound leads today, why the 5-second response time matters more than every other lead-management feature combined, and what the early data looks like when you actually build the thing instead of just talking about it.

The lead-decay problem nobody fixes

An interior design studio in Mumbai I've been talking to gets about 80 inbound enquiries every month from Instagram ads, their website, JustDial, and word-of-mouth referrals. The leads cost them roughly ₹600 each — about ₹48,000 a month in ad spend.

The leads come in at all hours. Some during business hours, sure. But the peak is between 9 PM and midnight, when people sit in bed scrolling Instagram and decide they want to redo the living room.

Their two sales BDEs work 10 AM to 7 PM. The 9 PM lead sits in a Gmail inbox until 10:30 AM the next morning. By the time someone calls, the customer has already messaged three other studios. The intent window is gone.

Across the design studios I've talked to in the last six months, the leakage number lands somewhere between 50% and 70%. Half to two-thirds of paid leads die before a human ever speaks to them.

Same story in real estate. Same story in B2B SaaS demo signups. Same story in coaching/edtech enquiries. The Indian SMB lead-to-call cycle is broken at exactly one place: the gap between intent and contact.

What the data says about response time

This isn't a new discovery. Harvard Business Review ran the original study in 2011: contacting a lead within 5 minutes is 21 times more likely to convert than contacting within 30 minutes. The 5-minute threshold isn't magic — it's the gap before intent decays.

HubSpot reproduced it. InsideSales reproduced it. Drift built a $1B company around the chat-based version of this insight.

Yet in 2026, walk into 100 SMB offices in Ahmedabad or Mumbai and 95 of them are still calling inbound leads with 1-to-4-hour latency. The bottleneck isn't insight. The bottleneck is staffing — you cannot pay humans to sit by a phone 24 hours a day at the unit economics of a ₹40 lakh-a-year design business.

So the question I had two years ago was: what if you didn't need humans for the first contact?

Why nobody had built this before

When I started building QuotaHit's voice agent, I assumed I'd just plug into one of the existing AI calling tools. The market is loud — 11x.ai raised $50M from a16z, Artisan raised $25M, Regie.ai has Salesforce partnerships, Bolna recently raised $6.3M from General Catalyst.

Every single one of them was wrong for an Indian SMB.

11x charges around $5,000 a month (₹4.2 lakh), annual contracts only, no Hindi, no self-serve. Artisan starts at $250/month per user but doesn't do voice calls at all. Regie.ai is per-seat at $499/user/month — adds up fast for a 3-person sales team. Bolna does have Indian-language voice infrastructure, but it's the infrastructure layer; there's no SDR product on top of it. You buy Bolna to build with, not to use.

None of them were priced for an SMB founder paying ₹35,000/month for a junior BDE. None of them switched languages mid-call. None of them integrated with IndiaMart, which is where most Indian B2B leads actually originate.

The white space was right there.

What we shipped

The product is simple to describe. The implementation took three months and roughly 30,000 lines of code.

There's a form at quotahit.com. You type your first name, phone, and email. You click "Try AI Calling — call me now."

Within 5 seconds, your phone rings. The caller ID shows up. You answer.

You hear: "Hi [your name]! This is Angelina from QuotaHit calling about the demo request you just submitted. Quick heads-up — this is an AI sales call and it's being recorded for quality. Got a quick minute?"

That sentence does four things in 12 words: (1) personalizes by your name, (2) acknowledges the form submission so you don't feel ambushed, (3) discloses it's AI per TRAI compliance, (4) discloses recording per DPDPA compliance.

If you say yes, she runs a 2-3 minute discovery conversation: what's your business, how many sales people, where are your leads coming from, what's the bottleneck. If you sound interested, she offers a 15-minute demo and books it. If you sound busy, she offers an email summary.

The technical stack underneath: Twilio for telephony, OpenAI's GPT-4o-mini Realtime API for voice generation, Modal for the Python bridge that mediates audio between them, Supabase for the database, Vercel for hosting, Next.js for the frontend. All commodity infrastructure. The moat isn't the stack — it's the configuration, prompting, and integration discipline.

Early signals from the first two weeks

I ran a batch of controlled test calls in the first two weeks — a mix of my own number on different phones, friends, and team members who agreed to be guinea pigs.

A few patterns surfaced.

Pickup rate ran much higher than cold calls. Not 30% like a cold-call from a telecaller. The reasons: people had JUST submitted a form (high intent), the call came in within seconds (memory still fresh), and the call rang once or twice with no spam-filter delay.

Average conversation length sat in the 2-minute range — long enough to extract real qualification data (business type, sales team size, current pain), short enough that nobody felt trapped.

Demo-to-meeting conversion was meaningfully higher than the industry baseline. Inbound-to-meeting benchmarks land at 5-10% across most SMB SaaS reports. Early QuotaHit numbers tracked roughly 3-5x that, though the sample size is small and we'll publish hard numbers once we hit production volume.

The unexpected thing: when callers found out it was an AI, the response wasn't "you're wasting my time" — it was almost always "this is incredible, who built this?" The novelty of the 5-second response time was more memorable than the AI-ness of the voice.

What's still broken

I want to be honest about what doesn't work yet.

Hindi recognition is rougher than English. The bot understands code-switching at the sentence level ("yes haan tell me more"), but a fluent Hindi sentence with regional vocabulary still hits a meaningful misrecognition rate. We're integrating Sarvam AI's Saaras speech-to-text next, which is purpose-built for Indian languages. Until that ships, the bot performs much better in English-default conversations.

Latency is real. Per-turn response latency runs in the 1-2 second range on Indian networks. Under 500ms feels human; over 1 second feels like a slightly laggy international call. We're moving the Python bridge from a US Modal region to Mumbai/Singapore to halve this.

Outbound from a US number is not ideal. Twilio doesn't sell India outbound DIDs anymore, so our caller ID is currently a US-prefix number. It still rings Indian phones, but Truecaller occasionally flags it as "international call." For production volume past 50 calls a day per customer, we're migrating to Plivo's Indian 140-series number infrastructure and registering with DLT (the Indian telecom Distributed Ledger Technology that whitelists legitimate business callers).

The conversation can't yet handle complex objection trees. It's great at "tell me about your business" and "want to book a 15-min demo?" It's less great at "we already have a CRM we love, why would we switch?" The system prompt is being tuned with a library of 20 India-specific objection responses in the next iteration.

Why I'm building this in public

Most AI startups stay quiet about their architecture until launch. I'm doing the opposite — every dispatch is documented, every stable tag is on GitHub, every research finding lives in a public-readable document.

Two reasons.

One, accountability. If I tell you on May 17 that the next 7 days ship Stripe + the dashboard billing page + the new pricing, you can hold me to it. The blog post and the GitHub commits don't lie.

Two, the moat for an Indian AI SDR is not the calling infrastructure — Bolna and Sarvam will commoditize that layer within 12 months. The moat is product, workflow, integrations, and trust. Trust comes from building in public. Workflow comes from talking to real SMB founders every week. Product comes from shipping every day.

If you're a founder in Mumbai or Ahmedabad running a real estate, interior design, or B2B SaaS business losing leads to slow follow-up, type your phone on quotahit.com. Angelina will call you in 5 seconds. You'll know in 60 seconds whether this thing is real.

That's the whole pitch. The math takes care of itself once you've heard her.