Why I Built Clara Instead of Hiring a Receptionist

I work 12-hour shifts in the NICU. When a call comes in at 2pm on a Tuesday, I'm elbow-deep in someone's airway. I can't pick up. And the people calling my business don't leave voicemails — they just call the next place on the list.

That's the real problem. Not "how do I scale customer service." Just: the phone rings, I can't answer it, and the lead is gone.

I looked at hiring a receptionist. A live answering service runs $200-400/month. A dedicated receptionist is $3,000+/month plus benefits. Neither of those works when you're a solo founder with zero revenue. So I built something instead.

The First Version Was Bad

I won't pretend it was good from day one. The first version of Clara used a basic Twilio webhook, sent audio to Whisper for transcription, then sent the text to an LLM, then sent the response back through a TTS engine. The latency was 4-6 seconds. Nobody wants to wait 6 seconds for someone to respond on a phone call.

It also hallucinated. Someone would ask "Do you have parking?" and Clara would say "Yes, we have free parking in our underground garage." We don't have a garage. We don't even have an office.

That's when I realized the hard part isn't the voice. It's the knowledge. You can't just throw an LLM at a phone call and hope it knows your business.

What Actually Works

The version that works looks different:

• Grounding context. Clara gets a knowledge base about the business — hours, services, pricing, FAQs. Every answer is grounded in that context. If it's not in the knowledge base, Clara says she doesn't know.
• Fast inference. I moved to Groq for LLM inference. Sub-200ms response times. The total round-trip from caller speaking to Clara responding is under 1 second. That's the threshold where it feels like a conversation instead of a wait.
• Escalation, not replacement. Clara handles the calls that were already going to voicemail. If someone needs something she can't handle, she takes a message and texts me. I'm not trying to replace a human receptionist — I'm catching the calls that were already falling through the cracks.

The "I Don't Know" Problem

This took me the longest to get right. Most AI phone agents either hallucinate when they don't know something, or they say "I don't know" to everything because their confidence threshold is too high.

The fix was a two-step approach: first, try to answer from the knowledge base. If the confidence is low or the question is clearly outside scope, say "I don't have that information, but I can have someone call you back." Then actually call them back.

Honesty is cheaper than hallucination. A wrong answer costs you a customer. "I don't know, but I'll find out" builds trust.

What I'd Tell Someone Building the Same Thing

Start with the knowledge base, not the voice. Spend more time on what your agent knows than how it sounds. A fast, natural-sounding agent that gives wrong answers is worse than a slightly robotic one that's always right.

Also: test with real calls. Not your own test calls — real people calling about real things. The first 20 calls will surface problems you didn't think of. "Can I bring my dog?" "Do you take insurance?" "What's your cancellation policy?" — stuff that's obvious in hindsight but invisible until someone asks.

And don't try to handle everything. Clara's job is to answer the easy 70% of calls and escalate the rest. That's a win. Before Clara, 100% of after-hours calls went to voicemail. Now 70% get answered. That's not replacing a human — it's catching what was already being dropped.

Where It Is Now

Clara's running. It answers calls, takes messages, and texts me when something needs my attention. I'm using it for my own business first — eating my own dog food before taking it to anyone else.

We're pricing it at $149/month. That's less than half what a live answering service costs, and it never sleeps, takes breaks, or goes on vacation. But I'm not going to pretend it's perfect. It still stumbles on accent-heavy callers. It still can't handle multi-party conversations. It's a tool, not a person.

The goal isn't "AI that sounds exactly like a human." The goal is: the phone gets answered, the caller gets what they need, and I don't lose leads because I was in the middle of a procedure.

That's it. That's why I built Clara. Not because I wanted to build an AI receptionist — because I needed one.

← Back to blog