Every week, small businesses lose jobs for the dumbest reason: nobody picked up the phone. The customer doesn't leave a voicemail — they call the next listing on Google.
Hiring a full-time receptionist costs $3,000+/month. An AI receptionist costs a fraction of that and never sleeps, never takes lunch, and never lets a call go to voicemail. Here's how I build them for my clients at TechMitten.
What the client actually gets
- Every call answered instantly, 24/7 — nights, weekends, holidays
- A natural-sounding voice agent that greets callers, answers common questions, and captures lead details (name, number, what they need)
- The moment a call ends, the business owner gets a text with a short summary, plus a detailed email with the full breakdown
- No app to install, no dashboard to learn — it works through the phone number they already have
The stack
RetellAI for the voice agent. This is the core — it handles speech-to-text, the LLM conversation, and text-to-speech with low enough latency that callers don't feel like they're talking to a robot. Voice quality is what makes or breaks the whole thing; a long pause and the caller hangs up.
Twilio for telephony. The client's business number routes to the AI agent. Nothing changes for the customer — they dial the same number they've always dialed.
A scheduled notification pipeline for lead alerts. After each call ends, the system waits for the call analysis to complete, then fires two notifications: a short SMS summary to the owner's phone and a detailed email with the call breakdown. I run this check every 15 minutes, so nothing slips through.
Building one for a client: the actual process
1. Conversation design. This is most of the work, and it's not code — it's writing. I map out: how should the agent greet callers? What are the 10 most common questions this business gets? What info must be captured from every caller? The agent gets a personality brief, not just a script.
2. Business knowledge. I feed the agent the details that make it sound like an insider: services and pricing, hours, service area, how to handle pricing questions versus "I need to talk to a human" situations. A generic agent is useless; a knowledgeable one books jobs.
3. Call flows and fallbacks. What happens when the caller asks something unexpected? The agent takes a message rather than inventing an answer. What happens when someone is angry? De-escalate and offer a callback. These edge cases are where cheap implementations fall apart.
4. Test calls. I call the number myself — repeatedly, with different scenarios. The friendly caller, the confused caller, the one who rambles. If the agent can't handle my test calls, it's not ready for real customers.
5. The notification loop. The owner needs to know about leads now, not tomorrow. The SMS is short enough to read at a glance ("New lead: Sarah M., wants a quote for a bathroom remodel"). The email has the full details. Speed of follow-up is what turns a lead into a job.
What I've learned
Latency is everything. The difference between a fast response and a slow one is the difference between "wow, that was smooth" and "this is a robot." A streaming voice pipeline is non-negotiable.
Owners don't want a dashboard. Every client I've talked to wants the same thing: "just text me when someone calls." The fancier the admin panel, the less they use it. SMS plus email beats any dashboard.
The agent should know when to stop talking. The best AI receptionists are brief. Greet, help, capture, confirm, done. Callers can smell a rambling bot instantly.
Test with real-world messiness. Accents, background noise, speakerphone — the real world isn't a quiet room. I test accordingly.
The bottom line
A missed call used to mean a missed job. Now it means the AI took the details, the owner got a text moments later, and the lead got a callback within the hour. That's the whole pitch — and it's why every client I've set one up for wonders why they didn't do it sooner.
I build these for small businesses at TechMitten — if missed calls are costing you jobs, that's literally the problem I solve.