Most conversations about voice AI start at the enterprise end — call centers, IVR trees, ten thousand seats. That’s where the money is. It’s also where the problem is the least interesting. A call center already has staff. Voice AI there is a cost-reduction lever, not a change in kind.
The interesting end is the other one. A single-location HVAC company that misses eight calls a week because the owner is inside a wall. A medical spa whose front desk closes at 6pm — meanwhile the ready-to-book lead is scrolling at 11pm. A law office where every incoming call is potentially a case, and the person who answers is also the one writing briefs.
For these businesses, the receptionist they need isn’t expensive — it’s impossible. You can’t hire a fractional receptionist. You can’t hire someone to work 11pm–1am on Tuesdays. So the phone rings out. The lead goes to the next Google result. The revenue never shows up in the dashboard because it never existed.
Where cloned-voice agents actually shine.
I’ve shipped nine of these in the last twelve months. The pattern is clear.
- Inbound qualification, not conversation. The agent asks four questions, books the calendar, and hangs up. Every project that tried to make the AI a full-service assistant went sideways. Every project that treated it as a smart intake form with a voice shipped in three weeks and stuck.
- After-hours and overflow, not primary. Route to human first when someone’s there. Fall through to AI. Customers don’t mind the AI when the alternative is voicemail. They mind it a lot when the alternative is a person.
- Voice cloned from the owner, not a stock TTS. This is the difference between “creepy tech” and “oh, that’s Kevin.” A four-minute sample and it holds up over a two-minute call. Every dollar you spend here comes back.
- Book straight to the calendar. Every action the AI takes has to be reversible from the CRM in two clicks. If you can’t undo it, don’t automate it.
Where it doesn’t work.
Same nine projects. Same clear pattern.
- Anything emotional or legally sensitive. A divorce lawyer. A hospice line. Not the place. The AI will handle the call fine — it’s that you don’t want it to.
- Businesses with high-context repeat customers — a bespoke tailor, a family CPA. The relationship is the product. Automating the front door destroys it.
- Any workflow that requires looking something up in a system the AI can’t see. If the receptionist has to alt-tab to Jobber, ServiceTitan, or the founder’s email, the AI will hallucinate. Give it read access or give it back to a human.
“The point isn’t that the AI is a good employee. The point is that the alternative is no employee.”
What I built, concretely.
A three-part stack. Vapi handles the telephony and the LLM turn-taking. A small backend in Python holds the business logic — qualifiers, calendar checks, CRM writes. Every call is transcribed and scored on a did-they-book-yes/no axis, and the ones that failed get emailed to the founder every morning with the transcript and my guess at what to fix.
The whole thing is under 800 lines of code, deploys to a $5 VM, and costs the business about $0.14 per call. A missed call in most of these industries is worth $200–$800 in booked work. The unit economics aren’t close.
What I’m watching next.
Two things. First, latency — the current stack is ~700ms of round-trip on a good day, which is fine but not human-fine. Second, the shift toward on-device inference. When Voice AI runs on the phone itself, the economics change again, and small businesses become the biggest beneficiary a second time.
If you’re running a business that misses calls: I can build you one of these in about three weeks. Write me a sentence about what you’re trying to catch and I’ll tell you if it’s a fit.