May 24, 2026 · 5 min read
What it actually costs to run an AI receptionist, and what I learned setting one up
Real numbers from setting up the Runari AI receptionist last week. What the per-call math looks like for a small business, what the vendor marketing oversells, and the four questions I'd ask before picking one.
The number first
A small business that misses 20 calls a month after hours can have an AI receptionist take every one of them for under $20 a month, all-in. A busier one, say 100 calls a month, runs around $50 to $100 a month.
For comparison: a human answering service for the same volume runs $200 to $500 a month. A part-time receptionist runs $1,500 to $3,000.
That's the kill shot of this whole category. The math is so favorable that the only honest questions left are whether you want one and whether the vendor you pick actually delivers. I'll get to both, but I want to start with the math because it's the part nobody publishes plainly.
What I actually did
Last week I set up an AI receptionist for Runari, the practice you're reading. I used Retell. Total setup time was about two hours, most of which was writing the system prompt and configuring the post-call data I wanted captured.
The result: when someone calls (877) 812-3360, an AI receptionist answers, captures the caller's name, email, and what they're calling about, takes a message if they want to leave one, and then drops a notification into my phone with the structured summary. The full transcript and the AI's interpretation land in my database. I follow up within a business day.
The whole thing (voice agent, phone number, post-call analysis, webhook, database, notification) cost me roughly $8 in the first week of testing, plus a $5/month toll-free number.
If I get one decent lead per quarter from it, the math is already done.
The actual breakdown
Here's what the line items look like, plainly:
| Item | What it costs |
|---|---|
| Retell per-minute rate | ~$0.07–$0.15/min (varies by voice + model) |
| Phone number (toll-free) | ~$5/month |
| Phone number (local) | ~$1–$2/month |
| LLM (built into per-minute) | included |
| Webhook + database + notifications | $0; runs on existing infra |
Plug in real volumes:
| Calls per month | Approx. minutes | Approx. cost |
|---|---|---|
| 10 calls × 3 min avg | 30 | $3–$5 + phone |
| 50 calls × 3 min avg | 150 | $15–$25 + phone |
| 200 calls × 3 min avg | 600 | $60–$90 + phone |
A typical local service business, plumbers, electricians, dentists, accountants, contractors, misses maybe 20 to 50 calls a month outside business hours. Putting an AI on the line for $15 to $30 a month is, frankly, an absurdly good trade. Even one captured lead a month makes it self-funding.
What the marketing oversells
I've now looked at three of these vendors in detail (Retell, VAPI, Bland) plus a handful of the "no-code AI agent" tools that show up in newsletter ads. A few things to be honest about:
The voice matters more than the model. Vendors love to talk about which LLM is behind their agent. In practice, what callers experience is the voice: its cadence, its naturalness, the latency between their words and the AI's response. Every vendor lets you pick from the same pool of voice providers (11labs, OpenAI, etc.). The LLM choice mostly affects how often the AI says something slightly wrong, not how the call feels.
The "AI agent that does everything" pitch is a lie. These tools are excellent at narrow, well-defined jobs: taking a message, qualifying a lead, answering basic questions you've trained them on, scheduling an appointment. They are not good at open-ended customer service, troubleshooting that requires looking up an account, or anything that touches a transactional system you haven't carefully wired up. The demos always show the polished version. Your first agent will not sound like the demo.
Post-call analysis is the actually-important feature. Most vendors advertise the live conversation. The thing that makes the difference between "voicemail with extra steps" and "an actual lead capture system" is what the vendor does after the call. Can it extract structured fields (name, email, intent) reliably? Does it send them to your system in a format you can act on? Retell does this well; some don't. This is the feature I'd evaluate first if I were picking again.
"Production-ready" means handling the boring failure cases. Webhook retries when your endpoint is down. Signature verification so you know the webhook is real. Graceful fallback when the AI doesn't understand. Most of the configuration time isn't writing the prompt. It's making sure the plumbing won't drop a call. If you're evaluating a vendor, ask them to show you the failure-mode docs, not the happy-path demo.
When this makes sense for an SMB
Good fit:
- You miss inbound calls regularly: after hours, during peak season, in the middle of a job.
- Most of your inbound is predictable. People asking about pricing, hours, services, scheduling, or wanting to leave a message.
- You can clearly say "if a stranger called this number, here's what I'd want them to walk away with."
Bad fit, in my opinion:
- Heavy customer service or complex troubleshooting.
- Regulated industries: legal, medical, financial, where an AI giving wrong advice creates real liability.
- Anything requiring account lookups, transactions, or sensitive customer data flowing back and forth.
- You only get one or two calls a month. The setup time isn't worth it.
The honest middle case is most of what I see: small businesses with too many missed calls to ignore but not enough to justify a real receptionist. That's the sweet spot for this tooling, and it's the gap the vendors are competing to fill.
A note on what to ask vendors
If you're picking one yourself, here's what I'd actually evaluate against, in priority order:
- How clean is the post-call data? Get a vendor to walk you through what comes out of the webhook for a real test call. If you can't get structured fields out, you're paying for an answering machine.
- What does the per-minute rate include? Some bundle LLM costs, some don't. Some charge extra for the premium voices that actually sound good.
- Can you change the agent without redeploying? You will tune the prompt every week for the first month. If the vendor makes that painful, you won't tune it, and the agent will stay bad.
- What are the failure modes? Specifically: what happens if your webhook is down for an hour? Do they retry? Do they queue? Or do those calls just vanish?
If you want me to actually do this evaluation for your specific situation, that's the Audit: $1,500 flat, written recommendation, real-cost projection, integration-risk callout. Otherwise the questions above will get you most of the way on your own.
If you've actually set one of these up and saw something I missed, I'd like to hear it.