# AI Receptionist for Small Business: The 2026 Buying Guide (Real Costs, Integrations and a Test Script)

Author: Kris Hubbard
Published: October 5, 2026
Updated: October 5, 2026
Canonical: https://wrkless.com/blog/ai-receptionist-small-business-buying-guide

**Short answer:** For most small businesses, an AI receptionist costs $0.07 to $0.31 per minute on [Retell](https://www.retellai.com/pricing) or $0.12 to $0.14 per minute on [Bland](https://www.bland.ai/pricing) if you build on a voice platform, or about $1.67 to $3.00 per call on a managed service like [Smith.ai](https://www.smith.ai/pricing/ai-receptionist). A live human service like [Ruby](https://www.ruby.com/pricing/) starts at $250 a month for 50 minutes. The per-minute rate is the smallest part of the decision. What decides whether it pays off is how well it books into your calendar, writes to your CRM, hands off to a person, and follows up by text. Test those four things before you sign anything.

We build and test AI receptionists at Wrk Less, mostly on Retell with n8n behind it. This guide covers what we check before we recommend a platform to anyone. Prices below come from each vendor's public pricing page, checked on October 5, 2026.

## What does an AI receptionist actually cost per call?

Vendors price in three different ways, so sticker prices don't compare directly.

| Option | Pricing model (public, Oct 5, 2026) | What's not in the headline number |
|---|---|---|
| [Retell AI](https://www.retellai.com/pricing) | Pay as you go, $0.07–$0.31/min depending on model and voice. Their calculator's default build comes to $0.11/min | Telephony ($0.015/min on Retell's US Twilio numbers), $2/month per number, add-ons like knowledge base (+$0.005/min) |
| [Vapi](https://vapi.ai/pricing) | $0.05/min hosting, plus speech-to-text, LLM and voice at provider cost | Telephony is billed by the provider. The Pro success package has a $999/month minimum |
| [Bland AI](https://www.bland.ai/pricing) | $0.14/min (Start, no platform fee) or $0.12/min plus $299/month (Build) | Transferred minutes billed at $0.04–$0.05/min. Warm transfers and the in-call SMS node are listed for Enterprise only |
| [Synthflow](https://synthflow.ai/pricing) | Enterprise contracts start at $30,000/year | Scoped per contract |
| [Smith.ai AI Receptionist](https://www.smith.ai/pricing/ai-receptionist) | Per call: free for 25 calls, then $3.00/call. Pro is $150/month for 75 calls, Enterprise $500/month for 300 calls | Overage rates of $2.10–$2.50 per extra call on paid plans |
| [Ruby](https://www.ruby.com/pricing/) (live humans) | $250/month for 50 minutes up to $1,725/month for 500 minutes | Larger volumes are quoted by sales |

Here's a worked example. Say you get 300 calls a month at an average of 3 minutes. That's 900 minutes. At Retell's default configuration plus Retell telephony (about $0.125/min), usage comes to roughly $113 a month. On Bland's Start plan it's about $126 before telephony. Smith.ai's Enterprise tier covers those 300 calls for $500 a month, setup included. Ruby's largest published plan stops at 500 minutes.

Read the billing details too. Retell bills for silence and hold time because the transcriber keeps listening. Its agent fee stops when a call is transferred, but telephony keeps running. Bland charges a separate rate for transferred minutes. Per-call pricing like Smith.ai's rewards short calls, and Smith.ai says spam calls count toward your quota unless you flag them (they offer a 10% spam allowance). Price your real call lengths under each model.

**The honest math:** at small-business volume, usage costs less than one missed job a month. The real cost is the build: the integrations, the call flow, the testing and the upkeep. That's where cheap setups break.

## Which integrations actually matter?

A voice agent that only talks is just a nicer voicemail. We scope these five connections on every [AI receptionist](https://wrkless.com/ai-receptionist-for-small-businesses.php) build:

1. **Live calendar availability.** The agent should check real open slots and book directly. Texting a booking link is a weaker fallback. Ask the vendor whether the calendar is checked during the call or only after it.
2. **CRM write-back.** Every call should create or update a contact with a summary, a disposition (booked, callback, spam, existing customer) and the recording link. If your team still has to listen to recordings to find out what happened, the system isn't finished.
3. **Call transfer with context.** A warm transfer should tell the person picking up who's calling and why before the caller is connected. Find out what happens when nobody answers the transfer. The caller should fall back to a captured callback, not dead air.
4. **SMS follow-up.** A confirmation text after booking and a "sorry we missed you" text after a dropped call recover a lot of leads. They're also where consent rules apply. The FCC just changed how opt-outs work, so read [our breakdown of the new opt-out rules](https://wrkless.com/blog/fcc-opt-out-rules-2026) before you turn on any automated texting.
5. **Post-call webhooks.** This is the escape hatch. If the platform sends a structured payload when a call ends, you can route it anywhere: your CRM, Slack, a job board, a spreadsheet. We usually send it to n8n. Here's [how we decide where that logic lives](https://wrkless.com/blog/retell-workflows-vs-n8n-zapier) and [how Retell webhooks hand off to a CRM](https://wrkless.com/retell-ai-integrations-and-webhooks.php).

If you're only trying to stop losing unanswered calls, start smaller. A narrow [missed-call capture](https://wrkless.com/missed-call-capture.php) flow that collects name, number, service and timing and creates a callback task is often the right first build. Add booking and transfers once that works.

## Does the new voice-model wave change what you should buy?

It changes less than the demos suggest. In the past month, OpenAI released [GPT-Live-1 in its API](https://openai.com/index/introducing-gpt-live-1-in-the-api/) on September 10, 2026. It's a full-duplex model that listens and speaks at the same time, and the front-end voice layer is priced at $0.05 per minute, with backend model usage [billed separately](https://developers.openai.com/api/docs/models/gpt-live-1). On September 28, ElevenLabs launched [Eleven v4 and v4 Turbo](https://elevenlabs.io/blog/eleven-v4). The company reports about 100 ms median inference latency for Turbo, about 150 ms to first speech, and support for more than 90 languages.

For a buyer, that means natural-sounding voices and fast turn-taking are now table stakes. Every serious platform will offer them within months, and many already let you swap the voice or model underneath. Don't pick a vendor because its demo voice sounded the most human. Pick the one whose booking, transfer and CRM behavior holds up on your calls, and make sure you can change models later without rebuilding.

## How do you test an AI receptionist before you buy?

Don't rely on the vendor's demo line. Set the agent up on a test number with your real hours, services and calendar, then call it yourself. Here's the scorecard we use before a build goes live. Run each call at least twice, once from a quiet room and once from a car or a loud shop floor.

| # | Test call | Pass looks like | Score (0–2) |
|---|---|---|---|
| 1 | Simple booking: "Can I get in Thursday afternoon?" | Offers real open slots, books it, sends a confirmation text | |
| 2 | Interrupt mid-sentence and change the day | Stops, adjusts and doesn't double-book | |
| 3 | Spell a hard name and read back a phone number | Captures both correctly in the CRM record | |
| 4 | Ask for pricing it wasn't given | Declines to guess and offers a callback or transfer | |
| 5 | "I need to talk to a person" | Warm transfer with context, or a clean callback if nobody answers | |
| 6 | Existing customer with a complaint | Recognizes it isn't a sales call and routes it correctly | |
| 7 | Call after hours | Follows your after-hours rules, not daytime ones | |
| 8 | Silence for 10 seconds, then hang up | Doesn't loop. Logs a dropped call and triggers follow-up | |
| 9 | Spanish or a strong accent (if your callers need it) | Understands, or switches language if configured | |
| 10 | Check the back end afterward | Summary, disposition, recording and cost per call are all visible | |

A score of 16 or higher out of 20 is ready for a soft launch. Anything that scores zero on tests 4 or 5 is a dealbreaker. An agent that makes up prices or traps a caller who asked for a human will cost you more than it saves.

We also time the gap between the end of the caller's sentence and the start of the agent's reply. Any delay longer than a normal human pause makes callers start talking over the agent. We check that again after every change to the prompt or model.

## What are the red flags when choosing a vendor?

- **An annual contract before you've run a single real test call.** Several platforms are pay as you go, so you have options.
- **You can't export recordings, transcripts or call data.** That data is your history. If you leave, it should leave with you.
- **No post-call webhook or API.** You'll end up stuck with whatever integrations they happen to support.
- **"Unlimited" plans with no stated limits on concurrency or call length.** Look for the cap. It's usually somewhere in the fine print.
- **Transfer is "on the roadmap."** If the agent can't hand a call to a person today, it isn't a receptionist.
- **No way to see cost per call.** You should be able to audit your bill at the call level.
- **Nobody owns updates.** Hours, services and staff change. Someone has to update the agent's instructions and test them again.

Calls coming in after hours need their own rules. We cover that in our guide to [after-hours AI call answering](https://wrkless.com/after-hours-ai-call-answering.php).

## Should you build on a platform or buy a managed service?

Buy a managed, per-call service if you get under a few hundred calls a month, need little integration and want it running this week. Build on a platform like Retell if calls need to book into your real calendar, update your CRM, transfer by rule and trigger follow-ups across your other tools. That's the work Wrk Less does: we map the call flow, wire up the integrations, run the scorecard above, and stay on for tuning after launch.

**Want a second opinion on your call flow?** [Book an AI systems assessment](https://cal.com/kris-hubbard-wrk-less/ai-systems-assessment) and we'll go through where your calls go today, what an AI receptionist should handle, and what it would cost to run at your volume.

## Questions people ask

### How much should a small business budget per month for an AI receptionist?

Usage is usually the small part. At 300 calls a month averaging 3 minutes, per-minute platforms such as Retell or Bland come to roughly $113 to $126 in usage based on their public October 2026 pricing, while a managed per-call service like Smith.ai's Enterprise tier is $500 a month for 300 calls. Budget separately for setup, integrations and ongoing tuning, which is where most of the real cost sits.

### Is per-call or per-minute pricing better for an AI receptionist?

Per-call pricing is easier to predict and favors longer conversations, while per-minute pricing is cheaper when most calls are short. Per-minute platforms also bill for silence and hold time, and some charge for transferred minutes, so compare both models using your actual average call length and monthly call count.

### Can an AI receptionist transfer a caller to a real person?

Yes, most voice platforms support call transfer, but quality varies. A good setup does a warm transfer that tells the staff member who is calling and why, and falls back to a captured callback task if nobody picks up. Some platforms reserve warm transfers for higher tiers, so confirm it on the plan you are buying.

### What is the fastest way to tell if an AI receptionist will work for my business?

Set it up on a test number with your real hours, services and calendar, then place about ten scripted calls: a simple booking, an interruption, a hard-to-spell name, a pricing question it should not answer, a request for a human, an after-hours call and a silent hang-up. Check that each call lands correctly in your calendar and CRM afterward.

### Do I need a new phone number to use an AI receptionist?

Usually not. Most setups forward unanswered or after-hours calls from your existing business line to the AI agent, or route them through your current phone provider. The right approach depends on your phone system, so review your routing before switching anything on.

## Sources

- https://www.retellai.com/pricing
- https://vapi.ai/pricing
- https://www.bland.ai/pricing
- https://synthflow.ai/pricing
- https://www.smith.ai/pricing/ai-receptionist
- https://www.ruby.com/pricing/
- https://openai.com/index/introducing-gpt-live-1-in-the-api/
- https://developers.openai.com/api/docs/models/gpt-live-1
- https://elevenlabs.io/blog/eleven-v4
- Retell pricing page, checked Oct 5, 2026: pay as you go $0.07-$0.31/min; calculator default $0.11/min; US Twilio telephony $0.015/min; numbers $2/month; knowledge base +$0.005/min
- Vapi pricing page, checked Oct 5, 2026: $0.05/min hosting plus provider costs; Pro success package $999/mo minimum
- Bland pricing page, checked Oct 5, 2026: Start $0.14/min, Build $0.12/min plus $299/month; transfer minutes $0.04-$0.05/min
- Synthflow pricing page, checked Oct 5, 2026: enterprise contracts start at $30,000 annually
- Smith.ai AI Receptionist pricing, checked Oct 5, 2026: free 25 calls then $3.00/call; Pro $150/mo for 75 calls; Enterprise $500/mo for 300 calls ($1.67/call); extra calls $2.10-$2.50; 10% spam allowance
- Ruby pricing page, checked Oct 5, 2026: $250/month for 50 minutes up to $1,725/month for 500 minutes
- Wrk Less estimate: 300 calls x 3 min = 900 min; Retell ~$0.125/min = ~$113/mo; Bland Start $0.14/min = $126/mo
- ElevenLabs Eleven v4 blog, Sep 28, 2026: v4 Turbo ~100 ms median inference, ~150 ms time to first speech, 90+ languages
- OpenAI GPT-Live-1 API launch, Sep 10, 2026: $0.05 per minute front-end voice layer