How to Set Up an AI Receptionist for Your Business in 2026: A Step-by-Step Guide

Missed calls are missed revenue. For most small businesses the front desk is the hardest seat to staff — and now it is one you can fill with an AI voice agent that answers on the first ring, takes messages, books appointments, and never takes a sick day. Search interest tells the story: "ai receptionist" queries are up roughly 125x year over year, and "ai call centers" about 300x, according to October 2026 search data from Rising Trends. This is not a futuristic experiment anymore; it is a procurement decision.

Why now: the numbers behind the rush

The steepest growth curves in AI search in 2026 belong to AI that works a shift: Rising Trends' October 2026 data shows "ai call centers" at about 300x a year earlier and "ai receptionist" at about 125x.

On October 21, 2026, Toku launched Makimoto Flows, a visual builder for speech AI pipelines inside its open-source Makimoto initiative, aimed at developers in Asia-Pacific.

Step 1 — Pick your route: hosted, developer, or open-source

The no-code route (Bland, Retell). You configure the agent in a dashboard: pick a voice, paste in a script, connect a phone number. Bland publishes itemized pricing with everything bundled — Start costs $0.14 per connected minute with no platform fee and no card required, with the LLM, transcription, and voice all inside that rate. Retell's official pricing page lists $0.07 to $0.31 per minute on pay-as-you-go: $0.055 for its voice infrastructure, $0.015 for text-to-speech, $0.015 for telephony, plus the model you choose — a Haiku-class model adds about $0.0032 a minute, so a lean build lands near $0.10 a minute all in.

The developer route (Vapi). Vapi charges $0.05 a minute for hosting and passes the model, transcription, voice, and telephony costs through at the provider's listed price. Assembled stacks typically land between $0.12 and $0.31+ per minute. This is the choice when you want to swap models and voices freely and you have someone who can read a billing breakdown.

The open-source route (Makimoto). Toku's Makimoto gives developers a drag-and-drop Flows canvas — Transcribe, Summarise, and Tags blocks that map to API endpoints and export as Python code — with 1,000 free minutes for every new account. It is not a managed receptionist product; it is infrastructure you assemble yourself, with audio and transcripts processed in Singapore.

A rule of thumb: if nobody on your team will touch code, pick a bundled platform. If you want model freedom and can read a multi-line bill, go developer. If you are a developer in Asia-Pacific who needs data residency, Makimoto deserves a look.

Step 2 — Get a phone number

Every platform lets you buy or port a number without leaving the dashboard. Retell sells phone numbers at $2 a month. Bland includes an inbound number valued at $15 a month with its free Start signup. Vapi sells numbers for $2 a month and charges Twilio pass-through rates — about $0.008 a minute inbound and $0.014 outbound in the US — or you can bring your own carrier and skip its markup.

Buy the number before you finish the script — you will use it for test calls. Porting an existing business number takes days, so don't port the number customers already dial until the agent has passed its tests.

Step 3 — Choose the voice and the brain

Two components decide whether callers tolerate your agent: latency and voice quality. Latency is the harder one — every 100 milliseconds of pause makes the conversation feel more robotic. That is why the model choice matters: fast, small models (Claude 5.5 Haiku at $0.0032 a minute on Retell, or GPT-4.1 nano at the same price) keep turns snappy and cheap. For context on where model speed is heading, see our Inception Mercury 3 vs MiniMax M3.1 comparison, and our Claude Haiku 5.5 pricing analysis for the cheap-model angle.

For the voice itself, pick one and stop fiddling. Retell offers platform voices at $0.015 a minute or premium voices up to $0.10 a minute; Bland bundles premium voices and cloning into its flat rate. Test two candidates with the same script and keep the one that sounds most natural over a phone line.

Step 4 — Write the receptionist's call script

This is where most setups fail. A good receptionist script is a job description, not a prompt-engineering exercise. Write it the way you would brief a new hire:

• Identity and tone: "You are Alex, the front-desk assistant at Bright Smiles Dental. Friendly, calm, unhurried."
• The first sentence: exactly what the caller hears, including the AI disclosure — "Hi, you've reached Bright Smiles, this is Alex, the AI assistant. How can I help?"
• The three jobs: answer hours and location questions, book or move appointments, and take messages with a name and callback number.
• The exits: what happens when someone is upset, asks for a human, or goes off-script — "I can transfer you to the front desk; one moment."
• The forbidden list: never quote prices you can't verify, never give advice outside its job, never guess — say "I don't have that; I'll take a message."

Keep the system prompt under 300 words. Long prompts slow down responses and the agent will ramble. Then record ten test calls and rewrite anything the agent says twice.

Step 5 — Connect the tools: calendar, CRM, transfer

• Calendar booking. Bland ships an appointment-scheduling node on Build; Retell and Vapi expose function calling you wire to Google Calendar or your booking system. Test booking, confirmation, and cancellation — double-bookings cost more than the platform ever will.
• Transfer to a human. Bland charges $0.05 a minute for transfer time on Start ($0.04 on Build); Retell stops the AI fee the moment the call transfers and only telephony continues. Set a warm-transfer rule: the agent announces the caller before handing off, and falls back to a message if nobody picks up.
• After-call notes. Every platform generates transcripts and summaries; push these into your CRM or a shared inbox so the front-desk conversation doesn't die in the dashboard.

Step 6 — Test, then go live

Never point your main number at an untested agent. Run at least 25 test calls covering: the greeting, a routine booking, a reschedule, an angry caller, a silent caller, and a request for a human. Retell ships pre-built templates and simulation testing; Bland says most teams deploy within a day on self-serve plans. Check every transcript — voice agents mishear names, dates, and phone numbers more often than you expect, and the fix is usually a confirm-everything rule ("always confirm phone numbers digit by digit").

Go live with a soft launch: point a secondary number or your after-hours overflow at the agent first, watch the first hundred calls, then port the main line.

What it actually costs: the honest comparison

Put real numbers on it. For 1,000 connected call minutes a month: Bland Start costs $140 flat plus transfer time. A lean Retell build (platform voices, a Haiku-class model) lands around $100–$120 plus the $2 number. A typical Vapi stack lands around $120–$240 before Twilio's roughly $8–$14 for the call legs. On Bland's Build plan ($299 a month + $0.12 a minute), the same 1,000 minutes cost $419 — which only pays off above roughly 3,700 minutes a month, where the lower per-minute rate finally offsets the platform fee.

Compliance add-ons are the hidden line item. Retell's pay-as-you-go plan is HIPAA-ready and SOC 2 certified with no extra compliance fee on its public pricing page; Vapi offers HIPAA as a paid add-on; Bland includes HIPAA with a signed BAA on Enterprise. If you handle health data, get the compliance answer in writing before you go live.

Quick pros and cons

• Bland: one predictable per-minute number; Start caps at 100 calls a day.
• Retell: no platform fee or minimums, billed per second; you still pay for the model layer separately.
• Vapi: total control over models and voices; the bill arrives in pieces.
• Makimoto: open-source, free minutes, Singapore data residency; you assemble the agent.

A word on disclosure and consent

Tell callers up front that they're speaking with an AI — the script above does this in the first sentence. Several US states require it. Keep recordings and transcripts somewhere you control: Retell offers PII redaction and custom data retention, and Bland's compliance page lists SOC 2, HIPAA, GDPR, and PCI DSS.

FAQ

How much does an AI receptionist cost per month?

For 1,000 call minutes a month, roughly $100–$200: about $140 flat on Bland Start, around $100–$120 on a lean Retell build, or $120–$240 on Vapi. Compliance add-ons are the one cost that can dwarf the calls themselves.

Do I need a developer to set one up?

No. Bland and Retell are dashboard products — pick a voice, paste a script, connect a number, and most teams are live within a day. Vapi is built for developers who want control over the whole stack. Makimoto sits in the middle: its drag-and-drop Flows canvas needs no code, but exporting and running the pipeline does.

Will callers know it's an AI?

They should — disclose it in the greeting. Modern voices are convincing on short calls, but longer conversations, interruptions, and unusual accents still give it away. Treat disclosure as a trust feature, not a limitation.

Can it book appointments and take payments?

Booking, yes — every major platform wires into calendars or booking systems via tools or native scheduling nodes. Taking payments over the phone is a different compliance story; Bland lists PCI self-attestation, but get written confirmation from any vendor before card data touches a call.

What about languages other than English?

Most platforms offer multilingual voices, and Makimoto was built for Asia-Pacific conversations — accents, multilingual speech, and code-switching are its stated design targets. Test your actual callers' accents, not the demo audio; quality varies a lot by language and dialect.

Key takeaways

• An AI receptionist needs six things: a platform, a phone number, a voice and model, a script, tool integrations, and testing — do them in that order.
• Budget roughly $0.10–$0.14 per connected minute all-in; at 1,000 minutes a month that's $100–$200, before transfer time and compliance add-ons.
• Bundled pricing (Bland) is predictable; modular pricing (Retell, Vapi) is cheaper but arrives in pieces — pick the bill you can actually read.
• The script is the product: a 300-word job description with exits and forbidden topics beats any model upgrade.
• Disclose the AI in the greeting, confirm phone numbers and dates digit by digit, and soft-launch on a secondary number before porting your main line.

Published October 21, 2026. Prices and features verified on the vendors' official pages the same day; vendors change pricing, so re-check the official pricing page before you commit.