How to build an AI receptionist: 8 steps from call list to launch

How do you build an AI receptionist?

To build an AI receptionist, list the calls it should handle, pick a voice platform, connect your phone number, write its instructions and knowledge base, connect your calendar and CRM, then test it on real call scenarios before it answers live calls. Start with after-hours calls, then widen it.

Each of those steps is simple on its own. What decides whether callers like it is the order: the call list comes before the platform, and the testing comes before launch. This guide walks the steps in that order, says where each platform's own documentation covers the detail, and names the mistakes that make AI receptionists sound bad.

We build these for clients, but this guide is written so you can do it yourself. If you want the finished system built for you, that is our AI receptionist service.

How does an AI receptionist work?

An AI receptionist answers your phone line, turns the caller's speech into text, lets a language model decide what to say or do using your instructions and business facts, and speaks the reply back. When it needs to act, it calls a tool, such as checking your calendar or saving the caller's details in your CRM.

The parts of one call, in order:

1. Phone line. The call arrives on a number the platform controls, either a new number, your number forwarded to it, or your number ported to it.
2. Speech to text. The caller's words are transcribed as they speak.
3. Language model. The model reads the transcript, your instructions and your knowledge base, and decides the next reply or action.
4. Tools. Booking, availability checks, contact updates or a transfer to a person run through connections you set up.
5. Text to speech. The reply is spoken in the voice you chose.
6. After the call. A summary and the details collected are written to the contact record, and a workflow sends a confirmation text or alerts your team.

Some platforms now offer speech-to-speech models that skip the separate transcription and voice steps. HighLevel, for example, sells them at one bundled per-minute rate and describes them as lower latency. The other parts of the system stay the same.

Step 1: Which calls should your AI receptionist handle?

Write down the calls your business actually gets before you build anything: the five or six reasons people call, the questions they ask most, and which calls a person must always take. An AI receptionist built from real calls sounds natural; one built from guesses sounds like a phone menu.

Pull a week of call logs or listen to recordings, and sort each call into a type. For a clinic or service business the list usually looks like this:

The last row is the most important one to write down. An AI receptionist earns trust by knowing what it should not handle.

Step 2: Which platform should you build an AI receptionist on?

If your business already runs on GoHighLevel, build on HighLevel Voice AI: it books into your calendars and updates contacts without an integration. If you need custom call logic or connect several systems, use a voice platform such as Retell, Vapi or ElevenLabs Agents and connect it to your CRM.

HighLevel's Voice AI overview, read October 17, 2026, lists what its agents do natively: answer and make calls, book appointments through an Appointment Booking action across one or several calendars, use a knowledge base, follow working hours, speak several languages, and run actions during and after the call, including transfers to a live agent.

Usage costs are close across the cheaper platforms. Our AI voice agent pricing research priced the same 1,000-minute month on each, and standard setups came to about $76 to $140. Pick on fit, not on a cent a minute.

Step 3: How do you connect your phone number to an AI receptionist?

You can keep your number, and the simplest start is forwarding: set your carrier to forward unanswered or after-hours calls to the AI receptionist's number. Later you can port your number into the platform so every call reaches the AI first. A new number works too, but customers already know your old one.

Outside the US the number step has extra requirements: a regulatory bundle for UK and Australian numbers, an ID check for Canadian ones, and no HighLevel number at all in the UAE. Country notes: UK, Australia, Canada and the UAE.

Three options, from least to most change:

1. Forward after hours or when busy. Your team still answers during the day; the AI catches the rest. It is the lowest-risk place to start.
2. Forward every call. The AI answers first and transfers when a person is needed.
3. Port the number. The number moves into the platform. Porting runs in parallel with the build and takes two to four weeks for under 50 numbers, from our own build timelines.

On HighLevel, you assign any available number in the account to the agent, as long as it is not already used for something like an IVR. Our answer on keeping your business phone number in GoHighLevel covers forwarding and porting in detail.

Step 4: What should the AI receptionist's instructions say?

The instructions, often called the prompt, should give the AI a role, a greeting that says it is an AI assistant, the questions to ask for each call type, the rules for booking and transferring, and what to say when it does not know. Short, specific rules beat long paragraphs about tone.

A structure that works:

• Role and greeting. "You are the virtual receptionist for Bright Smile Dental. Greet callers with: Thanks for calling Bright Smile Dental, this is the clinic's AI assistant. How can I help?"
• Call types. One short block per type from Step 1, with the questions to ask in order.
• Booking rules. "Always check availability before offering a time. Offer two times. Book only after the caller picks one, then read back the date and time."
• Transfer rules. "Transfer to the front desk for complaints, billing disputes, or when the caller asks for a person. Outside opening hours, take a message instead."
• Limits. "Never give medical advice or quote prices not in the knowledge base. If unsure, say a team member will call back today."

Saying it is an AI in the greeting is our recommendation for every build. The legal side of AI calls, including the FCC's 2024 ruling that AI voices count as artificial voices for outbound calls, is covered in is AI calling legal.

Step 5: What goes in an AI receptionist's knowledge base?

Put in the facts callers ask for and the AI may repeat: opening hours, address and parking, services and what each involves, prices you are happy to state, booking and cancellation policies, insurance or payment options, and directions for emergencies. Leave out anything you would not want read aloud to a stranger.

Write it as short, plain facts, one topic per section, and date it. "Cleaning: 60 minutes. New patient exam and cleaning: 90 minutes. Cancellation: 24 hours' notice or a $50 fee." The AI answers from what is there, so a gap in the knowledge base becomes a vague answer on the phone. After launch, every question the AI could not answer is a line to add.

Step 6: How do you connect the AI receptionist to your calendar and CRM?

Give the AI two calendar tools, one to check availability and one to book, and one CRM tool to create or update the contact. On HighLevel these are built-in actions. On Retell or Vapi they are API calls to your calendar and CRM, usually through n8n, Make or a small server.

The connections a booking receptionist needs:

• Check availability for the right calendar or staff member, in the caller's time zone.
• Book the slot with the caller's name, phone and reason, and return a confirmation the AI can read back.
• Create or update the contact so a repeat caller is recognised and the call history sits in one place.
• Transfer to a person, with a fallback when nobody picks up.
• After the call, write a summary and the collected fields to the contact, tag the outcome, and start the follow-up: a confirmation text, a reminder sequence, or a task for the team.

The after-call step is where an AI receptionist becomes part of the business instead of a separate gadget. If the booking lands in the calendar but nobody gets the summary, the team still has to call people back to find out what they wanted.

Step 7: How do you test an AI receptionist before it goes live?

Run scripted test calls for every call type, plus the awkward ones: a caller who interrupts, mumbles, asks for a person, books for someone else, calls from a noisy car or switches language. Check each booking and contact record afterwards. Fix the instructions and repeat until the results hold.

A short test plan:

1. Two calls per call type from Step 1, read from a script by someone who did not write the prompt.
2. Ten edge cases: angry caller, wrong number, silence, a request outside the rules, a question not in the knowledge base, a double booking attempt, a time just outside opening hours.
3. Check every result in the system: the calendar entry, the contact fields, the summary, the confirmation text.
4. Listen to the recordings. A correct booking can still sound rude or slow.

HighLevel's overview makes the same point about testing before launch: find gaps in knowledge or logic and confirm that workflows and actions trigger correctly.

Step 8: How should you launch and improve an AI receptionist?

Launch in stages: after-hours calls first, then overflow when your team is busy, then all calls if the results hold. Each week, read the transcripts of calls that ended in a transfer, a hang-up or no booking, and fix the instruction or knowledge gap behind each one.

Four numbers tell you whether it works: calls answered, appointments booked, calls transferred, and calls where the caller hung up early. Compare them with the weeks before launch. The AI receptionist ROI calculator turns those numbers into money, and our AI receptionist cost guide lists the running costs to set against them.

Should you build an AI receptionist yourself or hire someone?

Build it yourself if you have one location, a few call types and time to test and adjust each week. Hire someone when it must book across several calendars or staff, connect to more than one system, handle regulated calls such as healthcare, or when nobody on your team will own it after launch.

The platforms make a basic agent easy to set up. The work is in the parts this guide covers: the call list, the rules, the connections and the weekly review. When we build one, it is usually a small build (about $1,000 at the time of writing) or part of a full setup ($2,000 to $5,000), with six months of maintenance included. A small build takes days; a larger GoHighLevel setup with AI agents takes three to four weeks, as our build timelines show. The dental case study shows an AI calling agent, GoHighLevel and n8n working together for a real practice.

Questions people ask about building an AI receptionist

Can I build an AI receptionist for free?

You can try one for free, since Vapi gives $5 in starting credits, ElevenLabs has a free plan with 15 call minutes, and Retell lets you sign up without paying first. Running one is never free: every minute carries platform, voice and phone charges, which our pricing research puts at about $0.08 to $0.14 a minute for standard setups.

Do I need to know how to code to build an AI receptionist?

Not for a basic one, because HighLevel Voice AI, ElevenLabs Agents and Retell all have visual builders for the instructions, voice and knowledge base. You need code, or a tool like n8n or Make, when the AI must talk to a calendar or CRM the platform does not connect to natively.

Can an AI receptionist book appointments into my calendar?

Yes, if it has a booking tool: HighLevel Voice AI books into GoHighLevel calendars with its Appointment Booking action, including across several calendars. Other platforms book through an API call to your calendar or booking system. Always give it an availability check first, so it never offers a time that is taken.

Does an AI receptionist have to tell callers it is an AI?

We recommend it on every call, in the greeting. Rules differ by state and by call type, and outbound AI calls need prior express consent under the FCC's 2024 ruling. Saying it up front also sets callers' expectations, which makes the conversation go better. Our guide on whether AI calling is legal covers the details.

Can an AI receptionist speak more than one language?

Many can: HighLevel lists multi-language support for its Voice AI agents, and you should check your own platform's language list. Test each language you plan to support with real callers' questions, and put the key facts in the knowledge base for each language, since answers are only as good as what it has.

What happens when the AI receptionist cannot answer a question?

It should say so and offer the next step: transfer to a person during opening hours, or take a message and promise a callback. Set up a task or notification so the callback happens. Then add the missing answer to the knowledge base, so the next caller gets it.

Will an AI receptionist replace my front-desk staff?

For most businesses it takes the calls staff cannot get to, such as after hours, lunch and busy mornings, and leaves the front desk free for people in the room. Our answer to will an AI agent replace staff goes through where it helps and where a person is still better.

Related questions and services

• AI receptionist service
• AI voice agent pricing compared
• How much does an AI receptionist cost?
• GoHighLevel Voice AI setup
• AI calling agent service
• Is AI calling legal?

Sources and further reading

• HighLevel, Overview of Voice AI agents (capabilities, booking, knowledge base, working hours, languages, during and post call actions, phone number assignment), read October 17, 2026.
• HighLevel, AI products pricing (speech-to-speech description and rates), read October 17, 2026.
• Vapi pricing, ElevenLabs Agents pricing and Retell AI pricing (free starting options), read October 17, 2026.
• FCC, ruling that AI-generated voices in robocalls are artificial voices under the TCPA (February 2024).

Voice platforms add features every month. Where a step here differs from your platform's current documentation, follow the documentation.