AI Receptionist: How It Works and When to Build a Custom One

What an AI receptionist does

A receptionist's phone work is mostly the same few jobs repeated all day: answer, find out what the caller wants, give an answer or book something, and pass anything else to the right person. An AI receptionist does those jobs with the same building blocks as any modern AI voice agent, but tuned for the front desk: short calls, many callers, and a lot of calendar work.

The difference between a good and a bad one is not the voice. It is whether it has the right information, the right connections to your systems, and sensible rules for when to stop and hand the call to a person.

AI receptionist vs human receptionist, voicemail and IVR

An AI receptionist is not a straight replacement for any one of these. It sits between a phone menu and a person: it understands people like a person does for routine requests, and it scales like a phone system does.

The usual pattern is a mix: the AI receptionist takes after-hours, overflow and routine calls, and people take the calls that need judgement.

How an AI receptionist works

Every call passes through the same five stages. Understanding them helps you judge any product or proposal, because each stage is a place where quality, speed or cost is won or lost.

1. 1

   Telephony: getting the call to the AI

   Your number rings on your phone system as usual. A routing rule sends the call to the AI receptionist, either through a SIP connection from your PBX or carrier, or through a voice-AI platform's own number. This is also where after-hours and overflow rules live.

2. 2

   Speech-to-text: hearing the caller

   The caller's audio is streamed to a speech recognition model that turns it into text as they speak. Voice activity detection works out when they have finished a sentence, which matters for both speed and interruptions.

3. 3

   Language model and tools: deciding what to do

   A large language model reads the conversation, your instructions and your approved information, then decides on the reply. When it needs to act, it calls a tool: check the calendar, book a slot, look up an order, create a CRM record or transfer the call.

4. 4

   Text-to-speech: replying

   The reply is turned into speech and streamed back to the caller, starting as soon as the first words are ready rather than waiting for the full sentence.

5. 5

   Handover and follow-up

   When the call needs a person, the receptionist transfers it with a short summary. After the call it can send a confirmation text or email, log the outcome and add a transcript for your team.

Why response speed decides whether callers trust it

In conversation, people expect a reply within about a second. Longer pauses make callers talk over the system or assume it has failed. The total delay is the sum of the stages above. These are typical ranges we design for; real numbers depend on the models, the hosting region and the carrier, so we measure them on real calls.

What an AI receptionist does well, and what it shouldn't do

Being honest about the limits is what makes a deployment work. The best results come from giving the AI a clear, narrow job and fast routes to a person for everything else.

The other common failure is wrong information. A language model will fill gaps confidently if you let it, so the receptionist should answer only from your approved information and say it will check with the team when it doesn't know. That is the same principle we use for RAG chatbots.

Off-the-shelf AI receptionist app or a custom build?

For a single small office with simple needs, an off-the-shelf app is often the right first step: quick to try and easy to cancel. A custom build earns its cost when the receptionist has to fit your phone system, your software and your rules rather than the other way round.

Signs you need a custom AI receptionist

• You want to keep your existing numbers, PBX or call-centre setup
• It must book into a system the apps don't support, such as an in-house or industry-specific one
• You handle a high volume of calls, so per-minute fees add up
• Calls and recordings must stay in a particular country or on your servers
• You need languages, dialects or a voice the apps don't offer
• You run several sites or departments with different rules
• You want to resell the receptionist to your own customers
• The same assistant should also answer chat and WhatsApp

Connecting the receptionist to your phone line and calendar

Two connections do most of the work. The first sends calls to the AI at the right times. The second gives the AI safe, limited tools to act in your systems.

1. After-hours and overflow routing on Asterisk

On an Asterisk PBX, a few dialplan lines send calls to staff during opening hours and to the AI receptionist otherwise, or when nobody answers within 20 seconds. The AI endpoint here is a SIP trunk namedai-receptionist; it could equally be an AudioSocket or LiveKit SIP connection.

; extensions.conf — main number context
[from-trunk-main]
exten => _X.,1,NoOp(Inbound call from ${CALLERID(num)})
 same => n,Set(CHANNEL(language)=en)
 ; Mon–Fri 09:00–18:00 and Sat 10:00–14:00 ring the front desk first
 same => n,GotoIfTime(09:00-18:00,mon-fri,*,*?office)
 same => n,GotoIfTime(10:00-14:00,sat,*,*?office)
 same => n,Goto(ai,1)

 same => n(office),Dial(PJSIP/frontdesk1&PJSIP/frontdesk2,20,tT)
 ; nobody answered in 20 s, or all lines busy: send to the AI receptionist
 same => n,Goto(ai,1)

exten => ai,1,NoOp(Routing to AI receptionist)
 same => n,Set(PJSIP_HEADER(add,X-Caller-Number)=${CALLERID(num)})
 same => n,Set(PJSIP_HEADER(add,X-Route-Reason)=${IF(${EXISTS(${DIALSTATUS})}?${DIALSTATUS}:AFTER_HOURS)})
 same => n,Dial(PJSIP/receptionist@ai-receptionist,,tT)
 ; if the AI endpoint is down, fall back to voicemail
 same => n,VoiceMail(100@default,u)
 same => n,Hangup()

2. A booking tool the AI can call

The language model never writes to your calendar directly. It calls a tool you define, and your code checks the request, applies your rules and talks to the calendar API. This is the tool definition in the JSON Schema format that OpenAI-compatible and most other LLM APIs accept:

{
  "type": "function",
  "function": {
    "name": "book_appointment",
    "description": "Book an appointment after the caller has confirmed the service, date and time. Only call this with a slot returned by check_availability.",
    "parameters": {
      "type": "object",
      "properties": {
        "service": { "type": "string", "enum": ["consultation", "follow_up", "cleaning", "emergency"] },
        "slot_start": { "type": "string", "format": "date-time", "description": "ISO 8601 start time in the business time zone" },
        "caller_name": { "type": "string" },
        "caller_phone": { "type": "string", "description": "E.164 format, e.g. +14155550123" },
        "notes": { "type": "string", "maxLength": 300 }
      },
      "required": ["service", "slot_start", "caller_name", "caller_phone"],
      "additionalProperties": false
    }
  }
}

Your handler then books through Google Calendar, Microsoft 365, a booking platform or your practice-management system, sends a confirmation and returns a short result the AI reads back to the caller. Keeping the logic in your code, not in the prompt, is what makes the receptionist safe to give real access.

How to set up an AI receptionist in six steps

Whether you buy an app or build a custom system, the same steps decide whether it works.

Where AI receptionists are used

Any business that misses calls or ties staff to the phone can use one. These are the settings where the fit is strongest.

Related: AI voice agents for insurance · healthcare communication software · real estate

How much does an AI receptionist cost?

There are two different cost shapes. Apps charge a monthly plan or a per-minute fee, which is simple and cheap at low volume. A custom receptionist has a one-off build cost, then running costs that you pay directly to the telephony, speech and model providers, with no platform markup if it is self-hosted.

What makes up the running cost

• Phone minutes and numbers from your carrier or SIP trunk
• Speech-to-text, billed by audio minute
• Language model usage, billed by tokens
• Text-to-speech, billed by characters or minutes
• Hosting for the agent and media servers if self-hosted
• Monitoring, transcript storage and occasional prompt updates

Ranges match our AI voice agent development page. You get a fixed quote after a discovery call.

Disclosure, recording and data

Callers accept AI receptionists when they are told what they are talking to and can reach a person easily. The rules on recording and data vary by country and industry, so build them in from the start.

More detail: TCPA and AI outbound calling · HIPAA requirements for voice AI agents

Frequently asked questions