In short: an AI receptionist answers your phone line, converts the caller’s speech to text, sends that text to a language model that has been given your business’s information and a set of tools, and acts on what the model decides — checking a calendar, creating a booking, writing to your CRM, sending a text. The whole loop runs in well under a second so the conversation feels like a conversation.
Most explanations of this stop at “it uses AI.” That tells you nothing about whether it will work on your phone line.
So here is one call, from the first ring to a job on the schedule.
0.0 seconds — the call arrives
Your number is hosted by a telephony provider. When someone dials, that provider opens a connection and starts streaming the audio in both directions rather than sending it to voicemail.
This is the least glamorous layer and the one that decides whether anything else matters. If your existing number isn’t ported or forwarded correctly, none of the rest happens.
0.3 seconds — speech becomes text
A transcription model turns what the caller just said into text, continuously, while they are still speaking. Good ones handle accents, background noise and the caller talking over the agent.
Two things go wrong here more than anywhere else. Trade vocabulary — brand names, part numbers, street names — gets mangled unless the model has been given a hint list. And callers rarely speak in sentences. “Yeah hi, so, the upstairs unit, it’s not — it’s blowing but it’s not cold?” is a normal opening and the system has to cope with it.
0.5 seconds — the model decides what to do
The text goes to a language model along with two things you supply.
Your knowledge base
Services, prices, service area, hours, after-hours rates, what the maintenance plan covers, brands you will and won’t work on. This is the part clients underestimate and it is the part that determines whether the agent sounds like your business or like a chatbot.
Its tools
A list of actions it is allowed to take: check availability, create a booking, look up an existing customer, send an SMS, transfer the call. The model doesn’t just talk — it picks a tool and calls it. This is the difference between an AI receptionist and a recorded message.
So when the caller says the upstairs unit is blowing warm, the model recognises a repair request, works out which questions it still needs answered, and asks the next one.
0.8 seconds — text becomes a voice
A speech model turns the reply into audio and streams it back. If the total round trip creeps past about a second, the caller starts talking over the agent because they think it didn’t hear them, and the call falls apart.
That latency budget is why every layer above is chosen for speed as much as quality.
The next ninety seconds — qualifying
The agent works through whatever your dispatcher would ask. Is the address inside the service area. What is the equipment doing. Is anyone home during the day. Has this been looked at before.
Two behaviours separate a good setup from a frustrating one. It should confirm rather than assume — reading the address back, spelling the street. And it should skip what it already knows: if the number matches an existing customer record, asking for their address again is the fastest way to sound like a machine.
Two minutes — the booking
The agent calls its calendar tool, gets back real availability, and offers slots that genuinely exist. The caller picks one. The agent creates the booking, writes the customer and the job details into your system, and confirms by text before hanging up.
This step is where most deployments quietly fail. If the agent isn’t connected to the system your business actually runs on, it can only take a message — and then someone retypes it in the morning, which was the job you were trying to remove. An agent that books is a different product from one that transcribes, and the difference is entirely in the integration work.
The calls where it should give up
A properly built agent has rules for when to stop being clever and pass the call to a person.
Gas, smoke, carbon monoxide, water pouring through a ceiling. Hard keyword rules that transfer immediately and, if nobody picks up, tell the caller what to do. Not a judgement you delegate to a model.
A repeat visit for the same fault, a billing dispute, a complaint about a technician. Detect the tone and hand over. Software has never calmed anyone down.
A quote for a replacement system, a question that isn’t in the knowledge base, an unusual request. A confident wrong answer costs more than a transfer.
Ask any vendor what their transfer rules are. If the answer is that the agent handles everything, you’re being sold a support problem.
Want to see this on your own phone line?
We build and connect these end to end — see AI receptionist services, or start with an AI opportunity audit that counts what you’re actually missing.
What it costs to run one call
That whole ninety-second conversation costs somewhere around fifteen cents. A mainstream stack runs about nine to fifteen cents a minute once telephony, the platform and the AI models are added together — we broke every line of that down in what an AI voice agent actually costs, with each vendor price checked against its own page.
The part that isn’t technology
Everything above is solved. The layers work, the models are fast enough, the integrations exist.
What decides whether it works for your business is the unglamorous preparation: an accurate knowledge base, a real connection to your scheduling system, and honest transfer rules. An agent given none of those will disappoint you in a week, regardless of which platform it runs on.
If you’re still deciding whether a phone agent is even the right tool, start with the comparison against chatbots and live answering services. Plenty of businesses find their enquiries arrive typed rather than spoken.
Frequently Asked Questions
How does an AI receptionist actually work?
Four layers run in a loop. A telephony provider answers the call and streams the audio. A transcription model converts speech to text as the caller speaks. A language model, given your business information and a set of tools, decides what to say and which action to take — check a calendar, create a booking, write to your CRM. A speech model turns the reply back into audio. The full round trip stays under about a second so the call feels natural.
Can an AI receptionist book appointments by itself?
Yes, if it is connected to your scheduling system. It calls a calendar tool, receives real availability, offers slots that exist, creates the booking and confirms by text. If it is not connected, it can only take a message that someone has to retype later — which removes most of the benefit.
What does an AI receptionist need to know before it takes calls?
Your services and prices, service area, opening hours, after-hours rates, what any maintenance plan covers, and any equipment or brands you will not work on. It also needs live access to real availability rather than a generic calendar. The quality of this preparation, not the platform, is what determines whether it sounds like your business.
When should an AI receptionist transfer a call to a person?
On safety keywords such as gas, smoke or carbon monoxide, which should trigger an immediate hard-rule transfer; when a caller is angry or disputing a bill; and for anything outside its knowledge base, such as quoting a replacement system. A vendor claiming their agent handles every call is describing a support problem.
Will callers know they are speaking to AI?
Most will within a sentence or two. The setup that works is to say so at the start and offer a transfer to a person. Callers who are comfortable continue; those who want a human ask and get one. Disguising the agent as a named person tends to backfire.