AI phone receptionists are suddenly everywhere, and most demos sound impressive. A demo isn't the same as a phone line that real customers call every day, though. Here's how the one I built actually works, call by call, and the rules it follows that never make it into a demo.
When it picks up
The business keeps its existing phone number. When staff can't answer, the business line forwards the call. It travels over an encrypted connection to the AI agent, which picks up and greets the caller in the business's name.
It tells callers up front that it's an AI assistant. That's both honest and practical: people speak more clearly to a system they know is automated, and nobody feels tricked when it hands them to a person.
Two models, two jobs
Inside, the work is split between two AI models.
A real-time speech model does the talking. It listens, responds naturally and can be interrupted mid-sentence, like a person can. Its job is the conversation.
A separate reasoning model runs the actions. When the caller asks whether there's availability on Saturday morning, the speech model passes the request to this tool model, which looks it up and returns the answer.
Splitting it this way means the conversation stays quick and natural, and the parts that touch real data are handled by a model chosen for accuracy and cost.
It talks to the booking system, carefully
The agent never touches the business's database directly. It calls the business's own booking system through a signed connection that only allows specific actions: check availability, look up prices, create a booking link. Each request is verified on the business's side before anything happens.
That means the agent can only do what it's been explicitly allowed to do, and the booking system stays in charge of what's actually available.
What a typical call looks like
A caller asks about renting two bikes on Saturday morning. The agent:
- Confirms the date, the time and how many people.
- Checks live availability through the booking connection.
- Gives the price from the current price list.
- Texts the caller a booking link for exactly that slot.
- Answers a couple of questions, like where to park, from the business's own information.
- Ends the call politely.
The caller books and pays through the link, on their own phone, in their own time.
What it never does
It never takes card numbers. Payment happens through the secure link, never over the call. Card details never appear in a transcript or reach the AI.
It never claims something it didn't confirm. If the booking connection didn't confirm availability, the agent doesn't say it's available.
It doesn't pass spam through. Robocalls and sales pitches don't get transferred to staff.
It doesn't improvise on policy. Refunds, damage, complaints and anything unusual go to a person, with a summary of the call so far.
These rules are enforced in code, not just written in the instructions. The model can't talk its way around them.
After the call: every call reviewed
When a call ends, a second AI model reads the transcript. It writes a short summary, scores how the call went and specifically flags anything the agent said it did but didn't confirm through the booking system. The owner gets an email or text for anything that needs attention: a frustrated caller, a request it couldn't handle, a possible mistake.
That review step is what makes it safe to run unattended. Problems surface the same day, not when a customer complains.
Where it fits
An AI receptionist is a good fit when:
- You miss a meaningful number of calls, during busy periods, on jobs or after hours.
- Most calls are routine: availability, booking, hours, directions, prices.
- You have a booking system or calendar it can check.
It's a poor fit when most calls need judgment, sensitive personal information or a long relationship conversation. In those cases it can still answer after hours and take a detailed message, which beats voicemail.
The part that matters most
The voice is the easy part now. What makes a phone agent trustworthy is everything around it: limits on what it can do, a booking system that checks every request, a clear handoff and a review of every call. If you're comparing products, ask about those, not just how natural it sounds.