How voice AI books a taxi: what happens on the booking call, step by step

You don't need to understand the technology to use a voice booking agent, but knowing how it works tells you where it can go wrong and what to test. This guide takes a call apart, piece by piece, in plain words. There are no accuracy figures here, because any figure that matters is the one you get on your own callers.

What is voice AI taxi booking?

Voice AI taxi booking is software that takes a taxi booking by voice, on the phone, with no person on your end of the call. It's AI taxi booking by phone, and it's built from several pieces:

1. The phone line delivers the call to the agent, usually by a divert from your own number.
2. Speech recognition turns the caller's words into text.
3. A language model reads that text and works out what the caller wants.
4. The booking questions fill in whatever is missing.
5. The address check turns "the Red Lion" into a place on a map.
6. The fixed price lookup finds your fare for that journey.
7. Text to speech turns the agent's reply into a voice.
8. The booking system saves the job and sends the confirmation.

Done well, the caller hears a conversation, not a menu. The sections that follow take the pieces one at a time.

How does speech recognition hear the caller?

Speech recognition listens to the audio and writes down the most likely words. Phone audio is harder than a voice note. Google's speech-to-text documentation describes its telephony model as built for audio from a phone call, typically recorded at an 8 kHz sampling rate, which carries less detail than a recording made on the phone itself.

Less detail means more guesswork on sounds that are close together. B, D, P and T sound alike down a line, which is why controllers have always said "B for Bravo".

The fix is telling the recogniser what to expect. Google calls this model adaptation: you bias the model towards words it would otherwise get wrong, such as proper nouns and local names. For a taxi firm that means your streets, pubs, estates, hospitals and stations. A recogniser that has been told your local names is more likely to write them down correctly. That's the first thing to ask about speech recognition for taxi bookings: can it be given your place names?

How does it know what the caller wants?

A language model reads the transcript and sorts it into the details a booking needs. A caller who says "Can I get a car from the Red Lion to the airport at half four tomorrow, two of us" has given four answers in one breath. The model picks out the pickup, the destination, the time and the passengers, and notices what's missing.

Then it asks the booking questions it still needs. These are the ones a good controller asks:

The order matters less than the rule: never save a booking with a gap in it. If the caller won't or can't give a detail, the call goes to a person.

How does the address check work?

The address check turns what the caller said into a precise place. The usual tool is geocoding. Google's Geocoding API takes an address or a place name and returns map coordinates and a place ID.

Two parts of the answer matter for a taxi booking. The first is a flag called partial_match, which Google sets when it couldn't match the whole request, only part of it. The second is the location type: ROOFTOP means a precise street address, and APPROXIMATE means what it says.

A careful agent treats either signal as a reason to ask again. "There are two Station Roads in the area. Is that the one near the station or the one by the school?" is a better use of a few seconds than a driver at the wrong end of town.

Times get the same care. "Half four tomorrow" has to become a date and a clock time, and the agent should say both back in full: "Thursday the 15th at 4.30 in the morning." A time that has already passed, or a date that doesn't match the day the caller named, gets a question, not a guess.

How does it find the fixed price?

The fixed price lookup matches the journey to your price list. Once the pickup and destination are places on a map, the system works out which of your zones each falls in and reads the fare you set for that pair, plus any extras such as a bigger car or meet and greet.

The agent doesn't invent prices. If a journey falls outside your zones, or the route isn't priced, the right set-up hands the call to the office. Our guide to taxi zone pricing explains how to draw zones so the list covers the routes your callers ask for.

How does the agent speak back to the caller?

Text to speech turns the agent's written reply into a voice. Developers control how the words come out with a markup language called SSML. Google's text-to-speech documentation shows how it can read text letter by letter, read a phone number digit by digit, and say dates and times the way a person would.

That matters on the read-back. A postcode should be spelt out, letter by letter. A phone number should be read in digits, not as one huge number. "04:30" should come out as "four thirty in the morning".

What happens after the caller says yes?

The booking system takes over once the caller agrees. The booking is saved with the details your records need. Outside London, section 56 of the Local Government (Miscellaneous Provisions) Act 1976 requires operators to record the particulars of every booking before the journey. The Department for Transport's statutory standards add the time of the request and the staff member who took the booking, and suggest keeping records for at least 6 months, so the record should show the agent took it.

Then the passenger gets a confirmation by text or email. Our guide to taxi booking confirmation texts and emails covers what to put in it.

If you take payment up front, the agent sends a secure payment link, not a request for the card number. The PCI Security Standards Council's guidance on telephone payments warns about card data ending up in call recordings, and a payment link keeps it off the call. Our guide to card payments for taxi bookings goes into the options.

Where does voice AI go wrong, and what catches it?

Voice AI goes wrong in predictable places, and a well-built agent has a check for each one:

The read-back is the safety net under all of them. On a good agent, nothing is saved until the caller has heard the booking and said yes. Ring any agent you're considering and try to break it with your own awkward addresses.

Dispatch AI is booking and dispatch software with a voice booking agent of this kind. Its AI phone agent answers your booking line when your office doesn't, quotes your own fixed prices and books the job into your Dispatch AI account, and each agent takes up to 10 calls at once. You can hear it on the Dispatch AI demo line, and our guide to how an AI phone agent books a job covers the set-up.