A taxi booking call should put a usable job into dispatch: the right pickup, an agreed journey and a confirmed booking record. That is the work behind the AI voice agent we built for a UK taxi company.
Across roughly seven months, the deployment handled 130,000+ calls and 39,000+ successful booking conversations. These historical platform totals count calls that used the booking action and were marked successful; they do not count completed journeys.
The build connects ElevenLabs, n8n, local address lookup and SafeKab. Here is how a caller's request reaches dispatch, where the team stays involved and which failures shaped the booking safeguards.
Step 1: some calls never reach the agent
A phone router sits in front of the agent. Calls from hotel numbers and account numbers go straight to the operator's team. The router makes that decision before the agent speaks, so nothing depends on the agent recognising a special case mid-call.
For every other call, the caller's number loads their history and any existing bookings before the first question. If the call is about one of those bookings (chasing it, cancelling it, changing it), the agent does not start a new one. The call goes to the team or to an update of that booking.
Step 2: the pickup is a lookup, not a transcript
The agent resolves the pickup with a location lookup that uses landmarks, maps and the caller's history. If the result is unclear, it asks again once, then hands the call to dispatch. It does not keep guessing.
Three cases get their own handling:
- Airports. The agent asks where the flight is coming from, once.
- Hotels. The hotel is matched against a landmark list, and the room number goes into the driver notes.
- Far towns. The agent confirms with the caller before going on. An address that names only an area, with no street, is refused.
The hotel path exists because of a dated failure. An earlier hotel lookup returned internal dispatcher labels instead of usable pickup details. Hotels are now resolved server-side.
Step 3: what goes to the team before a booking exists
Journeys with several stops, wait-and-return jobs and bookings that need nine or more seats go to the team before any booking is made. A caller who asks for a person is transferred.
Step 4: one fare quote, one tariff, one clear yes
The agent quotes one fare from the operator's single tariff, recaps the journey and waits for a clear yes before it books into SafeKab. We put the quote the agent speaks and the price the booking carries on that one shared tariff, so both read from the same source.
- The router and the caller's history decide who handles the call before the pickup is discussed.
- The pickup comes from a lookup with one re-ask, then dispatch.
- Multi-stop, wait-and-return and 9+ seat jobs reach the team before booking.
- The caller hears one fare from one tariff and gives a clear yes.
- The agent says "booked" only when SafeKab returns a booking ID.
Step 5: "booked" needs a booking ID
SafeKab returns a booking reference when it accepts a job. The agent says the journey is booked only when it holds a real one. Otherwise the call goes to dispatch. We did not always run it that way.
- A confirmation without a booking. Callers were told "booked" while the booking call had returned nothing. The agent now fails closed to dispatch.
- A lookup release that lost coordinates. A release of the address lookup wiped the coordinates, so bookings were refused. We rolled it back. Contract tests (does the lookup's output still fit what the booking call accepts?) and sequence tests (do the steps still run in order?) are now required before any release.
Illustrative example — not a client transcript. The caller has heard the recap and the fare. The booking call to SafeKab returns nothing.
Caller: "Yes, go ahead."
Agent: "I haven't been able to confirm that booking, so I'm putting you through to dispatch now."
The wrong version is the agent saying "That's booked" and ending the call. That is the failure this rule now guards against.
What the operator gets
For a standard booking, the workflow brings the pickup, journey details, agreed fare and caller confirmation into the dispatch process. Complex trips, unresolved requests and callers who want a person have a staff route.
That division of work matters when you scope a deployment. Review the calls you want automated and the exceptions your team will retain, then test both paths through to their actual result.
Where this workflow fits
This implementation suits a SafeKab operator with routine journeys to automate and a team available for exceptions. SafeKab is the dispatch integration demonstrated here. A different dispatch platform needs its own integration and acceptance testing; hotels, account customers and complex trips remain staff work in this build.