Follow a single booking through the crew.
Descriptions of AI are easy to nod along to and hard to verify. This is one ordinary ride, minute by minute, with the agent handling each step named and the human decision points marked.
A ride is requested by phone
A rider calls rather than using the app — common for airport runs and older riders. The call is answered, transcribed and matched to an existing rider record. Dispatch AI reads the pickup, drop and any access notes from previous trips at the same address.
Allocation and confirmation
The booking is placed into the operator's dispatch system, a car allocated, and the rider given a vehicle, driver name and ETA on the same call. Nothing here needs a person.
The ETA slips
Watch AI notices the allocated car has not moved for four minutes against a live traffic baseline. It does not wait for the rider to complain — it re-checks allocation and prepares a proactive message.
The rider is told before they ask
A message goes out with the revised ETA and the option to re-allocate. Proactive contact on a slipping ETA is the single highest-scoring thing a mobility desk does, and it is exactly the thing a busy human queue skips first.
Driver reports a device problem
Mid-trip, the driver's app drops the fare meter. Driver Support AI recognises the error signature from the operator's own history, sends the recovery steps, and opens a fare-reconstruction case so the trip can still be billed correctly.
A trip anomaly fires
The route deviates significantly from the expected path and the trip has been stationary for six minutes in an unusual location. Safety AI classifies it, pulls the trip record, the driver history and both contact records — and stops.
A human has it, inside a minute
Your supervisor's phone is ringing. The full context is already on the board: trip, route, both parties, history, and what the crew did not do. The human decides everything from here.
Resolved — a wrong turn and a roadworks diversion
The supervisor confirms with the driver, notes the outcome and closes the flag. The trip completes normally. Safety AI logs the classification, the outcome and the human decision for the weekly sample.
The fare is disputed
The rider queries the final fare against the original estimate. Fare AI reconstructs it from trip data, the diversion, waiting time and the meter drop, and prepares an adjustment with the evidence attached.
Adjustment released, ride closed
Under the ceiling, so your supervisor reviews and releases. The rider has an answer the same hour. The reconstruction, the evidence and the release are all on the record.
What that trace actually shows
Six agents touched one ride. Two of the ten steps needed a person, and one of those was mandatory rather than discretionary. Nothing was deferred to a morning shift, and nothing about the hour of the day changed how any of it was handled.
Note also what the crew did not do: it did not decide whether the anomaly was serious, it did not contact anyone on the driver's behalf, and it did not close the safety flag. It collected context and got out of the way. That distinction is the whole architecture.
What that trace leaves out.
That ride went well. Most do. Here is what the same page would show on a bad one.
A rider whose accent the transcription mishandles, so the pickup address is wrong and Dispatch AI confidently books the wrong street. A driver dispute where the operator's own policy is silent and the crew has no precedent to retrieve. A safety flag that turns out to be nothing, twice a week, waking your supervisor each time — because the alternative is a false negative, and there is no acceptable rate of those.
All three land in your queue. That is the job you are taking on, and it is why the eligibility page asks about out-of-hours reachability before it asks about anything else.
Any vendor who shows you only the clean trace is showing you a demo, not a desk.
Watch this happen on a live board.
Including the part most vendors skip: a safety flag firing, the crew refusing to act on it, and the escalation landing on a human.