AI, explained

Can AI agents book appointments?

Yes, and the interesting engineering is not the booking. It is everything that happens when two people try to take the same slot four hundred milliseconds apart.

Appointment booking is one of the most requested agent workflows and one of the easiest to build badly. The failure mode is almost never the conversation. It is state.

The rule that makes it reliable

The model must never be the source of truth about availability. It reads availability from the calendar at the moment of the request and writes through an API that enforces conflicts. If you let a model remember what was free earlier in the conversation, it will confidently offer a slot that was taken two minutes ago.

What a correct implementation does

1. Query the live calendar for real free slots, filtered by the rules that actually apply: business hours, buffer time between appointments, resource availability, staff skills, and travel time where the appointment is on site.
2. Offer a small number of concrete options. Three works. A wall of times produces indecision and more turns.
3. Place an atomic hold when the person picks one, with a short expiry. This is what prevents two conversations claiming the same slot.
4. Confirm the booking and release the hold, or let the hold lapse if the conversation stalls.
5. Send confirmation through the channel the person actually uses, with a working cancellation and reschedule link.

Rescheduling and cancellation

These are more common than first bookings in most businesses and are usually the weakest part of the build. Get the identification path right: the person rarely has a reference number, so match on phone number or email plus the upcoming appointment. Cancel and rebook as one operation rather than two, otherwise a failure between them leaves the person with no appointment and no notification. Handle the cancellation policy explicitly, including whether a late cancellation carries a fee, because getting that wrong is a complaint rather than a bug.

Where a human still confirms

• Anything with a clinical triage element, where the appointment type depends on a judgement the agent should not make.
• High-value or long-duration bookings where a mistake is expensive to unwind.
• Any request that arrives with distress or urgency in it. Route those to a person immediately and do not attempt a booking flow.
• Cases where the person has failed the identification path twice, which usually means something is genuinely wrong.

What to measure

Booking completion rate, double-booking incidents (which should be zero, not low), no-show rate compared to the human-booked baseline, and the rate of escalation to a person. If escalation is very low, check that the agent is not pushing through cases it should be handing over.

Frequently asked questions

Related