Use cases
AI employee for appointment booking
Tobiloba Odejinmi · 3 Apr 2026 · 6 min · 1,121 words
Direct answer
A booking employee offers slots that exist, writes the appointment, and tells the truth when the week is full. Good looks like a calendar row the front desk can trust. Escalate clinical advice, payments, and anyone who needs a named clinician now. Do not automate cancellations of clinical care first. Measure bookings that still exist the next day and the no-show rate on those bookings.
- The job is real inventory, a written booking, and an honest no.
- A ghost slot is a reliability bug, not a chat miss.
- Clinical advice and money stay with a person.
- Measure surviving bookings and no-shows, not conversations.
What is the booking job, not the calendar feature?
The job is to put a person in a real slot and make sure both sides agree on the time. That means live inventory, identity, the reason for the visit, and a confirmation that matches the calendar. A widget that “helps people schedule” without writing the row is not an employee. It is a form.
I think about this the way I think about 10mg on a Monday. More than 6,000 providers. Loans that have to land. If the system is wrong, a patient does not get seen. Booking is the same class of problem. Pretty chat does not help a full waiting room.
Write the rules before you write the prompts. Who can book which slot. How far out. What needs a referral. What needs a deposit. If those rules live in someone’s head, the employee will guess. Guessing on a clinic calendar is how you get two patients at 10:00.
What does a good booking flow look like?
The caller or the chat asks for a time. The employee offers two or three real options. The person picks one. The row is written. The confirmation repeats the day, time, place, and what to bring. The front desk can open the record and see the same thing.
Good also means a clean no. “Nothing this week. Next is Tuesday at 9. I can waitlist you for a cancel.” That sentence saves more trust than a hold that was never a hold.
Keep the questions short. Reason, name, contact, the slot. If you need insurance details, collect the ones you already collect on the form. Do not invent a new intake because the model likes asking questions.
- Inventory is live, not a cached week
- Confirmation matches the calendar text
- No slot means no slot
- Front desk can see the note without asking
When do you stop and get a person?
When the person asks for a clinician, describes a symptom that your protocol treats as urgent, wants to change a clinical plan, or wants to talk about money. Also when identity is unclear. Also when the calendar and the employee disagree.
The handoff carries the slots already offered, the reason, and the contact. The person who picks up should not ask “what day were you hoping for” if that answer already exists. That is a state handoff. A transcript of the chat is optional. The fields are not.
Urgent medical is a hard gate. I will not let a booking bot talk someone out of coming in. It can offer a slot or a transfer. It cannot say “probably fine”.
What should you not automate first?
Do not start with cancelling clinical care. Do not start with overbooking logic you do not already run by hand. Do not start with a bot that “optimises the day” by moving patients around. That is operations research with a smile. It will anger the room.
Skip multi-location puzzles until one location works. One calendar. One set of rules. Then add the second site. I have watched teams try to launch five clinics at once and spend a month on timezone bugs.
Do not automate reminder copy that includes clinical instructions unless a clinician signed the template. “Don’t eat after midnight” is a medical instruction. Treat it like one.
How do you measure bookings that stick?
Count appointments that still exist the next day. Count double books. Compare no-show rate to the human-booked baseline. If no-shows jump, your confirmations are weak or you are booking people who did not mean to book.
Time to a confirmed slot is useful. Containment is only useful if ghost bookings are near zero. I would rather have a lower containment and a clean calendar than a chat that “handled” 90% and a Monday full of ghosts.
Sample ten bookings a week. Did the reason match? Did the time match? Would the front desk have booked this? Write the noes down. That list is the work.
What about no-shows and the after-hours slot?
Reminders help when they are true. Time, place, how to cancel. A reminder that cannot cancel into a real waitlist just nags. If you add an after-hours employee, let it book the next real slot and take a message. Do not let it promise a clinician who is asleep.
After-hours booking is useful. After-hours medicine is not a first job. I wrote about nights without a night shift elsewhere. The short version: take the work that can wait until morning, and wake a person when the protocol says wake a person.
If you serve clinics, reliability is the product. A booking employee that writes the wrong row is an outage. Treat it with the same seriousness you treat login. The slide about “AI scheduling” will not help at 8am.
Questions people ask
What should a booking AI employee do?
Read the calendar you already use, offer times that are actually open, collect the fields you need, and write the appointment. Then confirm the time in the same words the calendar uses. If it cannot see live inventory, it should not book.
What if there is no slot?
Say so. Offer the next real slot or a waitlist. Do not invent a hold. A fake Tuesday is worse than a wait. The front desk will spend the morning unwinding it.
Can it reschedule and cancel?
Reschedule into a real slot, yes, once the identity check you already use has passed. Cancel clinical care only with the rules a person wrote, and escalate if the rules say escalate. Do not let a model cancel a procedure because the caller sounded sure.
What about clinics after hours?
Take the booking or the message. Do not give clinical advice. If the protocol says wake someone, wake someone. At 10mg we care that a provider can open the system on Monday. A night booking that is wrong is a Monday incident.
How do we measure it?
Appointments that still exist the next morning. Double books. No-show rate versus the human-booked baseline. Time to a confirmed slot. Chat volume is not a booking metric.
Written by
Tobiloba Odejinmi
Head of Engineering at 10mg Health. I have run engineering at Zeeh Africa and sold Insurpass and Shopl. I still write the code. If you have one process that still runs on people copying things, we can look at it in thirty minutes.
