A dental office doesn’t have one kind of appointment. A cleaning takes 30 minutes and can go almost anywhere in the day. An endodontic consult takes 90 minutes and only fits the specialist’s hours.
Most AI booking setups handle that by building one agent per service, or one giant decision tree. On Myna, one agent can book all of them. It works out which appointment the customer means, books it on the right calendar under that calendar’s rules, and asks when it isn’t sure.
The calendar owns the rules. The agent owns the access.
The split that makes this work is simple.
A calendar owns its booking rules: its duration, its booking hours for each day, its timezone, how soon and how far ahead it can be booked, its location, and what it’s called (an appointment, a booking or a reservation). Each calendar can also check connected Google or other calendars for busy time, and write a copy of each booking there.
An agent owns which calendars it may use, whether it can offer and move appointments, and its scheduling instructions.
When the agent books, the calendar’s rules win. The agent’s own settings are never a source for duration, hours or notice windows. That’s what lets one agent carry calendars with completely different rules side by side. A calendar can also be shared by several agents, and every booking keeps a snapshot of its calendar’s details, so editing a calendar later doesn’t rewrite past bookings.
The description is the router
Here’s the part people don’t expect: none of the scheduling tools has a “which calendar” setting, and the agent’s planner never picks one. The planner only decides that it’s time to schedule.
The choice happens inside the scheduling step. A dedicated extraction reads the turn and returns two things together: which calendar, and what time the customer wants. It looks at:
- each linked calendar’s name, description and tags
- the latest message and the recent conversation
- what’s already known about the contact
- the step’s instructions, and today’s date
Then it applies a short set of rules:
- One calendar linked? Use it.
- Several? Match the latest message against each calendar’s name, description and tags. A service the customer newly asks for beats one they mentioned earlier.
- Still unclear? Use the rest of the conversation and the contact’s details. A field like “service interested in” can settle it.
That makes the calendar description the routing text. You don’t build a decision tree. You write a good description, like “For anyone asking about spinal decompression or ongoing back pain”, and that’s what the agent matches against.
It also means the offer, the booking and any later move all agree on which calendar is meant, because the choice is made once per turn and reused.
This is how La Perle Tours’ agent chooses between four calendars, from private airport pick-ups to boat rides, based on what the customer asks for and how many people are travelling.
Ask, don’t guess
When the calendar genuinely can’t be worked out, Myna doesn’t pick one at random. It also doesn’t guess a time, because without a calendar there’s no timezone to read “3pm” in.
So nothing is searched and nothing is booked. Instead the agent asks one ordinary question that would tell the options apart, without ever saying “calendar” or naming one. A customer asking “can I come in Thursday?” might hear “Happy to help. Is this for a cleaning, or about the tooth that’s been bothering you?”
The agent is also told that calendar names are internal, and that openings and hours only ever come from a real availability search, never from its own assumptions.
Book only an exact time
There are three scheduling actions, all built on one shared scheduling service:
| Action | What it does |
|---|---|
| Offer Appointment | Searches the chosen calendar and offers one real slot, plus the open time around it |
| Schedule Appointment | Books one specific time, never a window, a day or a range |
| Modify Appointment | Moves an existing visit, keeping it on its own calendar |
Booking only ever happens for an exact time the customer agreed to. “Sometime Tuesday afternoon” gets an offer, not a booking.
And if a booking fails, because the slot was just taken, is too soon or is already in the past, the agent doesn’t report an error. It offers fresh openings instead, in the same reply. Because booking and moving are treated as critical actions, a failure can never come out as “you’re booked.” That’s the same truth check we described in how tools work inside a Myna agent.
No double-booking, even under pressure
Availability is searched on the chosen calendar only. Busy time comes from that calendar’s existing bookings, any other Myna calendars it checks for conflicts, and live busy blocks from connected Google or other calendars.
One rule matters more than it sounds: a calendar that can’t be read is never treated as free. If a connected calendar doesn’t respond, the result is uncertain and nothing is booked. Guessing “probably free” is how double bookings happen.
Then, at the moment of booking, Myna checks everything again with the relevant calendars held so nothing else can book them in the same instant:
- The same booking can’t be made twice.
- The contact doesn’t already have an upcoming visit on that calendar.
- Nothing overlaps on this calendar or any calendar it checks for conflicts.
- Connected external calendars are checked one last time.
Only then is the booking saved. Copies to connected calendars are written afterwards, each tracked separately, so a slow external calendar never undoes a confirmed booking or creates a duplicate copy.
Three actions, deliberately
When we designed this, it would have been easy to add more pieces: a separate classifier to pick the calendar, a dedicated booking tool per appointment type, a table of appointment types, a tracker for every offer. We considered all of them and left them out.
Three actions around one shared service, with the calendar chosen inside it, turned out to be simpler to configure, easier to reason about, and harder to get wrong.
Want a stricter path?
Inference works well when your calendar descriptions are clear. Some businesses want a fixed path instead. Today you can get there by:
- writing tighter calendar descriptions and tags
- using a contact field that identifies the service
- giving each service its own agent with a single calendar, so there’s nothing to infer
Coming soon: multi-step workflows, where an intake step routes to dedicated booking steps, each with its own calendar, for businesses that want a fixed path instead of inference.
The short version
One agent, many appointment types, each with its own rules. The calendar description decides the routing, the agent asks when it isn’t sure, only exact times get booked, and every booking is re-checked right before it’s saved. It’s scheduling that behaves the way a careful front desk would.