By KYN AI Advisory Team — AI implementation specialists, Singapore
An AI agent that books meetings needs hard rules, not judgement. Set working hours, buffers, a minimum notice period, a time zone per attendee, a hold that expires, and a list of meeting types it may book without a person. Check availability from the live calendar at the moment of booking, not from a cached list. Anything that moves an existing meeting or involves a client's senior contact should wait for a named person.
What does a calendar-booking agent actually do?
It turns "can we talk next week?" into a confirmed event. That sounds small. It contains five decisions: which person, which meeting type, which slots are free, which slot the other side picks, and who is told. Most calendar tools expose it. The risk is not the feature. It is booking the wrong person into the wrong slot under your name.
Which availability rules should be fixed?
Write these as configuration, not as prompts.
- Working hours per person. Per day, in the person's time zone. Include public holidays.
- Buffers. Time before and after each meeting type, for travel or preparation.
- Minimum notice. No booking inside the next hours you choose.
- Maximum horizon. No booking beyond a set number of weeks.
- Daily cap. A limit on meetings per day per person.
- Meeting types. Each with a duration, a host list and a location or link.
The agent chooses among options that pass these rules. It never invents a slot because the customer asked.
How should the agent check availability?
Ask the calendar at booking time. Google Calendar exposes a free/busy query that returns busy intervals for the calendars you name, and Microsoft Graph has an equivalent for Outlook. Use it for every offer and again at confirmation.
Two failure modes come up repeatedly.
- Stale offers. The agent proposes three slots on Monday. The customer replies on Wednesday. One slot is gone. Re-check before confirming and say so if the slot changed.
- Double booking. Two conversations pick the same slot within minutes. Place a short hold when you offer a slot and release it on expiry. Confirm with an atomic write that fails if the slot is taken.
How do you handle time zones?
Store every time in UTC. Show each person their own zone. Put the zone in the confirmation text, for example "Tuesday 14 January, 3:00 pm Singapore time (SGT)". For clients across Singapore, Hong Kong, Indonesia and Malaysia, the offset can be the same or an hour apart, and Indonesia spans more than one zone. Do not let the agent guess a zone from a phone number. Ask, or read it from the contact record.
Which actions should wait for a person?
Our rule is that outbound, money and irreversible actions wait for a named person. For calendars that means:
| Action | Agent alone | Needs a person | |---|---|---| | Offer slots for a standard meeting type | Yes | | | Book a first call from a rule-matching reply | Yes, if enabled | | | Reschedule a meeting the agent booked | Yes, once | | | Move or cancel a meeting a person created | | Yes | | Book with a client's senior contact or a regulator | | Yes | | Overbook against the daily cap | | Yes | | Delete an event | | Yes |
A person approves exceptions in one click from a queue. The queue shows the request, the proposed slot and why the rule failed.
What should the agent write into the event?
Make the event useful on its own.
- Title with the company and meeting type.
- A description with the source: which message asked, and the account link.
- The agenda or prep note the host needs, marked as agent-written.
- The video link or address.
- A tag so you can filter agent-booked events later.
Then write the booking to your own log with actor and time. A calendar is not an audit trail. Build one in your database, as in Postgres as the agent's memory.
What about reminders, no-shows and cancellations?
Reminders are low risk and high value. Send one a day before and one on the morning, through the channel the contact used. Provide a one-tap reschedule link that goes through the same rules. If a contact does not show, the agent records it and proposes a rebooking, but a person decides whether to chase. Cancellations free the slot at once and notify the host.
Messaging channels carry their own rules. Check consent before messaging on WhatsApp and honour opt-outs. We cover the sales side in is it safe to let an AI agent send emails to customers.
What does a booking rule set look like?
Write the rules in a table that a salesperson can read and change. An illustrative set for a sales team follows.
| Rule | Setting | |---|---| | Meeting types | 20-minute intro, 45-minute review | | Working hours | Monday to Friday, 9:30 to 17:30 local time | | Buffer between meetings | 15 minutes | | Minimum notice | 4 hours | | Booking window | Up to 21 days ahead | | Who may be booked | Named reps only, not the whole directory | | Attendee outside the company | Allowed, with confirmation message |
Rules in a table are testable. Create ten test requests: a Friday evening, a public holiday, a slot with a clash, a request with no time zone. Check the agent's answers against the table. Any answer the table does not cover goes to a person.
How do you roll out booking?
Roll out in three steps. First, run the agent in suggestion mode: it proposes slots and a person clicks to confirm. Track how often the person changes the proposal. Second, allow direct booking for one meeting type with known customers, and keep confirmation messages going to a person for review. Third, widen to other types once the change rate is low. Keep a one-page note for the team on how to block time, mark leave and override the agent. The agent is only as accurate as the calendars it reads, so ask everyone to keep their calendar current. Review double-bookings, missed meetings and cancellations each month and adjust the rules table accordingly.
What happens when the agent cannot decide?
Some requests fit no rule: a customer asks for a Saturday, a meeting for six people across three calendars, or a time inside a rep's leave. The agent should not improvise. It replies that it will confirm shortly, and creates a task for the named person with the request, the nearest valid slots and the reason it stopped. That person answers in one click. Count these hand-offs each month. If one pattern repeats, such as Saturday requests, decide once whether to allow it and add the rule.
Where KYN fits
KYN builds booking into the sales or operations system rather than as a separate tool. Agents have their own restricted accounts with calendar access limited to the people and meeting types you approve. Every booking is logged with the actor, and outbound messages wait for a named person unless a rule you set allows otherwise.
The first 14 days cover your calendars, meeting types and the people who book them. A full build ships in under 90 days. See sales systems and integrations.
Frequently asked questions
Can an AI agent read my whole calendar?
It should not. Grant free/busy access where possible, so the agent sees when you are busy but not the event titles. Give event-level access only to the calendars it must write to.
How do I stop double bookings?
Check free/busy at the moment of confirmation, place a short hold when you offer a slot, and use an atomic write that fails if the slot is taken. Log each hold and release.
Should the agent reschedule meetings I created?
No. Let it manage events it created, once, under the same rules. Moving a meeting you set up yourself changes someone's day, so a person approves it.
Checked 30 October 2026.
Curious whether Marketing agents fits your business? Talk to KYN on WhatsApp — no forms, just a conversation.
