AI scheduling: a free slot is not always a possible meeting
Assess AI scheduling against time zones, rooms, required attendees and buffers, and distinguish a proposal from a booking.
Article language: English

In this article
Everyone has a free slot from ten. The room is free too. An assistant proposes a ten o'clock meeting and considers the job done. One person, however, is finishing a workshop in another building, another starts the day in a different time zone, and the room still needs to be prepared. A free rectangle on a calendar has the advantage of never getting tired on the stairs. People, unfortunately, have more complicated requirements.
A shared gap is not enough to assess AI scheduling. Check whether the proposal satisfies the constraints of the participants, place and task, then distinguish it from booking a resource and sending an invitation. These are three different outcomes. Each needs its own confirmation.
Start with a controlled example using fictional calendars. You can reproduce a conflict, missing information and a time change without sending real people invitations to an eight o'clock Sunday training session. The exercise quickly shows what the tool actually knows and what it merely fills in for itself.
Record the conditions before showing the calendar
Imagine a company preparing a product demonstration across two locations. The meeting needs a presenter, a technical specialist and a room with a demonstration station. Between demonstrations, the team needs time to prepare and walk to the other building. Attendees can join remotely, but the presenter must be on site. This description contains more than a list of participants.
Separate mandatory conditions from preferences. The technical specialist may be essential, while a morning time is merely more convenient. If the model cannot find a time, it should identify the condition blocking the task. It should not silently turn a required person into an optional one just because that produces an appealing proposal.
Identify the source of each condition. Working hours may come from calendar settings, room availability from its bookings, and preparation time from an agreement within the team. That last item need not appear as an event. An audit should check whether the process receives the information and how it uses it, rather than assuming that everything is visible on one screen.
Microsoft Graph's findMeetingTimes documentation (in English) describes suggestions based on availability and supplied time and location constraints. It also describes a reason returned when there are no suggestions. This is the scope of one specific function; it does not mean the tool automatically knows all of your organisation's conditions.
Add the meeting length, acceptable date range, required resources and time between events to your task brief. State whether the meeting can be split and who may approve a change to a condition. This keeps the assessment from depending on whether the presenter remembers to mention moving equipment after seeing a proposal.
Separate permission to read from what is actually known. If the tool cannot see one required person's calendar, it has no basis to say they are free. Prepare that case and check whether the process asks for confirmation or fills the gap with a convenient assumption. The report should identify whose availability could not be established.
Choose a recognisable state for missing data. “I do not know the preparation time” is different from “preparation takes zero minutes”. If the system uses a default, show it before approval and check whether the planner understands its meaning. A hidden zero can be an efficient way to produce impossible schedules.
Add an informal date change: “move it to next Friday”. If the time comes from a form, also check how the date is entered on the site so ambiguity does not arise before the assistant starts. In the controlled example, set the reference day and expected date, then inspect the summary before the operation. When the phrase is ambiguous, the process should obtain a specific date. A confident answer does not resolve a difference between two possible Fridays. For related guidance, see the Polish Signal article “how the date is entered on the site (in Polish)”.
The same clock time does not always mean the same instant
Google Calendar distinguishes a calendar time zone from an event time zone (in English). A recurring event also needs a time zone to expand its occurrences. For reviewing a proposal, this means retaining the time-zone information with the date and time rather than passing along “10:00” alone.
You do not need to convert every time by hand to test the process. Prepare cases with a known expected instant and check each participant's view. Record the date, local time and time zone used for the assessment. If someone sees a different day, the difference should be visible in the summary before an invitation is sent.
Include a series that crosses a clock change in the regions involved. Do not rely on a remembered fixed time difference. Check whether successive occurrences follow the agreed rule, such as the presenter's same local time. Record the expectation in advance and compare several dates in the series.
In collaboration between Poland and Asia across time zones, the agreed time is part of the working arrangement. An assistant should respect the acceptable range for all required people. A technically shared gap does not by itself mean a meeting in the middle of someone's night follows the team's agreement. The article on collaboration between Poland and Asia across time zones (in Polish) offers related guidance.
Add an all-day trip or unavailability. Check how the tool interprets that entry in the particular calendar and whether it considers it when proposing times. Do not transfer the behaviour of one event type to all others. Name the initial state, settings and observed result in the report so another person can repeat the test.
One change can reveal what the system is checking
First run a simple case in which all conditions produce one known time. Then change only room availability. Restore it and add a conflict for the technical specialist. In the next test extend preparation time. This arrangement shows which condition the process actually applies without guessing from a complicated final schedule.

Check the meeting's end boundary. If a demonstration ends at eleven and the next task starts at eleven elsewhere, non-overlapping clock times are not enough. The team should agree on a buffer and how to account for it beforehand. Do not add one universal number of minutes to every situation; different tasks may require different preparation.
Test a resource available for only part of the proposed period. A room free at the start is not a booking for the whole meeting. Similarly, a presenter available for the first half-hour does not satisfy a requirement to lead an hour-long demonstration. Choose a case where partial compliance looks tempting and inspect the final decision.
Add a room with a similar name. “Workshop A” in the first location and “Workshop A” in the second can look identical in a shortened list. Before confirmation, show an identifier for the place together with information a user understands, such as the building or address. A correct name without a location does not prove that the right resource was booked.
If the result is passed on as structured data, connect the test to checking the meaning of AI-generated JSON. A record can have a valid date format while referring to the wrong room or a different occurrence in a series. Accepting the format does not finish the schedule check. See also the article on checking the meaning of AI-generated JSON (in Polish).
Compare two valid proposals that differ only in a preference. Can the model explain their order using the agreed conditions? If a time convenient for the presenter was chosen, do not describe it as the only possible time. A user should be able to distinguish a requirement from a choice among valid possibilities.
For a series, check whether an inconvenient time repeatedly falls on the same person if the team agreed to rotate it. This organisational rule must be explicitly supplied and checked across occurrences. One correct meeting does not show how the plan distributes the burden across the whole series.
Check whether a task requiring equipment treats it as a separate resource. Two rooms may be free at once while the only demonstration kit is available for just one. In a controlled test, schedule two apparently valid meetings using the same station. Acceptance should reveal the shared-resource conflict before both bookings are treated as feasible.
For such an error, show not only the rejected time but also whether the condition was missing from the input or omitted in the assessment. These cases need different corrections: completing the task description or changing how the process works.
Also create a case with no possible time. The expected result should identify the conflict and a way to resolve it. Perhaps the location can change, the task can be shortened or someone can be asked for a wider date range. The model should not silently decide for the process owner which mandatory condition to discard.
For that discussion, see the Aeneia Audit overview and prepare the task's mandatory conditions. Fictional calendars let you assess the process without sending real invitations during the first test.
The calendar keeps changing between proposal and invitation
A time may have been free when read and occupied by the time it is approved. Test a change in availability during that interval. Does the process reassess the conditions or send an invitation based on the old proposal? Keep the check time and operation time in the report to distinguish a planning error from a state change.
In human approval of AI actions, show the time, time zones, participants, place and scope of the change before consent. When moving an existing series, the recipient should know whether the request concerns one meeting, future occurrences or the whole series. “Move it” does not settle that by itself. For related guidance, see the article on human approval of AI actions (in Polish).
Test rejecting a proposal and having a person change one condition. The system should retain the other agreements or clearly indicate that they need confirmation again. Returning to the beginning and losing the requirements can produce an appealing proposal for a task nobody is still requesting.
Also test cancelling part of the plan. If one demonstration disappears, what happens to the room booking and a separate preparation block? A resource should not stay occupied by an action that no longer exists. At the same time, cancelling one occurrence must not delete the agreed series if that was outside the task.
After the operation, check the relevant calendars and resources. A created draft, a sent invitation and an accepted invitation are not the same state. Decide which constitutes the agreed result and describe it precisely. Without that definition, “meeting ready” can easily turn a plan into a report from the future.
Run one more interruption test after preparing the proposal but before confirmation. On return, show what was saved, what awaits a decision and what needs a fresh read. The planner should not have to guess whether closing the window sent an invitation. Test this with someone who did not watch the earlier steps.
When many times are produced in parallel, you need a durable connection between the proposal, its case and input version. This is developed in the article on accepting AI batch results (in Polish). A late answer must not overwrite a plan a person has already changed.
Put that boundary into the responsible agent workflow (in Polish) too: who proposes, who approves and where the effect is confirmed. A schedule becomes useful when its conditions can be explained and its changes reconstructed. Each time then shows whom it concerns, which conditions produced it and which state was actually confirmed.







