Schedule with AI
Turn natural language into an event draft or use a conversation to act on your calendar.
Calendar has two AI workflows: filling a new event form from a sentence, and executing calendar actions through a conversation. These features are available when AI is enabled for your deployment.
Choose the right AI workflow
Start with the outcome you want. Drafting, searching, and executing an action have different effects on saved data.
| Goal | Use | What happens before you review the result |
|---|---|---|
| Turn a sentence into one new event | Natural-language filling in the event editor | The form changes, but the event is not saved |
| Find events by subject or constraints | Semantic search | Matching events are returned without changing them |
| Read and update a schedule through conversation | Ask AI | Tool calls can create or change saved data |
Use semantic search when you only need to find something. An instruction such as “move the meeting” belongs in Ask AI, where the assistant can act on the result.
Fill an event draft
Natural-language filling is available when creating an event, not editing an existing one. It prepares a draft for review rather than saving it for you.
Describe the event in its title
For example: Project review next Wednesday from 3 to 4 pm in Room A, with a reminder 10 minutes before. Include the date, time zone, or participants when they matter.
Press Enter to parse
Keep the editor open while the draft is filled. Fields you did not mention retain their current values. If you specify a new start without an end, the current duration is preserved.
Review and save
Check dates, time zone, end time, and recurrence before saving. The success notification includes an undo action to restore the form to its state before parsing.
Closing the editor cancels a parse that is still running. If parsing fails, your text remains available so you can finish the form manually.
Control what the draft replaces
Mention the details you want to change. A short title such as Lunch can still produce a valid draft without inventing a location, participant list, or recurrence rule.
If your current draft lasts one hour, describing a new start without an end keeps that hour-long duration. If you want a different duration, include an end time or duration explicitly, then inspect both fields before saving.
The undo action restores the form's pre-parse values. It is not an undo for a later saved event or for actions executed by Ask AI.
Act through a conversation
In the command palette's Ask AI mode, describe the goal, time range, and constraints. For example: Find next week's project reviews and move Thursday's review 30 minutes later. Leave the others unchanged.
The assistant uses tools to read or change calendar data. Destructive tools require confirmation; creating or updating events can execute directly. Review the events and tool results after an action completes.
State boundaries before asking for a change
Give the assistant enough detail to distinguish the target from similarly named events. Include the date range, category or participant when useful, and whether a recurring change applies once or from a particular date onward.
These examples make the intended scope explicit:
| Intent | Example request |
|---|---|
| Inspect before changing | List next week's client reviews with their dates and durations. Do not change them. |
| Move one occurrence | Move the October 14 project review from 10:00 to 11:00. Keep its duration and leave other occurrences unchanged. |
| Create a bounded schedule | Create a 30-minute reading session every Monday at 18:00 for six occurrences, starting October 12. |
| Find a gap | Find a 45-minute gap in my calendar next Tuesday between 09:00 and 17:00. Do not create an event yet. |
Use a year or time zone when the date could be ambiguous. A free-time search uses calendar records available to your account; it does not establish the availability of every person you might invite.
Read execution results alongside the answer
A conversation can contain several tool calls. Some can succeed while a later operation fails. Treat the returned events and operation results as evidence of what changed, rather than interpreting a fluent summary as proof that every action completed.
Confirm the affected date, recurrence scope, and participants in Calendar after a multi-event change. If the assistant selected the wrong target, stop issuing further edits until you have identified the saved state.
Recover from an interrupted reply
Closing the palette stops the current response but does not undo completed actions. The conversation remains when you reopen it until you start a new conversation.
After a network interruption, check whether an event was already created or changed before sending the request again. Read execution results alongside the assistant's summary; a failed operation is not a saved change.
Starting a new conversation clears the conversation context, not the calendar actions already performed. Closing and reopening the palette also does not roll those actions back.
Work when AI is unavailable
The event editor, ordinary keyword search, and command palette's non-AI actions remain separate workflows. If AI reports a configuration or service error, finish the event manually or use keywords to find it.
If you reach a request limit, wait before retrying. Repeatedly resubmitting a change after a timeout can create ambiguity about which attempts completed; inspect the calendar first.
Connect an external assistant
The built-in assistant and external clients have separate access controls. To let another client use Calendar tools, follow Connect with MCP, grant permissions for the task, and review its activity.