Most assistants answer questions. This one also makes changes, which is a materially different thing and is treated as one throughout.
What it will change
Meetings. Create one, move one, cancel one.
Approvals. Approve what is waiting on you.
Tasks. Mark one done.
Events. Cancel one, or push it out.
You ask in plain language. "Move my three o'clock to Thursday." "Approve the expense from Priya." "Mark the pricing task done." There is no syntax to learn and no command to remember.
Nothing happens without your tap
Every change shows a confirmation card first, and the card is specific: which record, what changes, what it becomes.
Nothing is written until you press confirm. Not the change, not a draft of the change, nothing. Reading the card back and closing it leaves the world exactly as it was.
This is the single most important behaviour on the surface, and it is not configurable. An assistant that acts on a confident misreading is worse than one that cannot act at all, because the misreading is invisible until somebody notices the consequence. The card is where the misreading gets caught by the only party who can catch it, which is you.
When it is not sure which one
Ask to move "the call with Ahmed" when there are two, and it shows you both rather than choosing.
Same principle. A wrong guess between two similar records is precisely the failure the confirm step exists to prevent, so ambiguity is surfaced rather than resolved by confidence.
This happens more than you would expect, particularly with recurring meetings, and the extra tap is cheap compared with moving the wrong one.
Analytics questions go somewhere else
Ask something genuinely analytical, the kind of question that wants a chart rather than a sentence, and instead of answering in prose it builds the thing and takes you to it.
This is deliberate and occasionally surprising. The reasoning is simple: a number in a chat message is a number you will re-ask for next week, and again the week after. A saved view is one you can open, share, and put on a dashboard.
If you genuinely wanted the sentence, ask more narrowly. A single figure asked as a single figure comes back as a single figure.
Speaking instead of typing
Everything above works over voice, including the changes.
Voice has no confirmation card, because there is nothing to tap. Instead the assistant reads the change back and waits for you to agree out loud before doing anything. The gate is identical, delivered in the medium you are using.
Listen to the read-back rather than agreeing over it. Transcription mishears names reliably, and the read-back is the only place you will catch it.
What a good request looks like
The confirmation card catches errors, but a clear request means fewer cards to reject.
Name the time, not the position. "Move my three o'clock" is clear. "Move my next meeting" is ambiguous the moment a meeting ends while you are typing.
Name the person, not the subject. "The call with Ahmed" resolves faster than "the pricing call", because people are indexed and topics are inferred.
Say the destination, not the delta. "Move it to Thursday at two" is unambiguous. "Push it a couple of days" invites a guess about which two days and what time.
One verb per request. "Approve Priya's expense and mark the pricing task done" is two changes, and it will produce two cards, so you may as well ask for them separately and read each one properly.
Where the confirmation actually saves you
Three failure modes it catches routinely, all of which look like nothing until they are not.
The wrong record of two similar ones, which is the common case with recurring meetings and with customers who have similar names.
The right record and the wrong day, usually because a relative date was interpreted from a different reference point than you had in your head.
A change that would have been correct yesterday. The world moves between your forming the intention and your typing it, and the card is drawn from current data rather than from what you remember.
None of these are exotic. All three happen within a fortnight of anybody using this seriously, and each one is caught by reading four lines.
What it will not do
It will not send anything on your behalf without you seeing it. Asking it to draft a reply puts the draft in your composer, exactly as if you had typed it. Sending remains yours.
It will not exceed your permissions. If you cannot approve something in the platform, you cannot approve it by asking, and it will tell you rather than failing quietly or pretending to succeed.
It will not chain changes. One request, one confirmation. Asking for three things produces three cards, which is slower and is the point: three changes confirmed in one tap is three chances to miss one.
It will not act on a schedule. Anything recurring belongs in AI Agents, which is configured deliberately and runs with its own permissions.
Where this is worth using
The honest answer is: when the change is small and the navigation is not.
Moving one meeting is four clicks across two screens. Saying "push my two o'clock an hour" is faster, and the card shows the same information you would have checked anyway before clicking save.
Approving a queue is not a good fit. Go to the queue, or better, use the Cockpit filter that shows only approvals.
Anything involving a judgement about a person is not a good fit either. The assistant does not know the relationship, and an approval is sometimes a conversation.
Two habits
Read the card, do not skim it. The whole value of the mechanism is destroyed by confirming reflexively, and after a fortnight of correct cards the temptation is real.
Ask for one thing at a time. Compound requests are where the ambiguity lives, and the extra ten seconds of typing costs less than the wrong meeting moving.
Did this answer your question?
No, ask a person