← Playground
04Interactive demoReversibility

Undo

You’re about to create an AI assistant’s permissions policy: which actions it may take on its own, and which wait for a person. One decision per row.

Judge each action however you like. The grid grades your policy when you’re done.

The premise

An agent never knows which of its actions was wrong.

Its own log looks clean the whole time. Humans can and should be able to catch those mistakes: a ping on someone’s phone, a line on a statement, a confused reply. So autonomy shouldn’t be granted on accuracy. It has to be granted on what you can take back.

Below: ten tasks an assistant agent can handle. Decide on each one.

The permissions review
0 of 10 decided
ActionYour call
Draft the replyEverything. The draft disappears.
Send the emailNothing. It’s delivered.
Sort the inbox into foldersEverything. Move them back.
Charge the client’s cardThe money. Refunds post in seconds.
Add a hold to your calendarEverything. Delete the hold.
Delete a message in the team channelThe message is gone, as asked.
Reorder the task queueEverything. Drag it back.
Update the shared docVersion history rolls it back.
Archive last quarter’s filesEverything. Unarchive.
Cancel the 3pm meetingThe event. Re-invite everyone.

Your grading happens in this page and stays here: nothing is sent, scored remotely, or stored. The rows are ordinary assistant permissions, not any particular product’s.