The approval card

An agent that can act on your accounts is only useful if you trust it, and you can only trust it if refusing is easy, fast and obviously safe. Most of the design work in an agent platform is not in the acting. It is in the stopping.

What a good approval shows

• The exact action, in the verb the user would use. Send. Publish. Merge.
• The exact target. This address, this channel, this branch.
• The exact content, rendered as the recipient will see it, not summarised.
• What happens if you do nothing, which should always be nothing.
• What the system was trying to achieve, so an odd looking action can be understood rather than only rejected.

The summarised version is the tempting one and the wrong one. "Send a follow up to 14 prospects" is not an approval, it is a description. If the approval does not contain the words that will be sent, the person clicking it has not approved anything.

Defaults matter more than warnings

A confirmation dialog that appears on every action is trained away within a week. The right approach is to gate on blast radius: reversible and small goes through, outward facing and irreversible stops. Then the stop means something because it is rare.

Nothing sends, publishes or merges without you. That sentence is only worth writing if the product makes refusing genuinely cheap.

The test we use: could someone approve fifty of these in a row and still be reading them? If not, either the batching is wrong or the gate is in the wrong place.