Overview
How might we design reversibility marking so people can trust and act on AI output?
When to use
- Essential for agent interfaces, workflow automation, and any system that performs actions on behalf of users where surfacing reversibility helps calibrate approval attention to real risk.
When to skip
- Flows with only reversible local edits where every tag would say the same thing.
- When you cannot honestly classify reversibility and labels would mislead.
- Emergency break-glass tools where speed beats taxonomy.
Rules
Uniform primary buttons for send, delete, and tweak draft.
Calling something reversible when undo windows are tiny or unreliable.
Hiding irreversibility in tooltip copy only.
Marking irreversible actions lightly because they are “rare.”
Evidence
| Product | Implementation |
|---|---|
| ChatGPT / Gemini agents | Stronger confirmation chrome for send and external actions vs drafts. |
| Cursor Agent | Diff review emphasizes file-mutating steps before apply. |
| Banking / payments AI | Irreversible transfer confirmations distinct from simulation steps. |
| Email agents | Draft vs send clearly separated with send as irreversible. |