What are the four action classes?

Reversible and visible: let it run, with a cheap undo next to the result. Reversible and invisible: surface what happened where the owner will see it. Irreversible and visible: confirm, stating specifics. Irreversible and invisible: do not automate it in that form. They come from the chapter Downside Exposure for Interfaces in my design handbook.

Where should the action class live?

On the step, in the product's configuration, not in the system prompt. A prompt line that says to be careful with money asks the agent to rate the impact itself, which is the judgment that did not transfer in the first place.

Which reused patterns are safe as they are?

In my classification, a mention answered in a thread everyone can read is reversible and visible, so it can just run. A ticket assignment changes fields quietly, so it needs a visible account of what changed. A step that publishes publicly needs a confirmation. A step that approves money nobody re-reads should not go to an agent as built.

What question should design review add?

What was the human doing in this seat that the pattern never said? Atlassian's own question covers the case where the agent behaves differently from a human. This one covers the case where it behaves the same and the human was quietly doing more.

Should AI Agents Use the Same UI Patterns as Humans?

Should AI agents use the same UI patterns as humans?

THE SHORT ANSWER

Mostly yes. Atlassian's design review rule, as Jason Lemkin reports it from Sherif Mansour's October 2026 SaaStr AI session, is that nine times out of ten an agent should follow the pattern already built for humans: you can mention it, assign it a ticket, and add it to a workflow step. The exception is a seat where the human was exercising judgment the pattern never recorded. I put an agent in a human approver's seat, auto-approving expenses under $500 with a matching receipt. Approvals went up 40%, about $2,400 was approved that should not have been, and finance turned it off within the month. The pattern transferred and the approver's sense of which requests deserve a second look did not. Jakob Nielsen reports a study in which the best 2024-era model matched human ratings of an action's impact 58% of the time, so the agent should not be the one grading consequences. Reuse the human pattern, then write an action class on the step: reversible or not, and visible or not.

Reusing a human pattern for an agent is the fast path and usually the right one. The seat already has permissions, history, and a place in the workflow.

The failure I've had with it wasn't the agent doing something strange. It did exactly what the human pattern described. The approver who used to sit there had also been noticing things, and that part was never written down.

Atlassian's rule, Nielsen's number, both of my failed agents, and the table that maps mentions, assignments, and workflow steps to the four classes are in Agents Fit the Human Pattern Nine Times in Ten. Design the Tenth. The classes themselves are from Downside Exposure for Interfaces.

Last reviewed 2026-10-13 · 1 min read

Should AI agents use the same UI patterns as humans?

SOURCES

THE LONG VERSION

RELATED ANSWERS