Agents Fit the Human Pattern Nine Times in Ten. Design the Tenth.

Most teams adding agents to an existing product are doing the sensible thing. They take what already works for people, an assignee field, a mention, an approval step, and let an agent occupy it.

Atlassian did this at scale and came out with a rule. I think the rule is right, and I have an old failure that says where it stops.

The short version

Atlassian's design review rule, as Jason Lemkin reports it from Sherif Mansour's SaaStr AI session, is that an agent should follow the pattern already built for humans nine times out of ten. Jakob Nielsen reached the same primitive from undo the same week: treat the AI as another user. Both hold. The tenth case is any seat where the human was carrying judgment the pattern never recorded. I put an agent in a human approver's seat for expenses under $500 and approvals went up 40% before finance turned it off. Nielsen reports a study in which the best 2024-era model matched human ratings of an action's impact 58% of the time. So when an agent inherits a human's step, write the action class on the step: reversible or not, visible or not. Then let the agent run inside it.

What Atlassian found

The source is Lemkin's SaaStr write-up of Mansour's session. I haven't watched the session, so everything here is as Lemkin reports it.

Mansour has been at Atlassian 17 years and runs AI across more than 20 apps. About two and a half years ago the team took the existing Jira automation designer and added one box: use a Rovo agent at this step. A literal bolt-on. Customers took it further than expected, adding branches and chaining agents. One pattern he described: a ticket comes in, an agent picks it up, a marketing agent calls Canva to make assets, and a social agent reads them and posts.

About six months ago the team landed on a principle that came out of that work. Every problem they had to solve for agents was one they'd already solved for humans. People need tools, context, goals and accountability, and visibility into what teammates are doing. Agents need the same four. So they went through the portfolio:

• You can @mention a human. Now you can @mention an agent.
• You can add a human to a step in a Jira workflow. Same for an agent.
• You can assign a ticket to a human. Same for an agent.
• You can chat with a human in Slack or Teams. The agents are there too.

And the review rule. When a team shows something an agent does that a human couldn't or wouldn't do in the product, ask why it's different. Sometimes there's a good reason. Nine times out of ten, per Mansour, the agent should follow the human pattern.

One of the mistakes on his own list is taking about two years to see this.

Nielsen got there from undo

Two days later Jakob Nielsen's roundup arrived at the same primitive from a different door.

His subject was undo. With an agent editing alongside you, "undo last" may reverse the agent's work when you meant your own. He points out that groupware researchers solved this in 1994 with per-user undo, and gives the instruction in six words: "treat the AI as another user."

Two people I read, one week, same conclusion. The agent is another user. Give it the seat a user would have.

Nielsen's piece also carries the limit. He writes that agents misjudge their own impact, and reports a study by Zhuohao Jerry Zhang and co-authors at Apple in which the best 2024-era model matched humans' impact ratings for phone UI actions only 58% of the time. I haven't read the study. It's about phone interfaces and a model generation that's already old. I'd still write the number on the whiteboard, because of what happened the one time I skipped this.

The seat I gave away

Number 1 in 10 AI Agents I Built That Failed. The Honest Retrospective. is an expense agent.

It read expense reports submitted in Slack, checked them against policy, and auto-approved anything under $500 with a matching receipt. Finance had been doing that review by hand, and the idea was to save them 10 hours a week. Same step, same policy. I put an agent in the seat.

It worked for two weeks. Then one of my engineers figured out that "dinner with customer" and a vague receipt passed every time. Approvals went up 40%. Spend followed. About $2,400 was approved that shouldn't have been, and finance turned it off within the month.

Look at what transferred. The step. The threshold. The policy text. All of it.

What didn't transfer was never in the pattern. A human approver who sees the fourth "dinner with customer" from the same person in a month slows down. Nobody wrote that into the workflow, because nobody had to. It lived in the approver.

Number 4 in the same post is the expensive version. A pricing agent, sitting where a human analyst would sit, adjusting prices inside a band. It anchored 14% below our benchmark and cost about $180,000 in annual recurring revenue from one cohort. Same shape. The seat moved. The flinch didn't.

That retrospective's own tally was that two of the ten failed from cost-compounding autonomy: an agent given authority to act, with no human gate. These are the two I'd point at, and both were agents doing a job a person had done.

Write the class on the step

So the tenth pattern has a definition. When an agent takes a human's place in a step, the step has to carry the judgment the human carried in their head, and the agent's own rating of impact isn't reliable enough to carry it.

The instrument is already in the design handbook. Downside Exposure for Interfaces sorts every action by two questions. Can it be undone? Would the user notice if it went wrong? Four classes come out, each with its own posture.

Here's how I'd apply them to the patterns Atlassian lists. These classifications are mine, not theirs.

The last row is my expense agent. Under this sort it never ships as built. An approval nobody re-reads and nobody can claw back cheaply is the one class that doesn't get an agent in the seat.

Two details matter in practice.

The class goes on the step, in the product. Not in the system prompt. A prompt line saying "be careful with money" asks the agent to do the rating, which is the thing Nielsen's number says not to lean on.

And the human pattern usually has a place for it already. A Jira workflow step has properties. An assignment has permissions. That's the pleasant consequence of Atlassian's rule: if the agent sits in a human's seat, the seat has configuration you can extend.

The second review question

Atlassian's question for design review is a good one. If the agent does something a human couldn't or wouldn't do here, why is it different?

I'd add one for the other nine cases, the ones where the agent does exactly what the human did.

What was the human doing in this seat that the pattern never said?

Most of the time the answer is nothing, and the pattern transfers clean. When the answer is "deciding which ones to look at twice," you've found the tenth.

I've made a related argument about the gates that already exist in Kill 'Human in the Loop' as a Control. Keep a Gate Ledger. That one measures the human gates you have. This one is about the gate you removed without noticing, because it was never drawn on the workflow.

It's also a small piece of the Enterprise AI Agents argument, that what counts is how many agents still do production work after ninety days. My expense agent lasted one month. It didn't fail on capability. It failed in a seat whose real job description was longer than the one I gave it.

This week: pick one step in your product where an agent now sits where a person used to. List the actions it can take there. Put each in one of the four classes. If any land in the bottom row, that's the redesign.

Related answer: Should AI agents use the same UI patterns as humans?

Sources: Jason Lemkin, "Atlassian's Head of AI: We Bolted AI Onto 20+ Apps, Almost Didn't Ship Chat, and Flipped Our Hiring Toward Juniors," SaaStr, October 10, 2026, read in full; Sherif Mansour's session is cited as Lemkin reports it. Jakob Nielsen, "UX Roundup: CAPTCHA, Undo, Robotaxi Safety, and more," UX Tigers, October 12, 2026, read in full; the study by Zhuohao Jerry Zhang and co-authors and the 1994 per-user undo work by Atul Prakash and Michael Knister are cited as Nielsen reports them. 10 AI Agents I Built That Failed. The Honest Retrospective., falkster.com, May 4, 2026, for the expense and pricing agents.