Over the last two years, every people leader I know has been chasing the same number: what percent of our employees are using AI?

It’s the wrong metric.

At Atlassian, we learned this through firsthand experience. We’ve gone through three evolutions of how we measure employee AI adoption, and each time we changed the metrics we learned something telling about what the old numbers were hiding.

Those shifts have changed how we hire, develop, and bet on talent. Here’s how our thinking evolved, and what it means for you.

Phase 1: “Are they touching AI?” (Weekly active users)

When generative AI first landed in the enterprise, the only question anyone could answer was the simplest one: did this employee log in to an AI tool this week?

Weekly active users (WAU) was a useful starting point. It told us whether our employees were trying the tools. It did not tell us whether anything was changing.

You can hit 95% WAU and still have a workforce that uses AI the way it uses spell-check: occasionally, passively, and without ever rethinking the work itself. While we saw our WAU percentage was rising, we quickly realized that increase gave us no signal about actual value.

So we raised the bar.

Phase 2: One company-wide “superuser” threshold

In the second half of 2025, we defined a superuser as anyone using AI tools more than ~40 times a week — the 90th percentile of company-wide usage from the prior half. We set one goal: double the percent of superusers within six months.

It worked, sort of. Superusers increased from ~14% in July to 34% by December. We blew through the target.

But we quickly saw that a single company-wide bar penalized everyone who wasn’t an engineer. Engineering’s volume of AI interactions is, by the nature of the craft, an order of magnitude higher than a finance analyst’s or a recruiter’s.

We were also still overly focused on volume. A salesperson who pasted 50 prompts into an AI chat cleared the bar. A lawyer who used one AI agent twice a week to redesign how contracts get reviewed did not. The metric was getting in the way of the truth.

So we changed the definition again.

Phase 3: A superuser is the top of their craft — measured by habit, not headcount

Today, an Atlassian AI superuser is defined per function. Each craft — Engineering, Design, Product Management, Marketing, Legal, Sales, Finance, People — has its own 90th-percentile baseline of weekly AI interactions, set from January 2026 usage data. The bar moves with the work.

A superuser engineer in our CTO org averages more than 270 AI interactions a week. A superuser in finance or legal looks completely different — different tools, different cadence, different signature. What they share is not a number. It’s a position: the top ~10% of how AI is actually being used in their function.

Our company-wide score is the average of each function’s score. That’s deliberate. It stops the largest org from drowning out the rest, and it forces every craft to lift its own ceiling. Our target for the second half of this fiscal year is for every function to double its January baseline.

This is the strategy I’d encourage every CHRO to adopt — or at least to argue with. Not because we’ve landed on the perfect metric (we haven’t, we’re already reworking it for next year), but because the underlying approach holds: measure by function, set the bar relative to peers, and keep raising it as your organization matures.

Why this matters for talent strategy

Evolving our AI performance metrics was never just about the numbers. It’s about what the company believes AI adoption is for.

  • If you believe AI adoption is about compliance, you’ll measure WAU and celebrate when the number goes up.
  • If you believe it’s about engagement, you’ll measure usage volume and run training programs to push the average up.
  • If you believe it’s about transformation — that AI’s value comes from people who actually redesign how work gets done — you have to find and develop the small group of employees who are already doing that, in every function.

Those people exist in your company today. We’ve watched them in ours. They aren’t always the most senior, the most technical, or the loudest. They’re the lawyer who built a self-serve intake bot for partner teams. The finance analyst who turned six disconnected systems into a natural-language data hub. The recruiter who designed a pipeline scorecard with an AI thinking partner instead of yet another spreadsheet. They sit in glue roles. They redesign workflows. They share what they learn.

They are your AI strategy. Most companies just don’t know how to find them yet.

A better metric still only tells you who. We needed to know what they were doing.

Once we could reliably identify superusers in every craft, our Teamwork Lab — Atlassian’s in-house research team studying how people work with AI — went looking for the answer to a harder question: what do these people actually do that the rest of the workforce doesn’t yet?

We asked superusers across functions, levels, and crafts to record short demo videos walking through how they use AI in their day-to-day work, and we paired those demos with our internal adoption data. We expected to find one or two clean archetypes. We didn’t. The work patterns were more diverse than we imagined, and there is no single AI superuser persona to hire for.

What did emerge — consistently, across functions — was a small set of common behaviors and the underlying traits that produce them. Those are what HR and people leaders should be hiring and developing for.

Five things AI superusers actually do

  1. They redesign the work, they don’t just use AI a lot. Superusers break tasks into AI-friendly and human-only pieces. They use AI to generate options, surfaces, and drafts, then spend their human time on judgment and refinement. They build reusable scaffolds (agents, prompts, checklists, templates) instead of treating each task as a one-off. For example, one of our finance superusers turned data from six disconnected systems into a self-serve hub that answers natural-language questions, so analysts now spend their time on interpretation instead of data wrangling.
  2. They build lightweight systems for others. They share agents, prompts, and workflows with teammates, demo their processes openly, and adapt how they use AI to how different teammates like to work. They become the “AI champion” others go to when they’re stuck. One of our talent acquisition superusers used an AI thinking partner to design a pipeline health scorecard that managers across the org now reuse, instead of building a custom spreadsheet for every role.
  1. They’re comfortable being seen mid‑thought. They don’t just show the finished product; they show the wrong turns, dead ends, and course‑corrections that got them there. By explaining what didn’t work and why, they pass their learnings on so others can experiment more safely.
  2. They don’t confuse “first draft” with “done.” They use AI to move faster, but they never treat the first output as fact. They scan for gaps, errors, and tone mismatches, and they’re explicit about what’s rough AI versus production‑ready. They spot and delete obvious AI slop — the low‑quality drafts a less‑skilled user might pass along — taking responsibility for the final call.
  3. They show up in unexpected roles and at unexpected levels. Superusers are not concentrated in technical or AI-specific roles, and they aren’t always the most senior person in the room. They’re curious problem-solvers who enjoy teaching themselves new tools and experimenting with how work could run differently. Many of them sit in “glue” roles connecting teams or processes, where they see end-to-end pain that more specialized colleagues miss.

Three traits worth hiring for

Behind those behaviors sit three underlying traits we now believe are predictive of who will become an AI superuser in your organization:

  1. Pattern-seeking pragmatism. AI superusers notice recurring friction in how their team works and ask, “How could AI take this out?” They run small, concrete experiments instead of waiting for a perfect AI program or perfect data.
  2. Bias for sharing, not gatekeeping. When they hit on something that works, their first instinct is to teach it. They turn personal AI wins into shared playbooks, demos, and templates — multiplying impact beyond their own role.
  3. Comfort with imperfection. They don’t expect AI to be right, and they don’t treat the first answer as the only one. They sift, push back, and ask for more, sparring with AI for different angles and better options. They treat imperfect outputs as raw material, not verdicts, and they make it safe for others to learn in public the same way.

If you’re interviewing or developing for AI talent today, these are more useful signals than tool fluency or certifications. Tools change every quarter. These traits don’t.

Three things I’d encourage you to do this quarter

  1. Stop reporting WAU to your board. Replace it with a percentile-based definition that’s specific to each function. Your sales team and your legal team should not share a bar.
  2. Make it safe to be mid-thought with AI. The employees who transform their work are the ones willing to show messy drafts, narrate why they kept or rejected an AI output, and invite critique early. That culture is a leadership choice.
  3. Find your superusers and study them. Not just to celebrate them, but to learn what they do that the rest of your workforce doesn’t yet. Hire and develop for those behaviors, not for AI tool fluency.

We’ve changed our definition of an AI superuser three times in 18 months, and we’re already evaluating it again for the coming year. The goal isn’t to land on a perfect metric. The goal is to keep asking sharper questions, and to build a workforce strategy around the answer.

— Alicia Lenart is VP of HR at Atlassian