Your AI Usage Is Being Ranked. Here's What That

Picture your name on a leaderboard. Not for sales numbers or shipping velocity—for how many times you talked to a chatbot last week. You're 174th out of 250. Your manager can see this. HR can probably see this. You don't know who else can see this, or for how long it's stored, or whether it's feeding into your next performance review. You just know you're 174th.

That's not a hypothetical. According to reporting by The Information, Meta has been running something like this internally—a ranked list of top AI power users drawn from across its workforce, complete with titles for those who land near the summit. The sourcing on the specific details (the exact headcount covered, the precise titles awarded) traces back to anonymous sources rather than official confirmation, so treat the granular specifics with appropriate skepticism. But the underlying practice—individual, named, ranked AI consumption data used as an internal performance signal—is documented across multiple companies. Business Insider reported on a similar AI adoption dashboard at Disney. The Financial Times reported that Amazon employees were automating non-essential tasks specifically to inflate their usage scores. Kevin Roose at The New York Times reported that Meta and Shopify managers have begun factoring AI use into formal performance reviews.

The debate this has kicked off is mostly framed as a productivity debate: are token leaderboards a smart way to drive AI adoption, or are they a fast track to Goodhart's Law, where the metric becomes the game and stops measuring anything real? That's a legitimate question. It's not the one I keep getting stuck on.

The thing that productivity discourse tends to skip is the data architecture you have to build to run a token leaderboard in the first place.

To rank employees by AI consumption, you need granular, persistent, individually attributed usage logs. You need to know not just that a user ID made 4,000 requests last month, but that the user ID maps to a specific person with a specific job title reporting to a specific manager. You need that data aggregated, stored, and surfaced in a format that a manager or an HR system can read. And once you've built that infrastructure—once you've established that individual AI interaction data flows upward in the org—you've made a set of architectural choices that are very hard to walk back.

What are people actually doing with AI at work? That question sounds innocuous until you remember that AI interactions can reveal a lot: what you're struggling with, what you're trying to learn, what work you're automating versus doing manually, what you're anxious enough about to ask a machine instead of a colleague. Aggregate enough of that, attach it to a name and a job title, and you have a behavioral profile that goes well beyond "this person used 40,000 tokens on Tuesday."

I'm not saying companies are doing anything nefarious with this data right now. I'm saying most of them haven't written down what they will do with it, or for how long they'll keep it, or who internally has access, or whether it could surface in a performance improvement plan, a layoff decision, or a legal dispute. That's not a conspiracy theory. That's just how enterprise data works. Data collected for one purpose gets used for adjacent purposes, because it's there and it's useful and nobody explicitly said not to.

If your company just rolled out an AI usage dashboard, the questions worth asking HR are: What data is collected at the individual level? Who can access it? How long is it retained? Does it feed into performance systems? Is it covered by your existing data governance policies? You probably won't get complete answers. But asking establishes that employees are paying attention, and that matters.

Now, the productivity argument. Because it's real, and it deserves more than a dismissive wave.

The companies deploying these leaderboards are responding to a genuine problem. We are in the middle of a shift from AI as a productivity assistant—something you use to do your existing job faster—to agentic AI, where your job increasingly involves setting up conditions for AI systems to do things autonomously. That shift is not incremental. It requires experimentation that most enterprises don't have a culture for. People have to try things, build half-baked agents, abandon them, try again. The Financial Times Amazon story is being read as evidence that token leaderboards are broken. It's also evidence that people are engaging—messily, imperfectly, but engaging—with AI tooling in ways they weren't before the leaderboard existed.

May Habib, CEO of Writer, the enterprise AI company, put the pressure companies feel about as directly as I've seen: "It's existential for us. We are in the most competitive space that has ever existed and will ever exist." That's the honest version of why this is happening. The companies that figure out agentic AI deployment first aren't just going to be more productive—they're going to eat the ones that don't.

The token leaderboard is a blunt instrument in service of that goal. And Goodhart's Law is real: as the Financial Times documented, when Amazon employees knew managers were tracking AI usage, some of them started running the tool on non-essential tasks just to push the number up. That's not a technology failure. That's a metrics failure. It's what happens when you measure effort instead of outcome—you get performed effort.

The stronger argument for token leaderboards isn't that high consumption equals high productivity. It's that in a period where nobody knows what good AI adoption looks like, high consumption is at least a signal that someone is trying. The organizations with the capability overhang problem—the gap between what AI can do and what their employees are actually doing with it—aren't going to close that gap by leaving people alone. Some form of nudge matters. The question is whether a public, ranked leaderboard is the right nudge, or whether it's the nudge that generates the most surveillance exposure for the least actual learning.

CNBC's Deirdre Bosa has argued that token consumption "has decoupled from actual economic value. Like page views during the dot-com era, justify the spend until they don't." That's a fair concern about aggregate demand metrics. It's less convincing as a critique of individual company adoption programs, where the goal isn't to justify Anthropic's valuation but to figure out which workflows actually get better with AI in them.

The viral Slack screenshot that circulated recently—purportedly praising a $600 Anthropic spend while chiding someone for a $23 Uber Eats order—almost certainly wasn't real. The resonance was. People recognized it immediately as the kind of thing that could happen, which tells you something about where corporate AI culture has landed.

Here's where I land, and I want to be direct about it because I think the productivity-versus-Goodhart debate is genuinely obscuring something important.

Token leaderboards are a symptom of a measurement problem that companies are going to keep running into as AI gets more deeply embedded in work. The underlying impulse—we need to know if people are actually using this—is reasonable. The implementation—rank everyone by consumption, make it visible, tie it to performance signals—creates surveillance infrastructure that outlasts the specific AI adoption moment it was built for. The leaderboard might sunset in a year. The individual usage data it generated probably won't.

Companies that want to drive real AI adoption without building a panopticon have better options: anonymized usage data that shows organizational patterns without naming individuals, outcome-based measurement tied to actual work products, opt-in communities where people share what's working. These require more effort than a leaderboard. They also don't require employees to hand over a behavioral log in exchange for not being 174th.

The real design question isn't whether to measure AI adoption. It's whether the measurement needs to be individual, named, and ranked to work—or whether that specific design is just the easiest one to build.

I'd like to see someone at one of these companies answer that question out loud.

Rachel "Rach" Kovacs covers cybersecurity and privacy for Buzzrag.