A few weeks ago, news broke that Meta employees were ranked on an internal leaderboard called Claudeonomics. The metric was simple: how many AI tokens did each person consume? Top performers earned titles like “Token Legend” and “Session Immortal.” Over thirty days, eighty-five thousand employees burned through roughly sixty trillion tokens. Forbes, the Wall Street Journal, and Fortune all covered the phenomenon, which acquired a name: tokenmaxxing. Maximize your AI usage, the thinking went, and productivity will follow.
A few weeks later, the leaderboard came down. Amazon quietly shut its own version after employees began running pointless agents to inflate their counts. Engineers wrote loops that did nothing but burn tokens. Uber’s COO admitted that the link between token spend and actual productivity simply was not there. Michael Burry called it “leaderboard-driven, management-mandated overconsumption.”
I read this story and felt a tug of recognition. Not because I work in AI, but because I work in learning. And anyone who has spent time in L&D will tell you we have been running our own version of this story for decades. Silicon Valley just gave it a name.
We have always called ours something else: completion rate.
The metric that wasn’t
Here is a familiar dashboard. The training platform shows ninety-four percent completion. Two thousand hours delivered. Engagement scores in the high eighties. The quarterly review goes well. The slide makes it into the all-hands deck. Everyone moves on.
Six months later, nothing in the business has changed. Sales calls sound the same. The new manager training did not produce better managers. The compliance course was passed by everyone and forgotten by most. The expensive leadership program is remembered as “those two days at the hotel.”
This is not a failure of learning. It is a failure of measurement. We measured activity because activity is easy to measure. We then mistook the measurement for the outcome it was supposed to represent.
There is a name for this mistake. It is older than L&D, older than AI, older than dashboards. It is called Goodhart’s Law: when a measure becomes a target, it ceases to be a good measure. The moment “completion rate” became something to chase, it stopped telling us anything useful about whether people had actually learned.
Why we do it anyway
I want to be honest about why activity metrics are so seductive in our field. They are not the result of laziness or bad intent. They exist because measuring real learning is genuinely hard.
You can count completions in a database query. You cannot count, in any clean way, whether a sales manager handles a difficult coaching conversation differently after a workshop. You can plot time-on-platform over a quarter. You cannot plot the slow rewiring of how a new engineer thinks about code review. The things we most want learning to produce are diffuse, delayed, and tangled up with everything else happening in the business.
So we substitute. We measure what we can measure. And in board meetings and budget reviews, the easy numbers crowd out the hard ones. A program with strong completion data survives. A program with messier evidence of real behavior change gets cut.
This is the same pattern that produced tokenmaxxing. Engineering output is hard to measure. Token consumption is trivial. The easy metric won, until it became the target, and then it stopped working.
What the research keeps telling us
McKinsey’s most recent State of AI report found that only about six percent of companies are capturing real financial impact from their AI investments. The thing those companies share is not heavier AI usage. It is that they have fundamentally redesigned how work gets done. MIT’s NANDA group found that ninety-five percent of enterprise AI pilots produce no measurable impact at all.
The numbers in L&D look uncomfortably similar. Industry surveys of corporate training effectiveness routinely find that only a small fraction of programs produce demonstrable behavior change. The bulk of learning investment evaporates somewhere between the LMS completion screen and Monday morning.
The honest reading of this evidence is not that learning does not work. It is that we have been investing in activity and hoping it would somehow turn into capability. We have built whole departments around the management of completions, hours, and engagement scores, none of which were ever the actual point.
A different question
What would it look like to measure what actually matters?
It starts with a different opening question. Not what training should we deliver? but what should be different six months from now, in terms a non-L&D person would recognize? Fewer escalations from new managers. Faster ramp time for new hires. A specific safety incident rate brought down. A specific decision made better, by specific people, more often.
This is not new. The Kirkpatrick model has been pointing at this for sixty years. The data, the tools, even the methods have been available for decades. What has been missing is something less technical and harder to fix: the system around us does not actually let us do this work.
A learning intervention designed to move a behavior is a different artifact from a learning intervention designed to be completed. The former is shorter, embedded in real work, supported by manager conversations, and evaluated by what changes in the workflow over months. The latter is a course. Most of us know the former is what works. Most of us cannot actually build it.
The reasons are not mysterious. Behavior-change programs require managers to coach, not just nominate. They require business leaders to define what “better” looks like in concrete terms, before the program starts, not after. They require finance to wait longer for the evidence. They require executives to accept that a smaller, slower intervention may produce more impact than a larger, faster one. They require HR systems that can track something other than seat time and scores.
None of this is the L&D function’s gift to give. We can design the right program, but we cannot, alone, build the conditions in which it will work. When the business asks for a leadership program by Q3, the manager has no time to coach, the budget review wants numbers in twelve weeks, and the LMS only reports completions, the rational response is to deliver something completable. We end up building courses not because we believe in courses, but because the system rewards what is countable.
This is the harder thing the tokenmaxxing story made visible. The Meta engineers gaming token counts were not stupid. They were responding rationally to a metric they had been handed. Pull the metric, and the behavior changes overnight. The same is true in L&D. As long as the board deck asks for completion rates, completion rates are what L&D will optimize for.
Changing this is not an L&D project. It is a conversation that has to happen with business leaders, with finance, with HR, with managers, with the people who decide what gets measured and what gets funded. The L&D function’s job is to keep raising the question. The system’s job is to be willing to hear it.
The conversation we have been avoiding
I have been in this work for fifteen years, and in the budget reviews I have been part of, a pattern keeps showing up. I come in with a story about behavior change, ramp time, capability building. The questions that come back are often about completion rates and engagement scores. The conversation tilts toward the numbers everyone in the room can hold in their head, and the deeper case gets compressed into a footnote. I am sure this is not every room, every time. But it has been enough rooms, enough times, that it shaped how I learned to plan and pitch.
This is not a failure of the L&D person, nor of the CFO, nor of the HR director. Each of them is responding to what their role rewards. The L&D person needs to defend a budget. The CFO needs comparable metrics across functions. The HR director needs something that looks like accountability. Completion rates serve all three, badly. But they serve.
The shift we need is not for L&D to be braver in the meeting at three o’clock. It is for the meeting itself to change. For the question on the table to change. For someone above the L&D function to say: I do not want a completion rate this quarter. I want to know what new managers can now do that they could not do six months ago, and I am willing to wait for a real answer.
That is the conversation we have been avoiding. And until it happens, the L&D function will keep optimizing for what the room asks for.
What I am taking from the tokenmaxxing story
The Meta and Amazon leaderboards came down quickly once the absurdity became visible. Engineers gaming token counts was too obvious to ignore. I doubt our completion-rate dashboards will come down with the same speed. The absurdity is too familiar, too embedded, too useful in the wrong meetings.
But I think the underlying lesson travels. More is not a strategy. More content does not produce more learning. More learning, in the hours-and-completions sense, does not produce better performance. The work is to figure out what better would actually look like, in concrete terms, in the lives of the people we are supposedly helping, and then to measure that.
This is harder. It is slower. It does not slide as cleanly onto an executive deck. It also cannot be done by L&D alone. It needs business leaders willing to define what they actually want, managers willing to coach, finance willing to wait, executives willing to ask different questions. The L&D function can advocate for the change. It cannot make the change by itself.
The tokenmaxxing story will fade. The next AI hype cycle is already starting. But the question it surfaces, are we measuring activity because it matters, or because it is easy?, is one every function in every organization eventually has to answer.
It might be time we asked it together.
Ilkem Kayican Dipcin is the supervisor of inspaire, an AI-powered learning design project and has spent 15 years designing learning programmes across higher education and corporate settings. This is “Learning That Feels Good,” a newsletter about what makes learning actually work: the science, the design, and the human experience behind it.
If this piece landed for you, send it to someone who’s caught between proving ROI and remembering why they got into learning in the first place
.