April 29, 2026
Many companies have started doing something unusual. They have started publishing weekly snapshots of how employees use AI. It often comes as a leaderboard where the employees are ranked. Or a personal email telling you how you used AI last week.
There is something very familiar about this whole exercise. Like a cover song you can’t quite remember.
For a long time, the standard unit of employee productivity was presence. Think about factory work, where hours logged were directly proportional to the output. It takes 30 minutes to produce a product. You work for 9 hours, and you produce 18 products that day. You tap in at the beginning of the day, and then tap out when it is time to head home. If the floor manager sees you near your machine, he knows you are working.
The knowledge economy broke that logic completely. When the knowledge economy became the dominant economy, managers (with the old way of thinking) started losing sleep. Your output is no longer directly proportional to your input or hours logged. A developer who stares at the ceiling for 40 mins, then ships a feature in 3 hours, is not less productive than someone who takes 6 hours or a night of coding. This should have forced a harder question: what does productive work actually look like? Most companies don’t have this figured out. So they relied (and still do) on the hours logged metric.
In Hinduism, Vishnu had ten avatars across the yugas (eras). Each one suited to the crisis of its age. Bad management has its own avatar tradition. The factory foreman becomes the incompetent knowledge-work middle manager, who becomes the incompetent AI-adoption evangelist. Same underlying worldview. Different costume.
Now that GenAI has arrived, a similar pattern is emerging. The companies are asking themselves one question: “How to gauge if the employee is actually using GenAI?” and a similar suboptimal solution is being developed. Let’s start counting the tokens consumed. Prompts fired. And let’s put it on a leaderboard to encourage more people.
Jensen Huang recently suggested that if a $500,000 engineer isn’t consuming $250,000 worth of AI tokens annually, something is wrong. There is a version of this argument that makes sense. Powerful tools should be used. Under-leveraging AI is a real problem. But the moment you turn that into a numeric expectation and tracking, you are back in Goodhart territory.
There’s a concept in economics called Goodhart’s Law. The short version: when a measure becomes a target, it stops being a good measure. The moment you tell people that token consumption is what gets rewarded, token consumption is what you’ll get. Not better thinking. Not faster shipping. Not cleaner decisions. Just more prompts. Longer conversations.
Honestly, you can’t blame the people gaming the system. When you set wrong incentives, it is bound to happen. This is a failure that is creeping from the top.
Measuring the real impact of GenAI at work is doable. But it is hard. You have to go back to your boardroom, look at old JIRA tickets, and analyse that. What was the average time to ship features with X degree of complexity in the past? How much time is it taking now? What used to take three weeks, does it take ten days now? What used to need four revisions, does it need two? That kind of measurement is doable. It just requires intellectual rigour and some patience, and most organisations are running low on both.
You can do this outside the engineering work too. Are decisions getting made faster? Are they getting reversed less often? Are first drafts closer to final outputs? Are we spending less time in meetings trying to figure out what to do next? The focus on “Are people using AI more?” is the wrong thing to do. The better question is: is the cost of getting to clarity going down?
Vishnu, at least, had a purpose for each avatar. The form changed because the problem changed. The incompetent manager’s avatars share no such logic. The problem keeps changing. The solution stays the same: find a number, trust the number, reward the number.
The tap-in, tap-out in the knowledge economy never really measured work. It measured the anxiety of managers who didn’t know how else to look. Tokenmaxxing is just the same anxiety, wearing a different UI.