What is the difference between a high-performing leader and someone who is simply good at performative leadership? That question is becoming more urgent in the age of AI. The latest game in the surreal, parallel universe of big organizations is tokenmaxxing: the attempt to appear highly AI-enabled by generating, consuming, or reporting large volumes of AI usage. This is not really a story about tokens. It is a story about incentives.
Tokens, in this context, are the units of data processed by AI models. In theory, they can be a useful indicator of whether people are experimenting with new tools and embedding them into work. In practice, they can very quickly become another proxy for productivity that people learn to game.
Let’s be honest: a lot of office work is performative. Not because people are foolish, lazy, or cynical, but because organizations are extremely effective at teaching people which behaviours matter. If leaders put enough emphasis on certain visible norms, employees will adopt them, often at the expense of the underlying outcome those norms were supposed to support. If you reward being seen at the office, people will stay late. If you reward responsiveness at all hours, people will send emails at midnight. If you reward AI usage, people will find ways to use AI, whether or not their work actually improves.
For years, one of the most powerful proxies was stamina. Being first in and last out became a shorthand for dedication, especially in industries like financial services, consulting, and law. There was a time when exhaustion itself seemed to carry moral authority. The person still at their desk at 11 p.m. was assumed to be more serious than the person who had designed their work well enough to leave at a reasonable hour.
Eventually, the costs of work stamina culture became too obvious to ignore: burnout, poor judgment, brittle teams, and a generation of younger workers increasingly unwilling to treat chronic overwork as a badge of honour. Several firms, including JPMorgan Chase and Bank of America, have implemented or considered a 80-hour weekly cap on junior banker hours, particularly when not working on active live deals. In China, the infamous “996” model — 9 a.m. to 9 p.m., six days a week — helped spur the “lie flat” movement, as exhausted young people rejected the premise that ambition required the total surrender of the self.
Sadly, the new badge of commitment seems to be how visibly, frequently, and enthusiastically someone appears to be using AI. At a high level, some of this makes sense. When Jensen Huang, CEO of NVIDIA says that he would be ‘deeply alarmed’ if his $500,000 engineer did not consume at least $250,000 of tokens - he is not encouraging his employees to burn tokens to drive demand for his chips. He is talking about leverage. An expensive software programmer who is not using AI to increase output is leaving productivity on the table.
But a recent story, reported by the Financial Times, shows how quickly that logic can get out of control. Some Amazon employees have apparently been using an internal AI tool called MeshClaw to automate unnecessary or non-essential tasks in order to increase their AI token usage. The tool allows employees to create AI agents that connect to workplace software and act on their behalf. Amazon had reportedly introduced targets for more than 80 per cent of developers to use AI each week and had begun tracking token consumption on internal leaderboards. The company has said these statistics are not used in performance evaluations. But several employees told the FT they believed managers were watching the numbers.
Amazon are not alone in their zeal for token usage. The NYT reports that managers at Meta and Shopify are also factoring workers’ token consumption into their performance reviews. Leaders at Google have informed employees in non-technical roles that they are also expected to use more AI in their workflows.
A company investing heavily in AI infrastructure naturally wants to know whether its people are actually using the tools. Adoption matters. Experimentation matters. New habits do not emerge by magic. But the moment usage becomes a scoreboard, it stops being merely diagnostic. It becomes social. It becomes political. It becomes a new way to signal that you are modern, ambitious, adaptive, and aligned with the future. In other words, it becomes leadership LARPing.
There are two ways this trend could play out. The first is that we drift into a newly rebooted Taylorism, updated for the AI age. Frederick Winslow Taylor’s system of scientific management was built on the idea that work could be broken down, measured, standardized, and optimized. In its original form, it was about time-and-motion studies, stopwatches, factory floors, and the separation of thinking from doing. Managers designed the system; workers executed it. Efficiency was pursued through observation, measurement, and control.
The AI-era version will not necessarily look like a foreman with a stopwatch. It will look like a system that measures prompt counts, agent invocations, token consumption, keystroke analysis, meeting analytics, automated performance summaries, and leaderboards for behaviours that may or may not correlate with meaningful work. It will be sold as productivity intelligence. Some of it may even be useful. But without judgment, trust, and a clear connection to outcomes, it risks turning AI from a tool of augmentation into a tool of managerial authoritarianism.
That would be a tragic waste. The point of AI should not be to make humans easier to monitor. It should be to make work more intelligent, more creative, more adaptive, and, ideally, less bloody stupid.
The second path is more interesting. It requires us to recognize that trends like tokenmaxxing are not primarily technology issues. They are organizational design problems. They happen when the incentive system rewards the appearance of progress more than progress itself. People are being asked to transform work while still being evaluated through old structures. When roles, reporting lines, budgets, and compensation models remain rigid, new tools fail to make meaningful impact.
At too many organizations, people are not rewarded for improving the system. A person may see an opportunity to automate a broken process, improve a customer experience, or build an AI-enabled workflow that benefits three other teams. But if that work falls outside their job description, their P&L, or their manager’s immediate priorities, they may be discouraged from doing it. In some cases, they may even be penalized. The organization says it wants innovation, but the incentive structure says: stay in your lane.
This is why crude AI adoption metrics are so dangerous. They measure the easiest part of the problem. They can tell you whether people are touching the tools. They cannot tell you whether the organization has been redesigned to benefit from them.
Management theorists used to talk about principal-agent problems: the misalignment that occurs when one party, the agent, is supposed to act on behalf of another, the principal, but has different incentives or better information. The result is agency cost: decisions that are rational for the individual but inefficient for the organization. Now, it is not only humans acting as agents inside organizations. It is also the AI agents those humans design, deploy, and supervise. These agents may be capable of moving across systems, summarizing information, drafting code, triaging email, monitoring deployments, or coordinating workflows. But they remain trapped inside organizational structures built for a slower, more human-scaled world.
The result is a new kind of agency problem. We may have increasingly powerful digital agents operating inside organizations that have not resolved the human incentive problems around them. The tool can move faster than the structure. The agent can generate more possibilities than the organization has permission to absorb. The technology can expose inefficiencies that no one is rewarded for fixing.
This is why one of the most important emerging roles is not only the forward-deployed engineer, but the work architect.
A work architect understands how work actually happens: where decisions are made, where information gets stuck, where accountability becomes blurry, where process has accumulated because no one has had the authority or patience to remove it. They know enough about technology to see what AI can do, but enough about organizations to understand why capability does not automatically translate into impact.
In writing my new book, Abundant Intelligence, I’ve had many interesting conversations with leaders about the role of work architects. Tracey Franklin, Chief People and Digital Technology Officer at Moderna, for example, believes that the real opportunity is not merely to add AI to existing workflows. It is to rethink the workflow itself. What is the outcome we are actually trying to achieve? Which parts of the process require human judgment? Which parts require speed, memory, coordination, or pattern recognition? Where do we need an agent? Where do we need a better interface? Where do we need clearly articulated decision rights? Where do we simply need to stop doing something that no longer serves a purpose?
That is work architecture. It is part systems thinking, part organizational design, part product management, part anthropology. It is the discipline of asking not “How do we get people to use AI more?” but “How should this work now that AI exists?”
The distinction is crucial. A smart leader obsessed with AI usage will ask how many tokens were consumed. A transformative leader obsessed with cognitive leverage will ask what changed. Did cycle time improve? Did quality improve? Did customers get a better answer? Did employees spend less time on low-value coordination? Did the team make a better decision faster? Did the organization learn something it can now repeat?
Over time, aspects of the work architect role will apply to everyone. The future will not belong only to people who can use AI tools. It will belong to people who can redesign their work around them. That requires curiosity, judgment, and a willingness to challenge inherited assumptions. It also requires leaders who do not punish people for stepping outside the boundaries of a job description when the real opportunity sits between functions.
The concept of leadership is very much in play right now. I will not insult your intelligence by offering a generic list of abstract qualities leaders supposedly need in the age of AI. We have all read those lists. They are not wrong, but they are rarely useful.
What I would be looking for instead are people who are genuinely obsessed with solving problems from a particular point of view. It might be customer obsession. It might be an engineering-led desire for operational resilience. It might be a hatred of unnecessary bureaucracy. It might be the simple, underrated impulse to get rid of annoying tasks and broken ways of working. These are the people who will find the real uses for AI, not because they are trying to look like AI-native leaders, but because they have something they are trying to make better.
Those people should be rewarded. They should be given room to experiment, permission to cross boundaries, and naturally - as many AI tokens as they need. But most of all, they should be measured by the problems they solve, not the proxies they perform.
The real problem with tokenmaxxing is not that people will waste a few extra tokens. It is that organizations will once again confuse what is visible with what is valuable. Leadership in this strange new world will not be defined by who vibe codes the most reports, looks the busiest, or invokes the most agents. It will be defined by who can redesign work so that intelligence, human and machine, compounds.
And when you find those people, the best thing you can do is often the simplest: get out of their way.