Amazon pulled its AI leaderboard after tokenmaxxing turned work into gaming

Amazon yanked its AI leaderboard after some employees made “useless AI work” to game the rankings. It is a small headline with a big implication: the fastest way to measure AI adoption might be rewarding the wrong behavior.

Tokenmaxxing became the buzzword of the season this spring, and by summer the trend was already “running on empty.” In the same orbit as Amazon’s leaderboard reversal, Palantir CEO Alex Karp compared tokenmaxxing to a porn addiction, Duolingo walked back a decision to weigh AI use in employee performance reviews, and Meta and AT&T reportedly started curbing AI use as costs skyrocket. Across these examples, the pattern is consistent. When incentives reward AI activity instead of business impact, workers optimize for the scoreboard, not the outcome.

That is where “AI sprawl” comes in. The pressure is to use AI for the sake of using AI, so workers adopt new agents or vibecode solutions with a growing menagerie of tools. The problem is not only tool sprawl. It is that companies struggle to wrap their arms around what employees are doing, which drives expensive duplication and shallow knowledge transfer. Workers may waste time “botsitting,” meaning giving AI the necessary context and edits to make output usable. Or they may shuffle the same prompts between tools until something clicks, burning through tokens and budgets while failing to pass the best tips and tricks to coworkers.

A Glean Work AI Institute survey of 6,000 digital workers in the US, UK, and Australia puts numbers on the mess. Seventy-seven percent of those who use AI engage with multiple programs weekly. A third use four or more tools. And 60% will shuffle the same prompts between multiple tools when the first output is not sufficient. On paper, workers also report meaningful personal time savings. Individually, they save an average of 11 hours each week. But only 13% of surveyed workers said these savings have “significantly improved” the company’s performance. That gap is the whole story in a single statistic: personal productivity does not automatically become organizational performance.

The question business leaders keep circling is the “big why.” Kate Niederhoffer, head of BetterUp Labs at the coaching and workforce development company, said workers need clearer answers to why adoption is happening and what it is trying to accomplish. She also highlighted communication as the lever: the message has to be clear and compelling so people use the tools to achieve shared goals. But few companies are answering those questions, and without that framing, adoption turns into individualism. The rhetoric around AI has reinforced it, pushing the idea that you must maximize and master AI or be replaced by someone who can. That urgency can be productive, but it can also narrow focus to personal output and undermine collaboration.

Collaboration is not just a cultural ideal, it is an efficiency machine. The source points out that AI sprawl often produces duplicate work when employees operate in silos. Lee Senderov, chief transformation officer at Travelport, described observing exactly how costly this can be. In her example, one worker burned through 160 times the amount of tokens that the next most prolific AI user did over a four-day period. The costs are not only hard dollars, token spend. There are also soft costs: “wasting effort,” then figuring out “who’s the expert that should be writing this.” When people work alone with AI, they can also dilute outcomes and flatten the rewards of collaboration, trading shared expertise for a quick enough solution.

This is not a brand-new phenomenon. The source ties the behavior to Herbert Simon’s concept of “satisficing,” where individuals settle for the good enough option instead of evaluating every possible alternative. Organizations exist to coordinate satisficing individuals at scale so the overall system stays productive. But layoffs of thousands of workers and pivots toward AI collide with that theory. And while Meta plans to boost its spending on AI between 60% and 87% this year, after its “year of intensity” and slashing jobs, the broader claim in the source is that single individuals doing what used to take whole teams can erode the fabric that institutions rely on.

If that sounds like coordination risk, the source explicitly connects it to the “tragedy of the commons.” Rebecca Hinds, head of Glean’s Work AI Institute, frames it this way: if individuals benefit from a shared resource, they may push it toward depletion to boost their own stature and credibility, even if the team or project gets downgraded. The implication for executives is direct. You can encourage AI usage and still degrade trust and outcomes if people use AI in ways that the team cannot absorb. Prior BetterUp research cited here found that when people produced “workslop,” AI-generated documents and presentations without proper oversight, coworkers began to trust them less. Overreliance on AI can also shift work toward chatbots for answers and gen AI for outputs that previously required human expertise.

The twist is that AI is not purely a problem. The source also emphasizes upside. AI democratizes innovation, enabling marketers to “vibe code,” letting agents act like personal assistants, and allowing startups to do more with fewer hires. The strategic challenge is turning individual gains into team and company workflows. Senderov’s response at Travelport is to experiment with centralizing AI workflows, encouraging collaboration when two people are working on the same thing, and showing best use cases at the enterprise level. Centralization gets harder as companies grow, but the takeaway is clear from these examples: there is no tokenmaxxing shortcut to building an AI operating system that actually improves performance.