In April, a Meta employee built a leaderboard. It ranked 85,000 colleagues by how many tokens they fed into Anthropic's Claude. The top user burned 281 billion tokens in thirty days. Cost to Meta: roughly $1.4 million for one person's prompt habit.
Around the same time, Jensen Huang told a room of investors that every Nvidia engineer should have a token budget. His framing was unambiguous: a half-million-dollar engineer who doesn't consume at least $250,000 in tokens is, in his words, deeply alarming.
This is the new performance review.
What you are actually being paid to do
Stanford and BetterUp Labs published a study late in 2025 that named the output of this regime: workslop. AI-generated documents that look polished, get circulated through email and Slack, and quietly offload all the actual thinking onto whoever opens them. Forty-one percent of knowledge workers had been hit with workslop in the previous month. Average cost per incident: nearly two hours of rework. Extrapolated to a 10,000-person company: roughly $9 million a year in lost productivity.
Then UC Berkeley ran an eight-month longitudinal study inside a tech firm. The finding was blunter. AI tools did not reduce anyone's workload. They intensified it. Time spent on individual tasks went up between 27% and 346% depending on the role. Sixty-six percent of workers reported spending six or more hours a week correcting AI output.
So the picture is this. You are measured on how many tokens you consume. The tokens produce slop. The slop lands on someone else's desk. They spend two hours fixing it. Their token consumption goes up, because they are now using AI to fix the AI. And so on.
The workslop is not a failure of the system. It is what the system is paying for.
The metric is the message
The standard reading of this story is that companies are using bad metrics. Salesforce and HubSpot have already lined up to sell the alternative: "outcome maxxing" instead of token maxxing. Measure what gets shipped, not what gets consumed.
That reading is too kind.
Token consumption is not a flawed proxy for productivity that smarter leaders will eventually replace. It is a perfectly accurate proxy for something else. It measures how committed a workforce is to the corporate AI strategy. It measures how loudly a company can tell its investors that AI adoption is real. It measures how plausible the layoff narrative becomes next quarter, when somebody has to explain why headcount is dropping while compute spend is climbing.
In Q1 2026, the tech industry cut almost 80,000 jobs. Around half were attributed to AI efficiencies. The share of layoffs justified by "AI investment" jumped from 5% in 2024 to 20% this year. Meanwhile economists looking at aggregate productivity data find no measurable relationship between AI spending and output. The story being told in earnings calls and the story showing up in the data are not the same story.
Why this matters for your business
If you are a CEO using token consumption as a proxy for AI adoption, you are not measuring productivity. You are measuring obedience.
Your employees are not stupid. They optimize for the metric. The metric rewards volume. Volume produces slop. Slop produces rework. Rework gets attributed to "AI inefficiency" and used to justify more aggressive automation, which drives more obedience to the metric.
The cycle is closed. It runs by itself.
The workslop is not what is going wrong with your AI strategy. It is what your AI strategy is producing on purpose, even if nobody on your leadership team would say it out loud.
The honest version of the question is not "how do we reduce workslop." It is: do we have a real use case for AI inside the company, or are we using AI to perform AI in front of the market? Most companies that asked this quietly in 2024 and 2025 are getting the answer back now, in 2026. A pile of documents nobody reads. A flat productivity curve. A board meeting nobody wants to chair.