There is a new word making the rounds in Silicon Valley, and it has already migrated from engineering Slack channels into boardroom performance reviews at some of the biggest companies on the planet. That word is tokenmaxxing. If you have not encountered it yet, you will soon, because the debate it has sparked cuts right to the heart of how organizations measure human productivity in an AI-first world.
First, what exactly is a token?
Before we get into the workplace drama, a quick foundation is worth building. When you type a message to an AI model like Claude, ChatGPT, or Gemini, the system does not process your words the way you wrote them. It breaks your text into smaller units called tokens, roughly three to four characters each, or about three-quarters of a word on average. A sentence like “the quick brown fox” becomes something like four or five tokens.
Tokens matter because they are the currency of AI. Every time a model reads a prompt and writes a response, it consumes tokens, and enterprises pay for them. When a developer asks an AI coding assistant to review a 500-line codebase, draft a solution, test it, and refine the output three times, they may burn tens of thousands of tokens in a single session. An AI agent running autonomously can consume millions of tokens without a human typing a single keystroke.
Token volume, therefore, became the most visible proxy for how intensively a team or individual is using AI. And that is exactly where the trouble starts.
Where does tokenmaxxing come from?
The “-maxxing” suffix has an unlikely origin. It started in online fitness and self-improvement communities. Looksmaxxing meant optimizing your physical appearance, gymmaxxing meant maximizing training volume. The logic was always the same: take a variable, make it as large as possible, treat the number as a proxy for progress.
Tokenmaxxing, then, is the act of maximizing AI token consumption, either genuinely, by delegating as much work as possible to AI agents, or performatively, by engineering high usage numbers to satisfy managers tracking those metrics. Both versions exist in real workplaces right now, and the fact that they are hard to tell apart from the outside is exactly what makes this trend so interesting, and so risky.
Why are companies suddenly obsessed with it?
The pressure came from the top, and it accelerated fast. Shopify’s CEO sent a now-famous internal memo declaring that AI usage was a baseline expectation, and that employees would need to demonstrate why a task could not be AI-assisted before requesting more headcount. Microsoft told managers that AI usage was no longer optional and began factoring it into performance conversations. Meta tied AI-driven impact to annual reviews, and an employee-built dashboard began ranking staff by weekly token consumption, awarding titles to the heaviest users.
The logic behind the mandate is not entirely unreasonable. Companies have invested enormous sums in AI infrastructure, and executives want evidence of adoption. Token usage is measurable, immediate, and easy to compare across teams. The problem is what happens when a signal becomes a target.
Why product managers specifically should care
Product management has always lived at the intersection of measurement and judgment. A good product manager knows how to read a metric, and knows, crucially, when a metric is lying. Tokenmaxxing is a masterclass in metric failure, and it is arriving in product teams right now.
Here is the core tension. AI is genuinely transforming how product work gets done. Discovery interviews can be synthesized faster. Research transcripts can be clustered in minutes. Specs can be drafted, challenged, and revised in a fraction of the time. But when token volume becomes a performance signal, the incentive shifts. A product manager who thinks hard and writes a sharp prompt that solves a problem in 200 tokens looks less productive than one who runs circular conversations, generates five redundant drafts, and logs 20,000 tokens. The second person used AI more. The first got more done.
When a measure becomes a target, it ceases to be a good measure.
This is Goodhart’s Law in action. The software industry learned it with lines of code in the 1990s: developers evaluated on code volume wrote longer, more redundant code. Tokenmaxxing is lines-of-code thinking for the agentic era.
The evidence that volume does not equal value
The data is catching up with the intuition. A randomized controlled trial published by METR found that experienced developers using leading AI coding tools took 19 percent longer to complete tasks while reporting they felt 20 percent faster, a wide gap between perceived and actual productivity. The confidence AI generates can be its own liability.
A 2025 analysis of over 200 million lines of code found that since AI assistants became widespread, code that is written and then reverted or rewritten within two weeks has roughly doubled, while careful refactoring fell by more than half. The code is generated faster. It is also discarded faster. And a Stanford and BetterUp study coined a term for output produced quickly but poorly: workslop. Each instance costs colleagues about two hours to identify and fix, and the people who send it are rated as less creative, less trustworthy, and less intelligent. High volume, low quality, reputational cost. That is the trap in practice.
The correction is already underway
Badly chosen metrics get corrected, sometimes through competitive pressure, sometimes through embarrassment. Duolingo’s CEO, who had made AI usage part of reviews, publicly reversed course, acknowledging that AI is not yet reliably better than skilled humans at core product work. Klarna, which had aggressively replaced customer service staff with AI, quietly began rehiring after quality declined. The companies getting ahead of this are not abandoning AI. They are abandoning the proxy metric. The question shifts from how much AI did you use to what did it help you deliver.
What this means for how product teams should think
Token volume is a reasonable signal during an adoption phase, when you are trying to understand whether teams are using the tools at all. After that, it degrades quickly into a measure of activity rather than output, and once people know it is tracked, it becomes a measure of willingness to game a dashboard.
The more durable approach is to evaluate AI use the way you would evaluate any other process: against the outcomes it produces. Did the AI-assisted research lead to sharper prioritization? Did the AI-drafted spec reduce back-and-forth with engineering, or generate more of it? The teams that compound value from AI are the ones that treat token volume the way they treat page views: useful for a quick pulse check, dangerous as a North Star. The real question is never how many tokens were burned. It is always what those tokens helped build.
What are you seeing in your organization? Are usage metrics driving adoption, or are they already driving the wrong behaviours? The conversation is worth having before the leaderboard arrives.