Somewhere inside Amazon’s corporate sprawl, an employee is asking a large language model to summarize a document they already understand, for a meeting they have already prepared for, generating tokens that will be logged, counted, and reported upward as evidence of transformation. The task is real. The need is not.
This is the texture of AI adoption theater at industrial scale — and the story now unfolding at Amazon is less about one company’s internal dysfunction than about a foundational assumption embedded in how every major enterprise is trying to win the AI transition. That assumption is this: that usage is a proxy for value. It is the assumption most likely to be wrong, and the damage from its failure will take longer to surface than anyone is currently pricing in.
The Metric That Became the Mission
Amazon has been explicit about its AI ambitions. CEO Andy Jassy has described artificial intelligence as a once-in-a-lifetime technological shift, and the company has invested heavily in both its own models and its Bedrock platform for enterprise customers. Internally, that ambition has translated into pressure on managers and individual contributors to demonstrate AI engagement — tracked, in part, through usage metrics. Workers, according to reporting by Fast Company, have responded by fabricating tasks: feeding AI tools documents that need no summarizing, asking questions they already know the answers to, running queries that serve no purpose beyond registering a session. The practice has acquired a name in some corners of the internet — “tokenmaxxing” — which manages to be both funny and precise.
Amazon’s official position is that AI usage metrics do not feed into performance evaluations. Workers clearly do not believe this, or do not trust that the policy will hold. That gap between stated policy and felt reality is itself diagnostic. When employees calculate that the safer bet is to game a metric the company says doesn’t matter, the metric has already metastasized into something the company can no longer fully control.
Why the Assumption Feels Reasonable Until It Isn’t
The logic behind tracking usage is not stupid. Early in any technology adoption cycle, inertia is the primary obstacle. People default to familiar tools; new systems gather dust. Measuring uptake creates accountability and surfaces which teams are genuinely experimenting versus which are waiting for the pressure to pass. There is a legitimate school of thought — backed by McKinsey’s generative AI research — that organizations which move early and broadly tend to build compounding advantages: institutional familiarity, workflow integration, proprietary fine-tuning data. If you want broad adoption, you reward broad adoption.
But this logic contains a hidden load-bearing wall: it assumes the activity being measured resembles the activity you want to encourage. Once employees understand what is being counted, they will produce more of what is being counted. At Amazon, what is being counted is sessions and tokens. What the company presumably wants is judgment — the kind that routes the right problem to an AI tool, extracts something genuinely useful, and integrates that output into a decision that would otherwise have been slower or worse. Judgment is not measurable in the same register as token generation. So the organization measures token generation and hopes judgment follows. It usually doesn’t.
This is not a new failure mode. It is Goodhart’s Law — when a measure becomes a target, it ceases to be a good measure — wearing a new costume. What is new is the speed at which the costume is being adopted across the entire corporate world simultaneously, and the specific way AI’s architecture makes the gaming so frictionless. Asking a model to summarize something costs almost nothing and leaves a clean trail of logged engagement. The gap between performed adoption and real adoption has never been cheaper to manufacture.
Thirty Percent Is the Number That Should Worry Amazon’s Competitors
Roughly 30 percent of workers in surveys across multiple industries report using AI tools primarily because they feel socially or institutionally obligated to, not because the tools improve their work. That number is probably an undercount — social desirability bias runs in the opposite direction when employers are watching. What it suggests is that a significant share of the usage data now flowing up through corporate dashboards is measuring compliance with an expectation, not engagement with a capability. Every board presentation citing strong internal AI adoption rates should be read with that in mind. The data looks like transformation. It may be documenting something closer to AI adoption theater at scale, a performance staged for dashboards rather than for customers.
The reason Amazon’s competitors should care is not schadenfreude. It is that they are almost certainly doing the same thing. The incentive structure that produces tokenmaxxing at Amazon is identical to the incentive structure inside every large organization that has announced an AI mandate without a clear theory of how AI creates value in their specific workflows. You can fill a dashboard with green checkmarks and build almost nothing.
What the Research Community Is Watching
For researchers and developers building tools on top of enterprise AI platforms, the tokenmaxxing problem is not merely sociological — it is a data quality crisis with compounding effects. Research on human-AI collaboration has repeatedly found that the quality of user feedback loops is central to how well AI systems can be refined for domain-specific use. When usage is manufactured, the signal that should be shaping model improvement and workflow integration becomes noise. Enterprises that produce enormous volumes of low-intent interactions are not building better AI capability — they may be actively degrading the feedback architecture that would let them do so.
Educators building curricula around enterprise AI platforms face a related but distinct problem. The implicit lesson currently being taught inside organizations like Amazon — that the metric is the goal — will shape how an entire generation of knowledge workers understands AI’s role in professional life. That is a difficult mental model to unlearn. The students who graduate into environments already saturated with AI adoption theater will need to be taught something harder than tool proficiency: how to distinguish between AI use that improves an outcome and AI use that documents an activity. Few curricula are currently built around that distinction.
The Fragility Hidden in Plain Sight
Here is the assumption that will not hold: that the gap between performed adoption and genuine capability-building is temporary — a transition cost that will resolve itself as familiarity grows and resistance fades. This is the story Amazon’s leadership is most likely telling itself. Early friction is normal. People resisted spreadsheets. They resisted email. Give it time and the behavior will become genuine.
The problem is that spreadsheets and email did not require users to construct a fake relationship with the tool in order to satisfy a compliance requirement. Tokenmaxxing is not just passivity or reluctance — it is active mis-training of the user’s own habits. An employee who spends six months asking AI questions they already know the answers to is not building toward fluency; they are rehearsing a posture of engagement that is hollow at its core. When genuine workflow integration is eventually required, that employee may be further from useful adoption than if they had never been pushed to perform it at all. The performance does not become real through repetition. It calcifies.
Reuters reporting on enterprise AI investment has documented widespread frustration among executives who have committed significant capital to AI infrastructure without seeing commensurate productivity gains. That frustration is the early signal of a reckoning. It does not yet have a name in most boardrooms. In time, it will — and the name will be some variation of what has already been coined in the margins: AI adoption theater.
Still, one genuine complication sits inside this critique. Some workers who begin tokenmaxxing do, by accident or curiosity, discover a workflow the tool actually improves. The performance occasionally becomes real. Whether that effect is large enough to redeem the strategy, or whether it simply generates a few converts while entrenching cynicism across a much larger population, is genuinely unknown. The optimistic case is not impossible. It is just not the most likely outcome, and building a company’s AI transformation strategy on accidental discovery is not a strategy.
“The danger isn’t that employees won’t use the tools — it’s that they’ll learn to use them performatively, and that will feel like success for long enough to cause real damage.” — Senior AI program lead at a Fortune 100 enterprise
The board-level question that is not yet being asked with sufficient seriousness is this: if your AI adoption metrics are strong and your productivity metrics are flat, which number do you believe? Most organizations are still answering: the adoption metrics. They have invested in those. They have announced those. The productivity flatline can be explained away by transition costs, learning curves, the need for more time. The AI adoption theater can continue quite comfortably as long as the second question stays soft.
Inside Amazon’s Seattle offices right now, someone is deciding whether to route a genuine problem to an AI tool that might actually help, or to close the laptop and do the thing themselves faster. That decision, replicated across millions of workers at thousands of organizations, is where the AI transition is actually happening — not in the dashboards, not in the token counts, not in the shareholder letters. The dashboards are moving in one direction. That decision, still unmeasured, is moving in another.
FetchLogic Take
Within 18 months, at least two S&P 500 companies will publicly revise their AI adoption frameworks away from usage-based metrics after internal audits reveal that a material share of logged AI activity generated no traceable business output. The revision will be framed as a maturation of measurement strategy. It will actually be the first corporate acknowledgment that AI adoption theater has a cost that eventually shows up in the numbers — and that Goodhart’s Law does not grant exceptions for technologies with good press releases.
Related Analysis
Amazon’s AI Mandate Is Breeding a Culture of Productive-Looking NoiseMay 15, 2026
Amazon’s Token Trap: What ‘Tokenmaxxing’ Reveals About the Limits of Mandated AI AdoptionMay 13, 2026
TanStack’s npm Account Hijack Reveals the 60-Second Window Security Teams Are MissingMay 12, 2026
Meta’s $50B AI Bet Is Burning Out the People Building ItMay 10, 2026