Tokenmaxxing: The New Face of AI Theater

Scene one. Executive committee, last Thursday of the month. The transformation lead pulls up a dashboard: the operations team burned 2.3 million tokens this month, up 40% from last quarter. Smiles around the table. Someone says "we're on the right track." Nobody asks what got done with those tokens. Nobody asks whether any process got shorter, whether any customer noticed anything, whether any decision got better. The token-consumption metric went up. That was enough.

Scene two. A data analyst gets the directive: "use it in your day-to-day." They start running emails through the tool before sending them, asking it to summarize documents they're going to read in full anyway, asking it things they already know. Their workflow didn't change. Their judgment didn't change. Their deliverables didn't change. They just added an intermediate step that lets them report usage. The tool entered their routine without touching their judgment.

Scene three. A development team adopts agentic coding. Code output triples. Features hit review at a pace QA can't sustain. One of the developers starts generating more bugs than before — not for lack of talent, but because the accelerated volume exposes every judgment call that used to pass by unexamined. The monthly report celebrates ship rate. The product stays the same, or worse… it silently degrades…

What do these three stories have in common? A lot more than it seems.

The Symptom Has a Name; the Antipatterns Don't

A few days ago, Jeff Gothelf published a piece that's rightly circulating in transformation circles. He titled it The rise of performative AI work. In it he names a phenomenon:

Amazon employees are reportedly inflating their token usage for no real reason, simply because the company started measuring AI usage as a performance indicator.

Tokenmaxxing is the practice of artificially inflating AI token consumption — creating unnecessary agents, generating superfluous queries, or forcing workflows where the tool adds nothing — to hit internal adoption metrics. The metric became the goal. The goal became theater.

Gothelf describes the symptom precisely and lists three signals any leader can use to diagnose their organization:

1. Measuring adoption instead of impact,
2. Pushing usage without asking why, and
3. Having no baseline for what "better" even means.

So far, the analysis is sharp. What it immediately gets us thinking is: what organizational practices produce those signals?

Because tokenmaxxing isn't an Amazon anomaly. It's the predictable consequence of three antipatterns that have been installed in how large organizations try to transform themselves for decades, now applied to artificial intelligence.

The urgency of the moment is real. McKinsey's latest State of AI, published in late 2025, reports that 88% of organizations use AI regularly in at least one business function. And yet only 6% manage to capture significant EBIT impact. That gap — between universal adoption and material impact — doesn't close with more usage. It closes by understanding what practices are producing it.

Antipattern One: AI Theater

Fifteen years ago we coined the term Agile Theater in the agile world to describe organizations installing visible ceremonies — Daily, Sprint, Retro — without touching the values or principles that sustain them. It was theater: it gave the illusion of transformation while shielding resistance to change.

What scene one shows is exactly the same movie with different technology. AI Theater is the adoption dashboard presented as evidence of transformation. It's the usage metric mistaken for an impact metric. It's the ritual that replaces the uncomfortable question: what changed for our customers since we started this?

The signal Gothelf names — measuring adoption instead of impact — is the visible symptom. The antipattern underneath is structural: we install the rituals and expect transformation to emerge on its own. It never emerged with Agile. It's not going to emerge with AI.

Antipattern Two: Confusing Adoption with Appropriation

Adopting a tool means folding it into your workflow. Appropriating a tool means your judgment changes because you use it. These are two different things.

Scene two describes adoption without appropriation. The analyst uses the tool, but the way they think about the work hasn't shifted. They report usage, not transformation. And that's exactly what happens when the organization pushes the wrong verb: it pushes people to use without accompanying the shift in judgment about when it's worth using, what problem it solves, what decision gets better by using it.

The signal Gothelf names — pressure to use AI stronger than the why — produces this antipattern with clinical precision. When "are you using AI?" becomes a performance question, the question that actually matters stops being asked: should I be using AI for this? And often the best decision is not to use it. That distinction disappears the day usage becomes the goal.

Antipattern Three: Moving the Bottleneck Without Moving With It

When execution speed triples with AI, the bottleneck moves. It's no longer about how much code you produce, how many emails you answer, how many reports you generate. It's in judgment: are we building the right thing? does this solve something real? what decision is implicit in this output I just generated?

Scene three describes a team that saw its execution accelerate but didn't move its judgment capacity along with it. More volume means more judgment calls per unit of time, and every decision that used to pass unexamined now becomes visible — sometimes as a bug, sometimes as a product that solves nothing.

The signal Gothelf names — no baseline for what's better — isn't an oversight. It's the consequence of not having migrated cognitively when the bottleneck migrated. If your team is still measuring what it measured before AI, it's measuring the part of the work that was already running fine. The part that now matters got orphaned.

The Question Before the Answer

There's a predictable temptation at the end of a piece like this: offer the recipe. Three steps to avoid AI Theater, five principles for real appropriation, a framework for moving judgment along with the bottleneck.

We're not going to offer one. For an epistemological reason: the solution isn't generic. It depends on the iceberg each organization has beneath the water, on the prior history of its transformation, on the real state of its teams. And for a practical reason: if you already know what to do after reading a blog post, you don't need anyone.

The question we do want to leave you with is this. Reread the three opening scenes. Did you recognize one? One and a half? All three? That's the first useful piece of information. Here's the second: does your leadership team know they're living these, or are they reporting them as adoption success?

Itera Is Already Working on This

At Itera we're working with teams that are starting to notice the difference between using AI and transforming with AI. We don't bring a manual or a packaged methodology. We bring the perspective that comes from having seen the iceberg upside down many times before, now applied to this new wave.

If any of the three scenes felt uncomfortably familiar, let's talk. The conversation itself is usually where something starts to move.