Here's something nobody at your company's all-hands meeting is going to say out loud:
A chunk of your AI spend this year is being consumed by people who are using the tools to look productive rather than be productive. They're not cheating the system. They're not bad employees. They're just doing exactly what humans do when you attach a visible score to a behaviour — they optimise for the score.
The name for this has arrived. It's called tokenmaxxing. And if your company rolled out any AI tools in the last 18 months without a usage strategy to go with them, it's almost certainly already happening.
What Is Tokenmaxxing?
Tokenmaxxing is what happens when employees use AI tools not to get work done faster, but to look like they're using AI more than everyone else.
The term picked up in 2026 after companies including Amazon, Uber, and Meta introduced internal dashboards tracking how many AI tokens each employee consumed. The thinking made sense at the time: if we're spending millions on enterprise AI licences, we want visibility into whether people are actually using them. Token consumption felt like a clean signal. High usage equals high adoption equals ROI.
What followed was more human than any algorithm could have predicted.
Employees started optimising for the number. They uploaded enormous documents when a paragraph would have done. They re-ran the same query five times with slightly different wording. They routed basic tasks through the premium model when the cheaper one would have handled it fine. Some ran background requests with no particular goal, just to keep their count moving.
Token counts climbed. Actual output quality stayed flat — or dropped, because the people generating the most tokens were spending more time fiddling with prompts than doing their jobs.
Several enterprises burned through their entire annual AI budget before February was over.
First, Let's Make Sure We're Speaking the Same Language
If you're not deep in AI infrastructure, the word "token" might feel abstract. Here's all you need to know:
A token is roughly three-quarters of a word. Every time someone in your company sends a message to an AI model and gets a response back, both the input and the output get measured in tokens and billed accordingly. A quick back-and-forth costs a few hundred tokens. A complex task — long document, detailed reasoning, lengthy output — can run into hundreds of thousands.
When one person is doing this, the cost is negligible. When five hundred people are doing it dozens of times a day, you're looking at tens of millions of tokens consumed daily. Add tokenmaxxing behaviour on top of that baseline, and you can see exactly how a year's budget disappears in four months.
The pricing context makes this more concrete: frontier AI models currently run between $5 and $30 per million output tokens depending on the provider and model tier. That's the list price. The actual bill is driven by volume — and volume is driven by behaviour.
The Leaderboard That Backfired
The dashboards weren't a bad idea in theory. If you've invested significantly in AI tooling, you want to know it's being used. Token consumption felt measurable, comparable, and objective.
The problem is that tokens measure activity, not outcomes. This is the same trap companies have sprung on themselves before. Measuring software teams by lines of code written. Measuring customer service by number of tickets closed. Measuring marketing by posts published. You always get more of what you measure. What you measure is rarely the thing you actually want.
When token usage became visible and rankable, the same psychology that makes people pad meeting minutes and send unnecessary reply-all emails kicked in immediately. Nobody decided to game the system. They just responded to the incentives in front of them, the way people always do.
Amazon's internal analysis found something striking: the employees with the highest token consumption were not the same employees producing the strongest work. There was essentially no correlation between the leaderboard position and actual performance impact. The leaderboard was measuring busyness. It had nothing useful to say about leverage.
There's a Bigger Thing Happening Here
Tokenmaxxing is a symptom. The actual disease is something worth naming directly: AI productivity theatre.
This is what unfolds when the pressure to be seen as an AI-forward organisation moves faster than the actual integration of AI into meaningful work. Leadership announces the initiative. Teams get the logins. Everyone performs enthusiasm in the right meetings. Usage numbers go up. Real productivity improvements are harder to point to, so people point to adoption metrics instead and hope nobody looks too closely.
It's not new. This exact cycle happened with social media in the early 2010s. Companies rushed onto every platform, measured success by follower counts and posting frequency, and quietly avoided asking whether any of it was driving business outcomes. A few years later, most of them abandoned half those channels and concentrated on the two or three that actually worked.
The AI equivalent of posting on Google+ just to say you're posting on Google+ is running expensive prompts through a premium model at 11pm just to keep your token count respectable.
The companies that are getting genuine, measurable productivity improvements from AI in 2026 — and there are many of them — are not doing it by maximising adoption metrics. They're doing it by identifying specific workflows where AI genuinely reduces effort or improves quality, deploying carefully in those areas, measuring the actual outcome, and expanding from there. That approach doesn't make for exciting all-hands slides. It works.
What the Smart Response Actually Looks Like
The answer to tokenmaxxing is not to cut AI access or punish people for using the tools. That's the wrong direction entirely.
The answer is to build the systems and culture that make sensible usage the natural default. Here's what that looks like in practice:
Tiered model access based on task type. Not every task needs your most powerful model. A well-designed internal deployment routes simple requests — reformatting, summarising, basic Q&A — to cheaper, faster models automatically, and reserves the premium tier for work where quality has a real business impact. This single change cuts token spend by 40 to 60 percent in most organizations without anyone noticing any difference in output quality.
Outcome measurement instead of activity measurement. Replace the token leaderboard with something that tracks what actually matters. Hours saved on specific tasks. Time to completion on repeatable workflows. Error rates before and after AI assistance. These numbers are harder to collect than token counts. They're also the only numbers that tell you whether the investment is working.
Actual usage education, not just access. Most AI rollouts consist of: here's your login, good luck. The companies with the strongest results treated it more like a skill deployment. What is this tool genuinely good for? What does it handle poorly? How do you write a prompt that gets a useful result without burning unnecessary tokens? Which tasks belong on which model tier? That kind of structured guidance prevents the behaviour patterns that tokenmaxxing thrives on.
Team-level budget visibility. Giving teams transparency into their own spend, relative to their allocation, creates natural accountability without needing surveillance or enforcement. When a team can see their usage in context, sensible self-regulation tends to follow. The problem with company-wide aggregation is that no individual feels responsible for the total number — it's always someone else driving it up.
If You're Building AI Products, This Affects You Too
There's a downstream consequence of all this that's worth paying attention to if you're a developer or founder building on top of AI APIs.
The enterprise backlash against uncontrolled AI spend is making buyers more sceptical of AI-powered software that can't clearly demonstrate what it does for the business. Companies that just discovered their internal teams blew the annual AI budget on leaderboard optimisation are not in a generous mood when an AI-powered SaaS vendor can't show a direct line between usage and outcome.
This is actually good news if you're building something that solves a specific, measurable problem. If your product saves a user two hours every week and you can show it, the budget conversation in 2026 is easier than it was in 2024, because buyers are actively looking for AI that justifies its existence. If your product is AI-flavoured without a clear outcome, you're going to run into walls that wouldn't have been there eighteen months ago.
The AI price war is pushing API costs down significantly — great for your margins. But buyer scrutiny is going up at the same rate. Both trends point in the same direction: specificity wins. Build for a measurable outcome, show the measurement, and you're in a much stronger position than the broad-purpose AI tools that ask users to figure out the value themselves.
The Uncomfortable Conclusion
Tokenmaxxing exists because companies moved faster than they built frameworks for. They bought access before they had strategy. They measured activity before they understood what outcome they were actually after. They created incentives that rewarded the performance of AI adoption over the substance of it.
None of this means enterprise AI is a bad investment or a failing experiment. The organizations extracting real value from these tools are everywhere in 2026, and they're accelerating. But the gains are coming from deliberate, focused deployment in specific workflows — not from giving everyone access and hoping emergent productivity appears.
The employee with the highest token count on your leaderboard is probably not your most AI-enabled person. They're just the one who figured out what the leaderboard was actually measuring.
That's worth thinking about before you design what comes next.
This Is Exactly the Kind of Problem We Help Organizations Navigate
At NEV Infotech, we work with businesses at every stage of AI adoption — including the ones who moved fast, got the tools in place, and are now trying to figure out why the results don't match the investment.
Getting AI right inside an organization is not a technology problem. The technology works. It's a strategy problem — knowing where to deploy it, how to measure it, how to build the internal culture around it, and how to avoid the traps that are swallowing budgets at companies much larger than yours right now.
If your organization is somewhere in that space — good intentions, real tools, unclear returns — that's a conversation we're built for. We help businesses move from AI theatre to AI that actually changes how work gets done.
Because the goal was never a higher token count. The goal was a better business.
NEV Infotech helps organizations across industries build AI strategies that deliver real outcomes — not just adoption metrics. If you're rethinking how your teams use AI, or building toward it for the first time, reach out and let's talk about what that looks like for your specific context.