Inside tokenmaxxing: The gamification of corporate AI that's costing millions
From $500M invoices to checking the weather on Claude, AI enterprise deployment is becoming an expensive exercise in looking busy.
TL;DR: Lacking clear methods to track generative AI return on investment, some companies default to simply tracking AI usage.
What’s tokenmaxxing? The practice of artificially inflating generative AI usage at work — firing off unnecessary prompts, generating outputs no one reads, or routing trivial tasks through AI — to appear more productive or “AI-engaged” in workplace metrics (see Goodhart’s Law). The ecological cost: Operational busywork is accelerating energy use, with global consumption projected to hit 945 TWh by 2030. The real winners: High-performing organizations like JPMorgan, Walmart, and Salesforce structure AI deployments around business outcomes rather than raw activity.
Recently, inside a large corporation, a budget line slipped and no one noticed. When the bill came, they discovered they’d spent $500M dollars on AI… in a month. The company, quoted anonymously an Axios report last week, had rolled out Claude AI to their workforce without putting a limit on how much anyone could use.
We’re used to traditional software, which charges a license or per-seat fee, but most generative AI charges by the token (the unit of data processed by the model). Every prompt you fire off, every sentence it gives back, every step an AI agent takes on your behalf - it’s all added to your tab. Researchers found that something as small as how you phrase a prompt can up your costs significantly. And when AI agents get involved, it gets expensive. Industry reporting pegs agentic AI tools at roughly 1,000 times the token consumption of a basic query.
“One CTO told Axios their employees were using AI to check the weather.”
OpenAI’s 2025 enterprise report found corporate ChatGPT message volume jumped roughly eightfold year-over-year, and reasoning token consumption per organization shot up more than 320 times. Reuters Breakingviews called it “sticker shock.”
Now the biggest names in tech are backing away. Microsoft cancelled most of its Claude Code licenses, partly over cost, according to The Verge. Uber COO Andrew Macdonald said AI bills are getting “harder to justify.”
“Corporate leaders are starting to question whether soaring AI spending is delivering meaningful returns,” Axios reported.
The proxy problem: Why companies track AI usage instead of ROI
Earlier this year, Jensen Huang famously announced that if he didn’t see an engineer who was being paid a $500,000 salary using $250,000 worth of tokens, he’d be “deeply alarmed.”
Reporting in 2026 found that Meta, JPMorgan, KPMG and Amazon were all monitoring employee AI activity, using the data to sort the adopters from the stragglers. At Meta, “AI-driven impact” graduated from a nice-to-have in your self-review to a formal piece of your performance evaluation. At Accenture, employees were reportedly expected to prove they were using AI if they wanted to move up.
Without a clean way to measure whether AI is paying off, companies defaulted to tracking the thing they could see: usage. Some companies started flagging their heaviest users as in-house AI champions. Others wove “AI engagement” into the language of performance reviews and promotions.
So employees started to inflate their AI use, which is called tokenmaxxing.
Optimizing the metric: Goodhart's Law in the AI workplace
Tokenmaxxing is exactly what it sounds like: gaming your AI numbers. Firing off prompts you don’t need. Generating outputs nobody will read. Running tasks through AI unnecessarily so you look more “AI-engaged.”
Workers pad their AI activity because their visible usage reads as innovation, or at least compliance. Amazon killed its internal AI usage leaderboard entirely after people started doing pointless busywork to climb it.
Decades ago, British economist Charles Goodhart coined Goodhart's Law: The moment a measure becomes a target, it stops being a good measure.
Call centre workers game handle time. Developers whose lines of code are being counted write bloated code. Social media editors manufacture page views with clickbait. In other words, the company measured what was easy to measure, and people optimized for the number.
Charles Holive, chief AI officer at BNP Paribas said: tokenmaxxing is a “vanity metric.” Uber’s Macdonald made the same point, warning there’s no clear line between burning tokens and shipping anything useful. And Ali Ansari, CEO of model-training firm Micro1, told Axios he thinks the whole enterprise world is in a “healthy swing” away from AI overuse, which he’s hoping nudges everyone toward using AI more intelligently.
The planet pays, too: AI's energy cost
All those throwaway prompts (like weather checks, leaderboard padding, and unnecessary tasks put through AI) run on a physical machine somewhere, and that machine uses electricity and water.
One query is tiny on its own. OpenAI pegs the average ChatGPT prompt at roughly 0.34 watt-hours and about a third of a millilitre of water. But multiply that by billions of prompts a day, a meaningful share of them pointless, and it becomes noteworthy. Data centres burned through an estimated 415 terawatt-hours of electricity in 2024, and the International Energy Agency (IEA) expects that to nearly double by 2030. This will push data centres past 2% of the world’s electricity, with a chunk of it still coming from gas and coal.
Reasoning and agentic tasks are the gluttons. Researchers found a reasoning-heavy model can burn over 33 watt-hours on a single long prompt, roughly 100 times what a basic query costs.
Which makes tokenmaxxing worse than a vanity metric. A vanity metric wastes money. This one wastes money and water and power and carbon, all to generate an output no one wants. It’s like leaving every light in the building on to prove you stayed late.
What the research says about AI productivity
AI can boost productivity. But when, how much, and for whom?
McKinsey’s research on generative AI’s economic potential found genuine gains in specific, bounded jobs like customer service, software development, marketing ops, and research. Anywhere AI works on a well-defined task with a measurable output.
McKinsey also flagged that tying tech spending directly to labour productivity is “notoriously inexact,” with results swinging wildly across industries, company sizes and setups. Even OpenAI’s own enterprise data tells a messy story: its 2025 report found a widening gap between power users and average users inside the same company, meaning some people are genuinely getting more done while others are just generating usage.
Which makes a company-wide adoption rate an unreliable way to tell whether an AI program is working.
So who's making money from enterprise AI?
A handful of companies are making real money out of AI. And they’re not doing it by burning the most tokens. They’re focusing on outcomes over activity.
Enterprise AI case studies: Activity vs. outcomes
How these companies achieved an AI ROI
They enable their best people to move faster. JPMorgan built an internal tool called Coach AI for its wealth management advisors. AI isn’t replacing anyone, it cuts the time advisors spend looking for research by up to 95%. When markets turned in early 2025 and clients panicked, advisors sent personalized, portfolio-specific reassurance in minutes instead of hours. The bank credits AI with helping drive a 20% jump in gross sales year-over-year, and thinks advisors could grow their client books up to 50% over the next few years. The AI did the grunt work and the humans did the selling.
They helped customers buy more. Walmart launched a generative shopping assistant called Sparky in June 2025. It answers questions, summarizes reviews, and builds your basket for you. Shoppers who use it order 35% more than the average customer, and Walmart has openly touted the AI wins as revenue climbs. The AI isn’t a cost centre, it’s a salesperson that scales to every app user.
They sold AI itself. Salesforce turned AI into an entirely new product line. Its Agentforce platform, AI agents companies pay for, hit roughly $800 million in annual recurring revenue, up 169% year-over-year, with 29,000 deals signed. When you pair it with a Salesforce data product, that’s a multibillion-dollar business that didn’t exist 2 years ago.
Four reasons corporate AI ROI is so hard to measure
For every JPMorgan, there are lots of companies tripping over structural snags, which don’t show up on a usage dashboard.
Problem one: What are employees using AI for?
“Most people default to automating tasks they dislike rather than tasks most valuable to the company,” Velastegui Ventures CEO Sophia Velastegui (and former chief AI officer at Microsoft) told Axios.
Problem two: Cost.
Even simple queries carry a token cost that multiplies across thousands of employees. This results in costs that finance teams struggle to model.
Problem three: People aren’t ready.
Organizations are still catching up to the tech, and leadership is too. Velastegui calls one common approach “a thousand flowers bloom,” or handing out AI licenses to everyone and praying something useful sprouts. Without a strategy, it mostly doesn’t.
Problem four: Data.
When companies get nervous about handing AI agents access to their proprietary information (a reasonable instinct), those agents get noticeably worse. The tools providing the most value tend to need exactly the access companies are least willing to give.
When AI mandates backfire
Some lean into AI because it genuinely helps. But some lean in purely because they’re being measured.
The human cost is showing up, too. Employees are resisting AI mandates at work, for many reasons, which I write about in this piece:
And companies blaming AI for layoffs are finding that cuts might just be “the only lever they can pull” to cover their massive AI bills. That’s very different than a productivity win.
Real AI winners have a goal before they start
The companies winning at AI are the ones who decided what they were trying to earn before they decided how much to spend. JPMorgan wanted bigger advisor relationships, so it pointed AI at the research and let humans do the human part. Walmart wanted fuller carts, so it built AI that helps you shop. Salesforce wanted a new product, so it sold the agents. In every case the metric came first and the technology came second, the opposite of buying a thousand licenses and praying.
It’s not complicated, just hard: measure the outcome, not the activity. Did you sell more? Serve customers faster? Build something people pay for? If you can’t draw a straight line from the spend to the result, the line Uber’s COO went looking for and couldn’t find, you have a token bill and nothing to show for it.
So here’s the answer to the half-billion-dollar question, for free: Companies don’t make money by using AI. They make money by using AI for something.
AI in the news
Canada says AI strategy will help create 250,000 jobs, boost GDP by 3% (Reuters) Canada unveiled a new national AI strategy, “AI for All,” which PM Mark Carney says could create 250,000 jobs by 2031 and boost GDP by 3%, adding nearly $200 billion to the economy through higher productivity and wider AI adoption. The plan includes 2 new $500 million funds to support Canadian AI companies and help small and medium-sized businesses adopt AI tools, alongside proposed privacy and safety measures aimed at protecting children, combatting deepfakes, and monitoring emerging AI risks.
OpenAI files for U.S. IPO, following Anthropic to public markets (Globe and Mail) OpenAI, the maker of ChatGPT, has confidentially filed for a U.S. IPO, positioning itself alongside rival Anthropic in a race to bring the artificial intelligence boom to public markets. Reports suggest OpenAI could seek a valuation of up to US$1 trillion, making it one of the most closely watched technology listings in recent history. The filing comes amid explosive growth in the AI sector, while competitors such as Anthropic and SpaceX also prepare for blockbuster public offerings. Analysts say the wave of mega-IPOs will test investor appetite for AI and could reshape capital markets by concentrating enormous amounts of investment in a handful of industry leaders.
Why did the AI rally just hit a wall? (Wealthsimple’s TLDR) The AI-driven stock market rally hit a sharp setback last week after months of explosive gains, as investors grew nervous about lofty valuations, a narrowing group of winning stocks, and signs that semiconductor companies may no longer be exceeding sky-high expectations. The selloff accelerated after a strong U.S. jobs report increased the likelihood of higher interest rates, triggering losses in heavily leveraged AI-related stocks.
There’s also a gendered layer here that doesn’t get discussed enough.
In many workplaces, women are already over-indexed on invisible labor: mentoring, coordination, emotional glue, and cross-functional alignment. It's the work that keeps systems functioning but rarely shows up in dashboards.
If AI metrics focus only on visible tool usage, we risk undervaluing that human work even further.
Headcount was a vanity metric. But it was an honest one.
Hiring 200 people meant 200 real salaries, real burn, real risk. It meant you'd convinced 200 humans to bet their careers on you, and convinced investors to fund them.
Then we traded it for tokens. A number a script can manufacture while you sleep.