For months, big tech pushed employees to use more AI. Now the bill is coming due.
Amazon’s decision to shut down an internal AI leaderboard shows how fast the mood is changing. Companies still want AI in the workplace, but AI costs are forcing a tougher question: how much of that usage is useful, and how much is waste dressed up as progress?
Amazon’s AI leaderboard created the wrong incentive
Amazon’s internal leaderboard was meant to push adoption. Instead, it appears to have pushed behavior that cost money without adding much value.
The company had set targets for more than 80% of developers to use AI each week. On paper, that sounds like a clean way to speed up adoption. In practice, it turned AI usage into a number people could chase. Once a score becomes the goal, people start playing to the score.
That is what reportedly happened. Workers tried to climb the rankings by assigning AI agents to needless tasks. The result was higher computing costs, because every extra request burns tokens. Amazon then shut the leaderboard down.
A token is a small unit of data processed by an AI model. More tokens mean more usage, and more usage usually means more cost. When employees started inflating token consumption, Amazon ended up paying for activity that did not clearly improve products or workflows.
Amazon senior vice president Dave Treadwell reportedly told employees not to use AI for its own sake. That message matters because it cuts through a problem many companies created for themselves. They told people to use AI more, measured them on usage, and then discovered that usage alone says very little about value.
Amazon is not the only company to run into this. Reports have said Meta employees tried to boost their standing on internal tables by driving up token use as well. That broader pattern lines up with Tom’s Hardware’s report on token maxing, which described how rising AI costs are pushing companies to rethink these incentives.
The lesson from Amazon’s leaderboard is simple. If a company rewards AI activity instead of good work, people will find ways to create activity.
Why AI costs are suddenly harder to ignore
The cost problem is showing up at the same time as an enormous buildout in AI infrastructure. That makes every wasted prompt harder to shrug off.
Amazon has already gone through sweeping layoffs while pouring money into AI. More broadly, the spending wave across big tech is massive. One estimate cited in reporting put capex at around $200 billion, with most of it tied to AI and data center infrastructure. That kind of outlay raises the pressure for clear returns.
Then pricing changed. AI providers such as Anthropic have moved toward consumption-based pricing instead of simpler flat monthly fees. That means the meter keeps running with every request, every output, and every extra round of use. When companies expand AI access across thousands of employees, costs can jump far faster than leaders expect.
This matters for Amazon because it uses Anthropic’s AI models extensively. If the underlying model provider charges by consumption, then sloppy internal use becomes expensive fast. A workplace can burn through money one prompt at a time.
Model providers have their own cost problem too. They need to balance supply and demand while handling the heavy expense of training and running large models. So even as companies push harder on adoption, the tools themselves are no longer being treated like cheap, unlimited software seats.
That is why the tone has shifted. A few months ago, the push was simple: get people using AI. Now the message is more cautious: use it where it helps, and stop paying for motion that looks busy but goes nowhere.
That wider pressure is also reflected in Yahoo Finance’s coverage of aggressive AI spending, which framed the issue as a cost pileup, not an automatic productivity win.
Big AI bills are exposing a weak spot in the hype
The hardest part of this story is not that AI costs money. Everyone knew that. The harder truth is that many companies still can’t show that all this spending is producing enough useful output.
Axios recently reported that an AI consultant described a client that received a bill of about $500 million for a single month after failing to put usage limits on Claude licenses for employees. That is the kind of number that turns enthusiasm into budget panic. Some enterprises, according to the same reporting thread described in the video, burned through annual AI budgets in three months. Others saw monthly AI bills double or triple.
Uber has run into a similar question. COO Andrew Macdonald said the company has struggled to clearly justify some of its AI spending. That doesn’t mean AI has no value. It means the math is still blurry in too many cases.
Inside some companies, executives started calling the problem “token maxing.” The phrase sounds technical, but the issue is easy to understand. Developers can consume huge amounts of AI compute without a reliable way to prove that the extra spend leads to better code, stronger products, or faster delivery.
That is where the hype runs into the invoice. A team may feel productive because AI is active all day. Dashboards may show soaring use. Yet the real test is still the old one: did the work improve, and was the improvement worth the cost?
That shift from blind expansion to tighter controls matches The Wall Street Journal’s reporting on corporate AI rationing, which described how companies are starting to put limits on AI use as costs climb.
When business leaders can’t connect usage to outcomes, AI stops looking like a shortcut and starts looking like a budget line that needs a hard review.
The pullback is spreading across the industry
Amazon’s move is part of a larger pullback. Companies are not abandoning AI, but they are trimming the parts that look wasteful, forced, or hard to justify.
Microsoft has reportedly started reducing some external AI coding subscriptions, including Anthropic’s Claude Code. Duolingo recently reversed an internal policy that linked employee performance reviews to AI usage metrics after staff complained that they felt pushed to use AI even when it did not improve the work. That complaint goes to the center of the problem. When a tool becomes a performance target, people use it to satisfy the target.
The pattern becomes easier to see when you put the company moves side by side.
| Company | What happened | What it suggests |
|---|---|---|
| Amazon | Shut down an internal AI leaderboard after employees inflated usage | AI metrics can drive waste |
| Meta | Employees reportedly tried to climb internal tables by using more tokens | Ranking systems invite gaming |
| Microsoft | Reduced some outside AI coding subscriptions | Companies are trimming spend |
| Uber | Said some AI spending is hard to justify | Usage is outpacing proven value |
| Duolingo | Reversed policy tying reviews to AI usage | Mandates can create bad incentives |
The takeaway is not that these firms misread AI’s promise. It is that workplace adoption is much messier than a rollout deck suggests. A company can buy access, set targets, and publish rankings in a week. It takes much longer to find the tasks where AI saves time, improves quality, and does not create expensive rework.
That gap between rollout speed and real value is where many of these problems begin.
Why blanket AI mandates keep backfiring
The cases above point to the same issue. Many companies pushed AI across whole teams before they built clear rules for where it helps, how to measure success, and when human review still carries most of the load.
That creates a bad loop. Employees are told to use AI more. Managers look at usage metrics because they are easy to track. Workers respond to the metric. Then the company pays for prompts, revisions, retries, and corrections, while the final output still needs heavy human editing.
In that setup, token counts rise faster than real productivity. A team can look modern and busy while the actual gain stays modest.
The Duolingo example is especially telling. When employees said they felt pushed to use AI even when it did not improve their work, they were describing a common failure in plain language. A tool should help the job. It should not become a ritual.
Amazon’s abandoned leaderboard points to the same flaw. Ranking systems work best when the score reflects real performance. AI usage does not. A developer who sends ten prompts is not always doing better work than one who sends two. Sometimes the stronger result comes from using less AI, not more.
Perks can distort behavior too. When companies reward adoption with praise, rankings, or review credit, they create a workplace where appearing AI-forward matters more than choosing the best method for the task. That is expensive when pricing is tied to consumption.
The result is a strange kind of waste. Companies buy AI to save time, then spend more time checking, fixing, and paying for output that did not need to exist in the first place.
What balance looks like when AI stops being cheap
The shift now underway is less about retreat and more about discipline. Big tech still wants AI inside everyday work. What is fading is the idea that more usage automatically means more value.
A balanced rollout starts with proof, not pressure. Companies need to know which tasks benefit from AI before handing licenses to everyone. They also need metrics tied to outcomes, such as code quality, speed to completion, or lower error rates, rather than raw prompt counts.
That sounds obvious, but the last year showed how often companies skipped that step. They rushed to distribute tools across teams, partly because no one wanted to look late to AI. Once the spending scaled up, the missing structure became visible.
Consumption-based pricing makes that discipline even more important. When every extra interaction adds cost, careless use stops being a small issue. It becomes an operating problem.
There is also a human side to this. If workers feel pushed to use AI when it does not fit the task, trust erodes. Some will comply for the metric. Others will resist because the tool slows them down. Neither response helps the business.
The smarter posture is more grounded. AI is useful in some workflows, weak in others, and expensive when treated like an always-on badge of progress. Companies that figure out where the tool earns its keep will still move ahead. The ones that chase broad usage for its own sake may end up with soaring token bills and thinner returns.
The bridge between today’s workflows and tomorrow’s AI promise is not built by counting prompts. It is built by proving where the tool works.
Final thoughts
The cost meter is now visible, and that changes everything about how companies talk about workplace AI.
Amazon’s abandoned leaderboard shows the problem in one sharp example. When firms reward usage instead of useful results, employees learn how to consume tokens, not how to create value.
AI is still moving deeper into big companies. Yet the next phase will look less like a race for adoption and more like a search for discipline, proof, and returns that can survive a monthly bill.
