AI automates the tasks employees dislike, not the tasks that generate revenue.
A few days ago, GeekPark reported that Microsoft, which has made a major bet on AI, quietly suspended most employees' access to Claude Code internally.
This is highly unusual, because in this wave of AI adoption, the biggest marketing point for enterprise users has been "increasing efficiency." If AI boosts efficiency, why is Microsoft stopping its employees from using Claude Code?
Microsoft is not the only company doing this; reducing token usage and no longer encouraging employees to engage in frantic vibe coding has become a new trend among major Silicon Valley tech firms.
Uber spent its entire annual AI token budget in four months. Salesforce writes Anthropic checks of about $300 million per year. One AI consultant revealed that one of his clients spent as much as $500 million on AI in a single month. Meta even quietly retired its internal “tokenmaxxing” leaderboard—a ranking system originally designed to encourage employees to use AI more.
Today, companies are doing something that would have been unthinkable just a few years ago:
Restrict and monitor employee use of AI.
Why are major companies all shifting direction?
"Tokenmaxxing," a reflection of the times
To understand today’s cost crisis, you first need to grasp what “tokenmaxxing” is.
This term began gaining popularity around 2025 and literally means "maximizing token usage." It reflects a management logic: since companies have invested heavily in AI tools, employees should use them aggressively—the more they use them, the more they demonstrate "digital transformation"; using them sparingly is seen as wasting resources. As a result, many companies have implemented usage quotas, leaderboards, and even performance evaluations to push employees to adopt AI tools.
What’s the result?
Employees began using the company’s enterprise AI model to check the weather, write birthday messages, and ask what to eat for lunch.
A study of 2,444 companies found that for every dollar companies spend on AI tokens, $0.44 goes toward fixing AI-generated bugs, $0.27 toward rewriting AI-generated code, and $0.11 consumed by review and merge delays.
In other words, behind every dollar of AI procurement cost lies nearly 80% in hidden losses.
Investor Shruti Gandhi offered a precise analogy: “A tokenmaxxing company is like a business that measures productivity by leaving all the lights on—spending more money doesn’t mean producing more.”
More ironically, most of these companies have no idea what their employees are using AI for, let alone whether AI has made any difference in completing those tasks.
This "burning money" race, which began in 2024 and carried into 2025, finally exploded this year. JPMorgan released a sharply worded report with a blunt, uncomfortable title—“AI Token Costs Are Consuming Internet Profits.”
Shopify, Spotify, ServiceNow, and Roku all mentioned on their earnings calls that AI has become a primary source of operating expense pressure. The overall industry sentiment is shifting from "how amazing AI is" to "whether this spending is worth it."
When the CEO begins to question the ROI
Only 14% of CFOs report seeing clear, measurable returns on their AI investments.
Uber’s Chief Operating Officer, Andrew Macdonald, said something very candid on a podcast—he found it difficult to connect improvements in individual employee productivity to the company’s overall business impact. “If you can’t see how AI has helped you deliver valuable features to users, it’s even harder to justify the token costs.”
This highlights the core of the enterprise AI dilemma: increased individual efficiency does not equate to increased company revenue.
Employees wrote weekly reports three times faster using AI, but company revenue remained unchanged. Engineers generated code twice as fast using AI, but the code “churn rate”—the proportion of code discarded or rewritten—increased by 800%.
Former Microsoft Chief AI Officer Sophia Velastegui said something that makes many managers uncomfortable: “Most people default to automating tasks they dislike, rather than those that add the most value to the company.”
In short, businesses automate the tasks employees dislike, not the tasks that generate revenue.
This isn’t a technical issue—it’s a priority issue. That’s why about 30% of generative AI projects are abandoned at the proof-of-concept stage: when costs and value are unclear, executives naturally stop funding.
Salesforce CEO Marc Benioff’s approach is representative. Facing an annual $300 million bill for Anthropic, he envisioned a “smart router” that could determine which queries warrant top-tier models and which can be handled adequately by cheaper, smaller models.
The idea itself isn’t new—since the era of cloud computing, “pay-as-you-go” and “resource optimization” have been standard practices. But the AI wave came too quickly, and many rushed to buy before thinking things through; now they’re playing catch-up.
A return to rationality, or the prelude to a bear market?
Microsoft recently canceled most of the enterprise licenses for Claude Code, citing cost factors as the official reason. This move has sparked considerable discussion in the industry—after all, Microsoft is OpenAI’s largest investor, yet it is also cutting subscriptions to a competing product. It’s difficult to determine how much of this decision is driven by cost considerations versus strategic positioning.
But regardless, it signals that companies are beginning to vote with their feet.
Harness and CloudZero released their AI cost management tools almost on the same day—May 28—each introducing a solution: one focused on real-time monitoring of AI spending and ROI, and the other launching an “AI financial control plane” to help businesses tie every dollar of AI expenditure to specific business outcomes.
The very existence of these two products demonstrates the point: there is demand in the market, and it is urgent.
Since April this year, HubSpot has adjusted its pricing model for AI agents, moving away from charging by tokens to billing based on the number of resolved conversations or generated leads—a strategic shift that aligns vendor incentives with the customer’s actual outcomes. ServiceNow is making similar changes. AI vendors are increasingly recognizing that if they continue to sell "usage" rather than "results," enterprise customers will eventually push back collectively.
Is this adjustment a necessary pain in the industrialization of AI, or the prelude to a greater crisis?
I tend to lean toward the former. But one detail is somewhat concerning: global AI software spending is projected to reach $2.59 trillion by 2026, a 47% year-over-year increase, yet 94% of engineering leaders say key ROI metrics are still missing. More money is being spent, but no one knows where it’s going or whether it’s worthwhile—unless this contradiction is resolved, the next “tokenmaxxing” moment is just a matter of time.
A Fortune magazine analysis put it bluntly: “Tokenmaxxing is easy; redesigning workflows is hard.” Most companies today are optimizing existing processes rather than reinventing their business models. This is where the true value of AI lies—and where most businesses have yet to arrive.
Rational regression is a good thing. But after rational regression, companies still need to answer a harder question: Should AI be a hammer for our business, or a new framework for thinking?
If you merely use AI to do old tasks faster, the bill will eventually force you back to this question.
This article is from the WeChat public account "GeekPark" (ID: geekpark), author: Hualin Wuwang, editor: Jingyu