From pushing employees to use AI to fearing burning too many tokens: more and more companies are tightening their internal AI usage quotas.

Companies spent a year forcing employees to use AI, now they have to stop them from using it too aggressively. From Accenture banning employees from using AI to convert PDFs, to Uber burning through its entire annual AI budget by April, to Amazon and Meta simultaneously tightening quotas, the era of "tokenmaxxing" is dead. Executives are still waiting for an answer on whether the AI business model is truly worth the price.
(Previous context: The end of tokens has arrived: GitHub Copilot token price hikes spark negative reviews, as the AI industry shifts entirely to usage-based pricing)
(Background supplement: GitHub Copilot's fee change reveals the "biggest lie" in the AI industry)

Table of Contents

Toggle

• From Leaderboards to Bans
• Why Are Bills Skyrocketing?
• The AI Business Model Called into Question

A year ago, companies told employees: Not using AI could cost you a promotion. Now the same companies are holding meetings to discuss how to stop employees from using AI to create PowerPoint presentations. Accenture, Uber, Amazon, Meta, Walmart, Cisco—these companies almost simultaneously announced tightening internal AI usage quotas in the first half of 2026. Their common predicament is simple: They've poured huge sums of money into AI but can't clearly explain what they've gotten in return.

From Leaderboards to Bans

The absurdity lies in the fact that many companies pushed employees into this trap themselves.

Not long ago, some companies even set up internal AI usage leaderboards to encourage heavy usage. Accenture even hinted to employees that not using AI "could cost you a promotion." This was a reasonable management logic—if you want to drive digital transformation, you need to create usage habits within the organization.

Habits were formed, but usage went astray. According to an internal Accenture meeting recording obtained by 404 Media, employees began using the company's token reserves for basic tasks, such as converting PDFs into presentation slides. These tasks create no business value, but each operation burns money.

Justice Kwak, head of agentic AI strategy at Accenture, pinpointed the core of the problem:

"We are now at a turning point where AI is starting to become a significant part of the cost structure; spending is becoming very difficult to predict, and senior management—especially at the CFO, COO, and CIO level—is still asking whether the money we are spending on AI is actually delivering value."

Uber's situation is even more extreme. The company burned through its entire annual AI budget by April 2026, forcing an emergency cap: each employee's monthly token limit for agentic coding tools (such as Claude Code, Cursor) was set at $1,500. Before the cap, individual software engineers' monthly bills ranged from $500 to $2,000. Uber President and COO Andrew Macdonald's assessment was bluntly uncomfortable: "To connect the company's heavy use of Claude Code with innovation that serves consumers, that link doesn't exist yet."

Why Are Bills Skyrocketing?

In 2025, Anthropic and OpenAI's primary business model was fixed monthly subscription fees. Simply put, companies paid a fixed fee, and employees could use the AI tools, much like subscribing to Office 365—no extra charges for overuse. This model encouraged heavy usage.

However, by 2026, both companies had switched most of their enterprise plans to token-based usage pricing. "Token" is the basic unit for AI models to process text. In simple terms, the model charges for every word it reads and every word it writes. Regular chat interfaces have limited usage, keeping bills manageable. But agentic AI—AI agents that can automatically execute multi-step tasks, such as writing code, searching data, or sending requests—can consume tens of thousands of tokens per task, leading to a completely different billing structure.

This is the root cause of the explosive corporate bills: tools that were previously charged monthly are now billed per computation, and the usage of automated agents is hardly naturally constrained by any human behavior.

"Token rationing" is a term circulating within companies. Simply put, AI usage quotas are being controlled, much like how companies manage travel expenses or software licenses.

The AI Business Model Called into Question

This is not just a cost-saving decision for a few companies; it is the first real stress test for the business model of the entire AI industry.

The New York Times has dubbed this trend "token-minimizing" and pointed out that companies are systematically re-evaluating the return on investment (ROI) of AI spending. Fortune's framing is even more direct: tokenmaxxing is dead; companies haven't gotten the returns they initially expected.

From a technical perspective, AI model capabilities are indeed continuously improving. But between "model capability improvement" and "actual corporate benefit," there remains a gap that has yet to be closed. Uber's Macdonald spoke for many CXOs when he said: Employees use Claude Code to produce a large amount of code, but whether that code actually improves the end-user experience—no one can clearly connect the dots.

The AI industry has moved past the phase where novelty and excitement could cover everything. It must now face a very mundane but critical question: ROI.