Accenture, Uber, Amazon, Meta, Walmart, and Cisco all moved to tighten internal AI usage limits in the first half of 2026. The shift came after a year of pushing employees to use AI more aggressively, only for companies to find that spending was rising fast while the business payoff remained hard to explain.
From internal rankings to hard caps
Many of these companies had previously encouraged broad AI adoption. According to the source material, some even created internal rankings tied to AI usage, and Accenture signaled to staff that avoiding AI could hurt promotion prospects. The goal was clear: build a habit of use across the organization.
Once that habit took hold, usage drifted into low-value tasks. Based on an internal Accenture meeting recording obtained by 404 Media, employees were using company token reserves for routine work such as turning PDFs into slide decks. Those actions may save time at a small scale, but they still consume paid model capacity and do not necessarily create direct commercial value.
Accenture agentic AI strategy lead Justice Kwak said the company was reaching a turning point where AI had become a significant part of the cost structure. Spending was growing difficult to predict, and senior executives including CFOs, COOs, and CIOs were still asking whether the money spent on AI was producing real value.
Uber hit its annual AI budget in April
Uber faced an even sharper version of the same problem. The company exhausted its full-year AI budget by April 2026 and then imposed emergency limits. Each employee was capped at $1,500 per month in token usage for agentic coding tools such as Claude Code and Cursor. Before the cap, some software engineers were generating monthly bills of $500 to $2,000.
Uber president and COO Andrew Macdonald gave a blunt assessment: the link between widespread Claude Code usage inside the company and innovation that serves consumers was not there. More code was being produced, but that alone did not establish a clear improvement in the end-user experience.
Why the bills jumped so quickly
The pricing model changed. In 2025, Anthropic and OpenAI mainly sold enterprise access through fixed monthly subscriptions. Companies paid a set amount, employees used the tools, and heavy usage did not always show up as a separate line item in the same way.
By 2026, both companies had shifted most enterprise offerings to token-based billing. A token is the basic unit used to price model input and output. Standard chat interfaces can be relatively manageable, but agentic AI tools are much more expensive because they execute multi-step tasks on their own, including coding, search, and request handling. A single task can consume tens of thousands of tokens.
That change reshaped the cost structure. What had looked like software subscriptions began to behave more like metered compute spending, and autonomous agents had little natural restraint on usage. Inside companies, the idea of “token rationing” started to circulate, treating AI capacity more like travel budgets or software license controls.
Enterprise AI ROI is now under pressure
This is larger than a few companies trying to save money. The source says The New York Times described the trend as “token-minimizing,” while noting that companies were systematically rechecking the ROI of AI spending. Fortune framed it even more sharply, saying the era of tokenmaxxing was over.
Model capability may still be improving, but stronger models do not automatically translate into measurable business gains. That gap is now at the center of the discussion. Executives are no longer focused only on whether employees are using AI. They are asking whether all that usage is tied to outcomes they can actually identify on the business side.