In the past few days, the AI industry was still encouraging companies to "spend their budgets", but now it is quickly turning to a "throttling" model, because companies have discovered that even if they only use AI to handle some small tasks, they may easily burn a large amount of token fees, but they may not get the same return. More and more companies are beginning to restrict employees' use of AI tools, and enterprises are entering a stage called "token rationing," which is the rationing of AI resources.
Consulting firm Accenture recently tried to prevent employees from using AI to do basic tasks, such as converting PDFs into presentations, to avoid using up too many tokens. Such tightening comes not long ago, when Accenture warned employees that failure to use AI could hurt promotion opportunities. In a leaked internal meeting recording cited by 404 Media, Justice Kwak, Accenture's head of agentic AI strategy, said the company has reached a tipping point where AI is starting to significantly impact the cost structure and management is still asking whether the investment is really worth the money.
Token costs have begun to shake the AI business model. In the past few months, the stock prices and valuations of AI-related companies have come under pressure, with some companies that rely heavily on AI bearing the brunt, especially memory chip manufacturers. The industry is beginning to realize that AI cannot be supported by "freshness" and "conceptual enthusiasm". It must ultimately prove that it can really bring financial value.
More broadly, companies are collectively adjusting their internal AI strategies. Many companies have begun to set weekly or monthly usage caps for employees, or allocate different token budgets to different positions; some companies will also send reminders when usage is close to the cap, allowing employees to apply for additional quotas. Behind this approach is the fact that service providers such as OpenAI, Anthropic and GitHub have recently adjusted their pricing methods, shifting from a model that was originally more like a "monthly unlimited usage" model to one that emphasizes billing based on actual token consumption.
The report also pointed out that many seemingly simple tasks are actually not cheap to implement on advanced models. For example, offloading complex analysis to a large model running for a long time can easily cost more than $100; a large-scale vulnerability analysis of the entire code base can even cost $50,000 to $100,000. This is why some companies have begun to offload basic tasks to less powerful and lower-cost models, and even mix products from different manufacturers to control expenses.
From a business management perspective, this change means that AI is shifting from “encouraging multiple uses” to “fine quotas.” For employees, access to AI is no longer just an efficiency issue but also a budget management issue; for CFOs, COOs, and CIOs, AI must now be measured as rigorously as other core costs. This also marks that the first round of enthusiasm for enterprise AI is receding, replaced by a new stage that is more pragmatic and cares more about input and output.