In the second quarter of 2026, Uber consumed its entire year's AI budget within three months. According to internally disclosed information, token consumption grew 700% year over year, yet brought no corresponding improvement in feature output and no optimization of any roles. This phenomenon is widespread across the industry. Over the past year, most companies pushing AI transformation have seen similar situations: rapid growth in AI usage alongside a synchronized surge in costs. Multiple executives from OpenAI and Anthropic have acknowledged at recent industry conferences that AI's substitution effect on white-collar jobs has fallen short of prior market expectations, and both companies' official narratives have shifted from "replacing human labor" to "expanding employees' scope of work." The core problem with enterprise AI adoption today is "Utility" and "Economic Value" creates a disconnect. For individual users, using AI to debug a piece of code or generate a document—saving a few minutes—is clear value. But for an enterprise, if a single employee's time savings cannot be converted into increased overall output, reduced costs, or revenue growth, then the token fees paid for it are pure cost. When a company opens AI to all employees as a general entitlement without any usage controls, runaway costs become inevitable.
The Chaos of Unregulated AI Usage
At present, most companies' approach to managing AI is essentially a "credit-card style" openness: grant employees usage rights, set no limits, track no purposes, assess no output. This model directly leads to systemic resource waste. Statistics from third-party organizations show that about 30% of code generated by enterprise employees using AI never goes live. A Microsoft engineer mentioned in a public talk that some teams engage in "token farming to prove AI-native status," and some companies have even set up internal token-usage leaderboards, folding AI usage into employee performance reviews. Duolingo was among the earlier companies to try incorporating AI usage into KPIs. It once required all employees to use AI tools at work and treated usage volume as a performance metric. But this policy was scrapped after just three months, because employees fell into a "using AI for the sake of using AI" loop, workflows became cumbersome, and overall efficiency did not improve.
The Governance Framework Taking Shape
In response to the above problems, some leading companies have begun exploring the establishment of an AI governance system, with the core being to transform AI from "employee benefit" into "auditable production resource". The governance frameworks proven effective so far mainly consist of four components. First, a task-screening mechanism. The AI return on investment varies significantly across different tasks. High-ROI AI scenarios typically share three traits: clearly defined tasks, easily verifiable results, and long manual completion times—for example, complex code debugging or bulk document information extraction. By contrast, simple repetitive tasks or vaguely specified tasks usually have a low AI ROI, such as routine document lookup or simple email drafting. Second, value-assessment criteria. Token usage is a core metric for AI companies, but it should not become the standard by which enterprises assess AI effectiveness. Truly effective evaluation dimensions include: whether delivery quality improved, whether rework rates fell, whether project delivery cycles shortened. If a hundred-person team spends over a million dollars a year on AI but shows no improvement in core business metrics, that investment is ineffective. Third, decision-rights allocation. The choice of AI model cannot be fully delegated to individuals, nor uniformly mandated by headquarters for all scenarios. The reasonable approach is to build a dedicated AI governance team that matches model resources to task complexity: use strong models for core tasks above a complexity threshold, low-cost lightweight models for simple auxiliary tasks, and restrict AI use for tasks with no clear value. Finally, a feedback-loop system. Mandatorily record the deliverable and business outcome corresponding to every token, build an AI-effectiveness heatmap, and clearly visualize which scenarios create value and which waste it. Use real data to guide subsequent AI investment decisions, rather than relying on experience or intuition.
The Deeper Logic of AI Transformation
The underlying logic of enterprise AI is completely different from that of consumer (To C) internet. The logic of To C internet is "more is better," with user count and usage duration as core metrics. But the logic of enterprise AI is "precision deployment", with the core being to obtain maximum business value at minimum cost. Many companies reduce the challenge of AI transformation to a technology problem, believing that success just requires buying the best model and having enough compute. But in actual implementation, the core challenge of AI transformation is organizational and managerial capability. When AI becomes a basic production resource like water and electricity, how to allocate resources, how to assess value, and how to control costs are what determine the success or failure of transformation. Recently, leading AI companies have adjusted their narratives ahead of IPOs, shifting from "general AI changes everything" to "enterprise AI application deployment," a change that also reflects an industry consensus: the current value of AI technology is not yet enough to justify unlimited investment. Future AI competition between enterprises will not be about who uses more AI, but about who can manage AI usage more efficiently. Organizations that have not built a complete AI governance system will continue to face the tension between "employee usage experience" and "enterprise cost control."
