Enterprise AI spending hits a wall: companies ration tokens, redirect budgets

Some enterprises have run through their entire annual AI budget in three months. Others have watched their monthly AI bills double or triple. The Wall Street Journal reported in May that companies across industries, after spending freely to signal AI ambition to Wall Street, are now scrambling to ration access, redirect employees toward cheaper tools, and build internal systems to track what AI spending actually returns.

The publication named top technical executives at Uber Technologies, Meta Platforms, Microsoft, Salesforce, and DoorDash as among those who have either publicly addressed the need for AI use to generate measurable productivity or taken concrete steps to restrict access to certain AI tools for parts of their workforce. The dynamic reflects a broader reckoning: the experimentation phase of enterprise AI is giving way to something that looks a lot more like conventional capital discipline.

Token costs, not just subscriptions, are driving the squeeze

The unit economics of enterprise AI are more complex than a software license fee. Tokens, the fundamental unit of AI computing that governs how much text or code a model processes per interaction, accumulate rapidly across large organizations. According to The Wall Street Journal, AI model providers are actively balancing supply and demand, which has pushed token costs higher at exactly the moment enterprise consumption is peaking.

The Icecat Blog noted this week that high token consumption, premium model subscriptions, and intensive use of coding assistants have together generated operational costs that many businesses find difficult to justify absent clear business outcomes. The question circulating in boardrooms has shifted from how much AI a company can deploy to where AI demonstrably creates value.

That is a meaningful change in posture. Early enterprise AI adoption was largely decentralized: employees received broad access, usage limits were minimal, and the priority was encouraging experimentation. Centralized cost oversight is now replacing that approach at the companies The Wall Street Journal identified.

Faster output is not the same as stronger performance

One of the central tensions in enterprise AI right now is the gap between activity metrics and business metrics. AI demonstrably accelerates certain tasks: drafting content, generating code, summarizing documents, handling repetitive workflows. But as the Icecat Blog observed, faster output does not automatically translate into higher revenue, better customer experience, or improved operational efficiency.

If an organization's AI deployment allows employees to produce more material without improving the outcomes those employees are measured on, recovering the investment becomes difficult. Executives evaluating multi-year AI strategies are increasingly drawing that distinction. The shift is from measuring prompts and sessions to measuring the downstream business results those prompts were supposed to drive.

The Wall Street Journal reported that corporate leaders are looking at homegrown, lower-cost tools as one lever to bring expenses down while preserving capability. That points toward a bifurcated AI stack inside large organizations: premium frontier models reserved for high-value, well-defined tasks, and lighter, cheaper alternatives handling volume work.

E-commerce teams face a specific version of the ROI problem

Retailers and digital commerce operators are among the heaviest enterprise AI users, applying it to product content generation, customer service automation, merchandising, translation, and shopping recommendations. The Icecat Blog noted that these use cases can deliver strong returns, but only when they are tied to clearly defined business problems rather than deployed opportunistically across every available workflow.

One variable that consistently shapes AI effectiveness in commerce, according to the Icecat Blog, is the quality of the data fed into AI systems. Structured product specifications, accurate attributes, high-resolution imagery, and consistent categorization allow AI to generate more relevant content and recommendations on the first attempt, reducing the volume of iterative prompting needed to reach a usable output. Fewer prompts means fewer tokens, which directly reduces cost.

That connection between data quality and cost efficiency is becoming operationally significant. As organizations apply stricter budget oversight to AI, the teams that can demonstrate a clear link between AI output quality and input data quality will be better positioned to justify continued or expanded investment.

What this means for your team

• Audit current AI tool access by role and use case. If broad access was granted during the experimentation phase, map which use cases are generating measurable outcomes and which are primarily generating token spend.
• Benchmark your token costs against budget on a monthly cadence, not annually. The Wall Street Journal found that some companies consumed a full year's budget in a quarter, a gap that quarterly reviews would have caught far earlier.
• Evaluate whether premium frontier models are the right fit for every workflow, or whether lower-cost, internally built alternatives could handle volume tasks while reserving expensive models for high-complexity, high-value work.
• For e-commerce and product teams: assess product data completeness and structure before scaling AI content or recommendations. Richer input data reduces iterative prompting, lowering per-outcome token cost.