As Uber and Microsoft reassess heavy AI use, rising token bills are forcing tech companies to ask whether agentic tools are delivering enough value to justify their costDennis Bihler|Major technology companies that rushed to embed generative AI across their operations are beginning to confront a less glamorous question: how much does all that intelligence actually cost, and what does it deliver in return? The issue has become sharper with the rise of AI agents, autonomous tools that can write code, search files, run commands and repeat tasks with limited human prompting. Unlike a simple chatbot query, these systems can burn through large numbers of tokens, the units of text processed by large language models, as they reason through a task, check their work and call other tools.2 View gallery How much does all that intelligence actually cost, and what does it deliver in return? That spending is starting to matter. According to recent reporting, Uber has been reassessing its AI usage after internal concerns that the company had exhausted its 2026 AI budget within months. Uber President and Chief Operating Officer Andrew Macdonald said the company has yet to see a clear link between higher token usage and better consumer-facing products, even if more code is being shipped.The question, Macdonald suggested, is not whether employees are using AI. They are. The harder question is whether that use is producing features and improvements that customers can actually feel.Microsoft is facing a similar calculation. The company has reportedly begun pulling back access to Anthropic’s Claude Code for some developers and moving them toward its own Copilot CLI tool. Microsoft has framed the move as a consolidation around internal tools, but it also comes as the cost of AI coding assistants becomes harder for large companies to absorb at scale.The economics are shifting because AI tools are moving from short prompts to longer, agent-driven workflows. Goldman Sachs, according to a report cited by Tom’s Hardware, has estimated that agentic AI could dramatically increase token demand in the coming years. A recent academic study on agentic coding tasks reached a similar conclusion about the direction of travel, finding that such tasks can consume far more tokens than ordinary code chat or reasoning, while higher token use does not always produce better results.That creates a growing tension inside the AI boom. For the past two years, companies have often treated AI adoption itself as a success metric. Executives have boasted about the share of code written with AI, the percentage of engineers using assistants and the scale of internal AI deployment. But the first wave of adoption was often subsidized, either by cloud providers, AI vendors or corporate budgets willing to tolerate high experimentation costs.2 View gallery Microsoft has reportedly begun pulling back access to Anthropic’s Claude Code for some developers Now, tokenized billing is forcing a more disciplined conversation. If an AI agent generates more code but also consumes enormous compute resources, requires human review and does not clearly improve products, the business case becomes less obvious.The problem is especially sensitive for software companies. AI coding tools were supposed to make engineering teams faster and cheaper. But if their usage costs rise faster than measurable productivity gains, companies may find themselves trading one expensive input, human labor, for another expensive input, inference compute.Supporters of heavy AI use argue that this is a temporary phase. They say better models, improved tooling and more efficient chips will drive costs down over time, just as cloud computing became cheaper and more efficient at scale. Nvidia’s next-generation platforms and other inference-focused chips are expected to improve performance per watt and reduce the cost of running large AI workloads.But hardware improvements may not arrive quickly enough to offset the surge in demand. Companies have already committed billions of dollars to data centers and AI infrastructure, and replacing newly deployed chips with newer, more efficient systems is expensive. At the same time, the move from chatbots to agents means that each user action may trigger many more model calls than before.That leaves companies like Uber and Microsoft trying to find a middle path: keep AI tools where they clearly improve productivity, but impose tighter controls where usage becomes disconnected from measurable value.The debate is also likely to reshape how companies talk about AI success. Instead of celebrating raw adoption numbers, boards and investors may increasingly ask harder questions: Did AI reduce costs? Did it improve customer experience? Did it accelerate product development? Or did it simply shift spending from payroll to tokens?For the AI industry, the answer matters. The business model of many AI companies depends on rapid growth in usage. But if customers begin limiting agentic workflows because the bills are too high, the path to profitability could become harder.The AI boom is not necessarily slowing. But it is entering a more expensive and more accountable phase, one in which “more tokens” will no longer be enough. Companies will have to prove that the intelligence they are buying is not just impressive, but worth the bill.