The assumption that artificial intelligence would slash corporate costs is colliding with reality. Several of the world’s largest companies have imposed spending limits on AI tools after seeing bills surge far beyond projections. Uber exhausted its entire 2026 AI budget by April and capped monthly spend at $1,500 for some tools; Meta introduced usage limits following “exponential” cost growth; Microsoft cancelled most internal licences for Claude Code; and Walmart restricted its own agents. Nvidia’s head of applied deep learning research told Axios that compute costs in his team now exceed employee salaries. OpenAI’s chief executive, Sam Altman, acknowledged that the cost of AI has become “an enormous problem” for many clients in 2026, a concern barely voiced a year earlier.

The root cause is a shift from predictable subscription pricing to usage-based models that charge per token consumed. As companies deploy autonomous agents that read documents, write code and execute complex workflows, token consumption—and therefore cost—has grown much faster than anticipated. Workato, a software firm with 1,300 employees, saw its AI bill rise sevenfold in a single day after Anthropic altered its tariff structure. This financial friction is reshaping the calculus around AI and labour. The notion that AI offers a cheap substitute for human workers is being tested, even as some firms continue to link job cuts to automation. Amazon and Intuit have trimmed thousands of positions while increasing AI investment, and Goldman Sachs estimates that accelerating adoption could raise unemployment in the most exposed sectors.

Yet the broader employment picture resists a simple narrative of displacement. An analysis by the Yale Budget Lab found that AI has had a modest impact on America’s job market since the release of ChatGPT, changing jobs more than eliminating them—a pattern similar to the introduction of computers and the internet. PwC’s 2026 global jobs barometer, which analysed over a billion job advertisements, shows that companies most exposed to AI are increasing headcount and wages faster than less-exposed peers. However, the same data reveals a sharp divergence for entry-level roles: AI-exposed junior positions that added senior-level skills grew 35% between 2019 and 2025, while comparable roles that did not “seniorise” fell 10%. PwC itself plans to cut US entry-level hiring by a third over three years, even as it searches for hundreds of engineers.

In Brazil, where payment-by-tap adoption jumped from 3.9% to 72.8% of in-person purchases in five years, analysts at BTG Pactual note that traffic from generative AI to retail sites has surged 4,700% in the past year. They project AI agents could intermediate a quarter of all transactions by 2030, forcing retailers to optimise for algorithms while confronting the risk of radical price transparency and margin compression. Meanwhile, a study published in Human Systems Management, drawing on Italian manufacturing firms, links perceived AI-driven job insecurity to higher burnout, mediated by reduced psychological safety. Workers who fear obsolescence become less likely to voice concerns, amplifying stress.

The next factual milestone to watch is how companies reconcile the push to deploy AI agents with the mounting evidence of their cost volatility. As usage-based pricing becomes the norm, the economic case for substituting human labour with AI will depend on whether productivity gains can consistently outpace the token bill—a question that remains open.