The bill has arrived. For two years, companies handed employees access to AI tools and told them to use as much as they could. Now the invoices are landing, and they do not match the budgets.
Uber's chief technology officer admitted the company was "back to the drawing board" after burning through its entire planned 2026 AI budget in just four months. The company has since capped monthly spending at $1,500 per employee per AI coding tool, with employees able to request more if they can justify it. Royal Bank of Canada disclosed that its AI usage surged 500% since last year. Meta, facing internal AI costs climbing toward billions of dollars, is building a real-time spending dashboard and plans to impose formal budgets by 2027. An unnamed company reportedly spent $500 million on Anthropic's Claude in a single month after giving employees access with no spending limits at all.
The reason the numbers are so surprising is the shift from chatbots to what are called "agentic" AI tools: software that works on its own, making decisions and completing tasks in a series of steps without waiting for a human to guide each one. Every step in that process generates a charge. Gartner's analysis found these tools use 5 to 30 times more tokens per task than a standard AI chat tool. Goldman Sachs projects total token consumption across all businesses will be 24 times higher by 2030 than it is today, driven almost entirely by this shift.
The cost structure is also unpredictable in ways traditional software was not. Per-token prices have actually dropped about 75% over the past year, yet total AI spending for most businesses is still rising sharply. One spending tracker found that token usage among businesses grew by over 1,000% between January 2025 and April 2026, with total dollar spend up nearly 500% in the same window. Cheaper per unit, but used so much more heavily that bills go up anyway. That pattern has caught many finance teams off guard.
What makes it harder still is model choice. There is a wide range of AI tools at different price points, and employees tend to reach for the most powerful, most expensive ones, often without realising the cost difference. One analysis found that the most expensive AI model outputs cost 83 times more than the cheapest available option. When a team quietly upgrades to a more capable model for quality reasons, the finance team only finds out when the bill arrives.
Some companies are doing this better than others. Software firm 8x8 says it has saved roughly $5 million a year by cancelling software subscriptions it no longer needs, and its AI bill is still comfortably below that figure. It monitors usage through a shared dashboard and is considering requiring staff to prove that cheaper AI models cannot do the job before approving access to the more expensive ones. That is a practical approach: treat model selection as a spending decision, not just a quality decision.
Clothing brand Baseball Lifestyle 101 is going further in the opposite direction, actively pushing its managers to spend heavily on AI and treating it as a revenue investment rather than a cost to minimise. The company says the AI recently identified a retailer running low on a popular product, turning a fast inventory insight into a $1 million order. Whether that approach works at scale across an entire workforce is the open question most companies are still trying to answer.
The honest picture is that most companies cannot clearly link higher AI spending to better results. Uber's own chief operating officer said it is "very hard to draw a line" between AI usage and actual consumer features shipped. An academic study tracking more than 100,000 developers found that AI coding tools produced a 741% increase in lines of code written, but only a 20% increase in software actually released. More activity is not the same as more output.
What separates the companies managing this well from those absorbing bill shock is not whether they use AI. It is whether someone is watching what it costs and tying that cost to a measurable result. The companies setting spending dashboards, requiring justification for premium models, and reviewing which tasks actually benefit from autonomous AI will get more value for less money than those that simply handed out access and hoped for the best.