The promise of AI was a worker who never sleeps, never books leave, and never asks for a raise. The reality is messier: for a growing number of firms, the machine meant to replace payroll is starting to look like a payroll of its own

Uber spent its entire 2026 budget for AI coding tools in four months. By April, after an assistant called Claude Code had spread across its engineering teams, the money set aside for the whole year was gone. Somewhere in a finance review, someone had to say that sentence out loud. The company then capped what any single employee could spend on such tools at $1,500 a month.

Picture the meeting. A line item that was supposed to last until December is empty by April. Engineers are still typing, the tool is still billing, and someone in the room has to explain how a productivity tool ate a full year of budget in a single quarter. Uber is not a small shop caught off guard. It is one of the more sophisticated engineering organisations in the world, and it lost track of the meter.

That number cuts against the whole reason these tools got inside so many companies in the first place.

The pitch that sold a generation of CFOs

Through 2024 and 2025, AI was sold as a tireless stand-ins for human workers. The pitch was simple and seductive: a worker that runs around the clock, needs no salary, no benefits, no holidays or sick leave, and never asks for a raise. On a spreadsheet, a fixed fee against a full human salary looked like an easy win.

But the way these tools charge is having an effect. Most bill by usage for the sort of applications many companies need: you pay for each small unit of work the tool does, measured in what the industry calls tokens. The more your engineers use it, the more it costs, and there is no natural ceiling.

Gartner expects global spending on AI agent software to reach $207 billion in 2026, up sharply on the year before. A tireless worker turns out to bill by the minute.

Where the maths started to slip

The clearest signal came from inside the industry building the tools. Bryan Catanzaro, a vice president of applied deep learning at Nvidia, was blunt about his own group: “For my team, the cost of compute is far beyond the costs of the employees.” That is one executive describing one team at one moment, not a law of AI economics. But it is a striking admission from a company that sells the chips those costs run on.

The billing data points the same way. Figures from Gartner Peer Insights, cited in industry coverage, suggest 23% of tech leaders already spend between $200 and $500 per developer each month on these tools. Another 6% of companies spend more than $2,000 per developer a month. At the top end, the tool now costs more than a junior hire in some markets.

Gartner’s headline forecast makes the direction clear. Its analysts predict that by 2028, AI coding costs will overtake the average developer’s salary, working from a global average of about $2,000 a month. Gartner treats this as a management problem, not a certainty. As analyst Nitish Tyagi puts it, “Token discipline will not emerge through developer choice alone, as developers tend to optimize for speed and convenience over cost efficiency.” Left alone, people reach for the fastest answer, not the cheapest one. That is not a flaw in the workforce; it is what the tool was designed to encourage.

The firms now counting the cost

Uber is perhaps the clearest example, and its internal numbers show why the budget vanished. Monthly Claude Code costs per Uber engineer ran $150 to $2,000, and the company’s own chief technology officer reportedly spent $1,200 in a single two-hour demo. Uber’s president and chief operating officer, Andrew Macdonald, has been candid about how hard it is to link that spending to a return, saying “it’s very hard to draw a line” between the money going out and the value coming back.

Microsoft hit the same wall from a different direction. Its division covering Windows, Microsoft 365, Outlook, Teams and Surface set engineers a deadline of June 30, 2026 to move off Claude Code and onto GitHub Copilot, after usage billing chewed through the annual budget ahead of schedule. When a company that part-owns a major AI lab is switching tools to manage its own bills, the cost pressure is not a fringe complaint.

Tyagi’s warning is worth keeping in mind. “Without a governed engineering operating model, costs can escalate faster than the productivity gains these tools are designed to deliver,” he says. The word doing the work there is “can”. The tools are not always more expensive than the staff they replace. They become so when nobody is watching the meter.

What the reckoning actually looks like

The running cost is only half the story. The other half is showing up in hiring, where firms that cut people for AI are quietly reversing course. A survey cited by CNBC found that 55% of employers who laid off staff and replaced them with AI now have some regrets about the decision. Separately, Robert Half reported that 32% of U.S. hiring managers who cut a role mainly because of AI later rehired for the same or a similar position.

The savings, on closer inspection, were often smaller than promised. Analysis cited by Forbes estimates that once you count severance, lost productivity and the cost of replacing people, companies spend about $1.27 for every $1 they save by cutting staff. To state the obvious, a cut that costs more than it saves is not a cut.

There is also a longer worry beneath the short-term budget pain. IBM’s chief human resources officer, Nickle LaMoreaux, frames it as a question about the future rather than a settled outcome. If companies stop hiring entry-level workers because AI can handle the junior tasks, she asks, what happens in a few years? Her own answer is stark: “There’s no pipeline; the well simply dries up.” That is a caution about a possible future, not a measured result, but it names a cost that no meter captures.

The firms handling this well tend to treat AI coding tools less like a hire and more like an electricity bill: a metered utility that needs budgets, caps and someone accountable for the spend. Uber’s $1,500 monthly limit is roughly that idea in practice. It also concedes the original pitch was wrong. A worker that never asks for a raise was never the right analogy for a service that charges by the token; the correct analogy was the taxi meter, and the industry chose not to use it. The interesting question is no longer whether a machine can replace a developer. It is whether the companies buying these tools were ever really pricing them, or just believing the sales deck.