Even Uber's leadership didn't see it coming: a report that the company's annual AI budget had run dry in just four months. Earlier this year, the company rolled out a policy encouraging employees to use AI tools freely. The policy was supposed to deliver a leap in productivity — instead, it took less than a season to swing from encouragement to restriction. That rapid reversal is worth a closer look.
"Spend as much as you like" came back as a bill
In early 2026, Uber introduced a policy encouraging employees to freely use AI productivity tools — GitHub Copilot, ChatGPT, and others — on the company's dime. The direction itself wasn't unusual. Global tech companies like Microsoft, Google, and Salesforce have made similar declarations of company-wide AI tool support. The logic goes: invest in AI so employees can boost productivity, automate repetitive work, and spend more time on high-value tasks. Companies were racing to signal — to employees and investors alike — that they were all-in on AI adoption.
What made Uber's case distinctive was how quickly the results showed up in the numbers. According to publicly available information, the budget the company had set for the full year was exhausted in four months. The exact size of the overrun hasn't been disclosed. But for a company with tens of thousands of employees worldwide, the likeliest culprit is API usage fees ballooning faster than flat-rate subscriptions. API costs for coding assistance, document summarization, and data analysis are billed in proportion to usage. Double the usage per person, and the total cost doubles too. When tens of thousands of people start using these tools enthusiastically at the same time, it's not hard to calculate how fast the bill piles up.
Once Uber saw what was happening, it revised the policy, setting per-employee limits on AI usage. From encouragement to control — the gap between the two was four months.
When the person spending the money isn't the one accountable for it
Look closely at this situation and a pattern emerges. A "spend freely on the company's account" policy delegates the judgment of rational resource allocation to each individual employee. From the individual's perspective, it's company money — there's no reason to economize. A sense that you're losing out if you don't use it sets in naturally. Regardless of how much the tool actually improves your work, a usage pattern of "it's there, so I use it" hardens into habit.
Viewed through a lens that treats business and finance as one continuous flow, the problem is that the point where costs are incurred has been separated from the point where someone is accountable for them. The ability to understand where and how spending occurs — and to design that flow in advance — matters just as much as growing revenue. When the people spending a budget don't directly answer for the results, spending tends to drift past the optimal level. It's similar to what happens when a corporate card has no restrictions on where it can be used: charges unrelated to work quietly accumulate. Uber's version of this was vastly larger in scale, and because AI costs are billed by usage, the amplification happened that much faster.
Here's the point worth dwelling on: Uber could likely have avoided this outcome if it had set usage guidelines and departmental limits from the start. The encouragement of "spend freely" and the standard of "spend within this range" can coexist. Having only the former without the latter was probably the direct cause of a policy reversal in four months. With purpose-specific guidelines in place from day one, employees might well have used the tools in a more deliberate way, within their limits.
Caps may actually slow an organization's AI learning curve
That said, it would be hasty to accept Uber's spending caps as simply the rational move. The opposing view deserves a careful look, too.
AI tools can only really be learned by using them. For someone encountering them for the first time, it takes time and experience to construct effective prompts, understand a tool's quirks, and adapt it to their own way of working. Impose cost limits from the start, and employees begin calculating "is this worth my quota?" before they've even tried a new tool. Less experimentation means less learning. A team's collective fluency with AI only sharpens as hands-on experience accumulates — and when the density of that experience thins out, the growth of organizational capability slows with it.
Some companies account for this by deliberately allowing unrestricted use for the first two to three months, then setting team- and role-specific limits based on the usage data collected during that window. The idea is to first learn which teams actually use which tools, and how much, before drawing the lines. If Uber switched to blanket caps after four months, it may come at a cost to how quickly its employees develop proficiency with AI tools.
One more thing worth noting: blowing through a budget doesn't necessarily mean waste. If Uber set limits by looking only at the costs, without measuring how much those four months of AI usage actually contributed to productivity, then the organization may have made a choice that cuts spending while also capping the productivity gains that came with it. A cost cut made without weighing return on investment is hard to classify — is it savings, or a forfeited opportunity? I have no way of confirming whether Uber did that math. But the question itself is one that every organization reviewing its AI spending needs to ask.
Pull up your own list of AI subscriptions
Uber's story may sound like a big-company problem, but it applies directly to solo entrepreneurs and small teams as well.
ChatGPT Plus, Claude Pro, Perplexity Pro, Notion AI, GitHub Copilot… Write out the list of AI services you're currently subscribed to. Add up the monthly fees and see what the total comes to. Each one runs roughly $15 to $30 a month, but stack four or five together and it's not unusual to land at $75 to $110 or more. Then check how many of them you actually used every week over the past month. Your login history or app usage records will tell you quickly. Odds are there's at least one service in the mix that looked useful when you subscribed but that you actually open once or twice a month.
If you use APIs directly, spending alerts are especially worthwhile. OpenAI, Anthropic, and Google Cloud all offer the ability to set monthly or daily usage limits in advance. Without a cap in place, you find out that an automation script processed far more tokens than expected only when the bill arrives. What took Uber four months to discover, an individual can experience within a month, at a much smaller scale. Setting up an alert ahead of time is all it takes to prevent it.
Beyond subscriptions and API usage, there's one more thing to audit: how much time your AI tools actually save you. There can be a gap between a vague sense that something is "useful" and the hours genuinely recovered. If a $25-a-month service saves you one hour a month, it's worth calculating whether that hour is worth more than $25 to you. Run this check once a quarter, and you can keep habitual subscriptions from quietly piling up.
Encourage unlimited use, and spending climbs fast. Impose caps, and learning can slow down. Somewhere between the two is the balance point that fits you. Finding it starts with deciding — before you start using a tool — what you'll spend, and on what. Set the purpose first, choose the tool that fits it, and put a usage ceiling in place ahead of time. Reverse that order, and like Uber, you'll find yourself resetting the limits four months later.
