Our Take
Uber's budget blowout wasn't an accident—it was the predictable result of encouraging unlimited use, then discovering that usage volume and business value are not the same thing.
Why it matters
Enterprise AI spending is hitting a wall. When a company with Uber's scale and engineering maturity burns through annual budgets in four months with nothing to show for it, the ROI question stops being theoretical. Other Fortune 500 companies are quietly asking the same question.
Do this week
Finance: audit your actual AI tool usage this week against measurable output (features shipped, bugs closed, revenue impact) before requesting next quarter's budget allocation.
Uber burned its full annual AI budget in four months
In April 2026, Uber's CTO disclosed that the company had exhausted its entire annual artificial intelligence budget within the first third of the year. The ridesharing company then responded by implementing a monthly cap of $1,500 per employee on agentic coding tools, including Anthropic's Claude Code and Cursor (per Bloomberg). The cap is enforced via internal dashboard tracking, though exceptions can be granted on request.
The budget overrun followed a deliberate company strategy. Uber had actively encouraged staff to use AI "as much as possible" and even gamified adoption by ranking employee usage on internal leaderboards, according to reporting from The Information.
The move signals growing uncertainty about AI productivity. During a recent podcast appearance, Uber's COO Andrew Macdonald stated that the company found it "very hard to draw a line" between AI tool usage and actual new consumer features or business results.
The ROI gap is widening across enterprise AI
Uber's situation exposes a structural problem in how enterprises have approached AI spending. Companies have poured billions into tool adoption with the assumption that usage itself would drive value. When Uber discovered that four months of unlimited access produced a budget crisis but no clear product or revenue impact, the underlying model broke.
This is not a Uber-specific failure. It reflects a broader industry pattern: AI has become an expense category that enterprises are learning to manage, not an investment category with clear returns. The cap is less a technical fix and more an admission that encouraging unchecked adoption was a mistake.
For Uber specifically, the inability to trace tool usage back to shipped features raises a second-order problem. If engineering leadership cannot articulate what Claude Code or Cursor have actually produced, how do they justify continued spending? The monthly cap is a rational response to that absence of evidence.
Track usage to outcome, not vice versa
The lesson for other engineering organizations is straightforward: do not replicate Uber's initial approach. Gamifying AI tool usage without tying it to measurable output (lines of code reviewed, deploy frequency, bug resolution time, customer-reported defects) produces exactly what Uber got: high consumption, low accountability, and a budget crisis.
Teams should start by defining what success looks like before any tool is widely deployed. That might be "reduce code review time by 20%" or "increase deployment frequency from 3 to 5 per week." Then measure against that baseline. Unlimited access to Claude Code or Cursor is not a feature; it is a cost center that only becomes useful when connected to a specific outcome.
The cap itself is not the story. The story is that Uber's CTO and COO now publicly acknowledge that usage volume is decoupled from business value. Every other large engineering organization should treat that as a warning, not an anomaly.