Uber caps staff AI spending after blowing its budget

TL;DR:

• Uber has introduced a $1,500 (~£1,180) monthly cap per employee for each agentic coding tool, including Anthropic’s Claude Code and Cursor.
• The move follows the company burning through its entire annual AI budget in just four months.
• Usage is tracked on an internal dashboard, with caps exceedable only by permission — a sharp turn from earlier encouragement to use AI “as much as possible”.

The reversal is striking because Uber had actively pushed adoption, reportedly ranking employees on internal usage leaderboards. Having told staff to lean in, the ridesharing firm is now rationing the very tools it championed, a swing that captures how quickly enterprise enthusiasm for AI can collide with the bill.

The ROI question, made concrete

Uber’s CTO revealed in April that the company had exhausted its annual AI budget in four months; COO Andrew Macdonald has since cast doubt on the productivity payoff, saying it is “very hard to draw a line” between AI usage and new consumer features. That candour is unusual. Across the industry, returns on heavy AI spending remain largely anticipated rather than demonstrated, and Uber’s caps are an early sign of firms growing restless while they wait.

For UK businesses, the lesson is less about the specific dollar figure than the governance gap it exposes. Agentic coding tools bill by consumption, so costs scale with use in ways traditional per-seat software licences never did — and “use it as much as possible” is a dangerous instruction when each prompt has a price. UK finance and engineering leaders watching their own AI invoices climb will recognise the pattern: enthusiasm without metering produces exactly this kind of overshoot, and it reinforces the case for tracking value, not just spend.

Looking forward

Expect more organisations to move from open-ended access to metered budgets and per-tool caps as the novelty of agentic tooling wears off. The harder discipline is measurement — tying spend to demonstrable output rather than activity. Until firms can reliably draw Macdonald’s missing line between AI use and results, caps are a blunt but rational way to keep experimentation from outrunning the returns it is meant to deliver.