What Uber and Tesla's AI spending caps teach finance teams.

01What happened.

The sequence was the same at each company. Leaders encouraged staff to use AI, usage climbed quickly, and the bill arrived ahead of the plan.

In April, Uber's chief technology officer, Praveen Neppalli Naga, told The Information that the AI budget he had planned for was already spent. Fortune later reported that Uber had used its entire 2026 AI coding tools budget in four months, after teams were ranked by how much they used AI tools. Asked whether the spending was showing up in productivity, Uber's president and chief operating officer, Andrew Macdonald, said: "That link is not there yet."

The caps followed within weeks. On 1 June, Bloomberg reported that Walmart had given each employee a fixed token allotment for its in-house coding agent, where use had been unlimited. The next day, Uber set a monthly cap of $1,500 per employee for each agentic coding tool, including Claude Code and Cursor, which staff can exceed with permission, TechCrunch reported. Later in June, The Information reported a Meta memo to about 6,000 staff describing a platform to track spend in real time, set budgets and cap token spend. From 6 July, Tesla capped employee AI spending at $200 a week, with sign-off required above it, according to an internal memo reported by The Information and Electrek.

TakeawayUber, Walmart, Meta and Tesla each encouraged AI use, saw costs outrun the plan, and answered with caps in June and July 2026.

02Why the budgets ran out early.

Coding agents, like other agents, are priced by use. A seat licence costs the same whether it is used or not. An agent's bill grows with every task it takes on, every file it reads and every retry. Anthropic's documentation puts the average cost of Claude Code at around $13 per developer per active day, or $150 to $250 a month, with 90 percent of users under $30 a day. Averages hide the heavy users, and the leaderboards rewarded heavy use.

Gartner expects the pressure to grow. In June 2026 it predicted that by 2028 the cost of AI coding tools will exceed the average developer's salary, measured against a global average, as token consumption surges. Its analyst Nitish Tyagi described budgets being depleted earlier than expected. In August, Gartner predicted that inference costs per agentic workflow will increase more than fivefold through 2028.

Finance has already taken the work on. The FinOps Foundation's State of FinOps 2026 found that 98 percent of respondents now manage AI spend, up from 31 percent two years earlier.

TakeawayAgents are priced by use, so a budget written like a licence budget runs out early, and 98 percent of FinOps teams now manage AI spend.

03What the caps got right.

• A number per person and per period. Everyone knows the limit before they start, and a weekly period, as at Tesla, shows an overrun within days.
• A route for exceptions. Uber and Tesla both let people go past the cap with permission or sign-off, so work that justifies more spend can still get it.
• One rule for everyone. A plain limit is easier to explain and apply than a review of every request.

TakeawayA known number per person, a short period and a route for exceptions are the right base.

04What a cap per person leaves out.

Each of these caps follows a person using a tool. That fits a developer at a keyboard. Once agents do more of the work, four gaps open.

• Agents spend without a person. A scheduled agent, or one triggered by an event, can run at three in the morning on a shared key. A cap on people never sees it.
• Caps per tool add up. $1,500 per tool across three tools is $4,500 per person, and the total is visible only when the invoices are added together.
• The cap arrived after the bill. Each company set its limit once the money was spent. A limit written into the budget at the start of the year holds the spend it was meant to hold.
• Spend is half the picture. The leaderboards counted usage and the caps count dollars. Neither shows what the spend bought, which is the question Uber's chief operating officer was asked.

TakeawayCaps on people miss the agents running on their own, the total across tools, the timing, and what the spend bought.

05How to set AI budgets for 2027.

Budget season is the moment to put limits in before the spend, at every level the money moves through.

1. Set the ceiling. The organisation's total AI spend for the year, across every provider and tool, approved like any other line.
2. Allocate it to teams. Each team's share, owned by its lead, in monthly amounts so an overrun shows within weeks.
3. Cap people and agents. A limit per person for the tools they use, and a limit per agent for the work that runs without them, both counted across every provider.
4. Name the approver. Every increase carries a named person, a reason and an expiry, and approval takes minutes, so nobody routes around the cap.
5. Review cost per outcome. Each month, compare what each team and agent spent with what it delivered, and move budget towards the work that returns the most.

Our comparison of AI spending limits shows what each provider lets you cap natively, and the guide to token budgets and spending limits for AI agents covers what should happen at each cap.

TakeawaySet a ceiling, allocate it to teams, cap people and agents across every provider, name the approver, and review cost per outcome monthly.

06Questions for the budget meeting.

• What did we spend on AI this year, across every provider and tool, and how long did that number take to produce?
• Which teams went over, and when did we find out?
• How many agents run without a person attached, and what do they spend?
• Who approves an increase, and how long does it take?
• What did each team's AI spend deliver?
• If usage doubles next year, which limit holds first?

TakeawayIf any answer takes more than a day to produce, the budget is being read after the spend.