Your AI Agents Have a Budget. Do They Have a Meter?

Uber's coding tools ran out of their entire 2026 AI budget by April. Four months.

Andrew Macdonald, Uber's COO, said the quiet part out loud in May. The company cannot draw a line from what it spends on AI coding tools to what customers actually get. "It's very hard to draw a line between one of those stats and okay, now we're actually producing 25% more useful consumer features," he told reporters. Uber runs 30,000 engineers and one of the best data operations in tech. Its own COO still couldn't tell you what the spending bought.

That should stop you. Not because Uber is careless. Because Uber is good at this, and it still couldn't see it.

Here's the number sitting one layer up from Uber's. WRITER's second annual AI adoption survey, published April 7, 2026, drew on 2,400 responses from executives and employees across the US, UK, and Europe. It found 97% of executives deployed AI agents in the past year. Only 23% report significant ROI from those agents. Almost every company bought the tool. Fewer than one in four can say, with a number behind it, that the tool paid for itself.

I've watched this exact gap before, from inside a Fortune 100 innovation budget. You don't run out of enthusiasm for a new technology. You run out of the instrumentation to know if the enthusiasm was earned.

Here's what's actually happening. This is a measurement problem dressed up as a productivity story. It shows up in finance first, because that's where unmeasured spending always shows up first. J.R. Storment, executive director of the FinOps Foundation, said it plainly at FinOps X in June. Mature cloud teams forecast their spending within 1% to 3% of where it lands. With AI, he said, "it's totally blown out." Per-developer token use rose roughly 18.6 times in nine months, per TechCrunch's investigation into the industry's cost scramble. One company ran up a $500 million bill on a single AI vendor after never setting a usage limit. One Faros AI case found a single engineer spent $40,000 on tokens in a month.

This is a story about invisibility more than waste. You cannot govern what you cannot see. Most companies deployed agents faster than they built the ability to see them.

This has happened before. The shape is always the same. Paul David studied the electrification of American factories. He found that companies owned electric motors for decades before productivity moved. Nobody had rebuilt the accounting, the layout, or the workflow around the new power source yet. The motor showed up as a cost on the balance sheet long before it showed up in output. Owning a technology and measuring what it does for you are two separate purchases, made at two different times. The gap between them is where budgets go to die.

Agents widen that gap. An agent doesn't clock in like a motor or an employee. It runs a task, calls a model, sometimes calls another agent, and the bill lands as a token count with no face attached to it. Jellyfish's engineering data found something telling here. Engineers with the highest token use were roughly twice as productive as their peers. They also spent 10 times more tokens to get there. Is that a good trade? Most finance departments can't answer that question yet, because the systems that would connect token spend to business outcome mostly don't exist. That's why the Linux Foundation is launching a Tokenomics Foundation in July 2026, backed by Oracle, Microsoft, IBM, JPMorganChase, and a dozen others, to build shared definitions for exactly this. A whole standards body, built because nobody can agree on how to count what an agent costs, let alone what it's worth.

Zoom out and this stops being a CFO's problem. Every agent a company can't measure is a job getting rewritten in the background, with no number attached to the decision. The 52-year-old analyst whose queue an agent now handles. The new hire who never gets made because an agent covers the workload. The team lead asked to "own AI outcomes" with no dashboard to see them. All of that runs through the same invisible pipe Uber's COO described. Skills expire fastest where nobody can see them expiring. That's the relevance cliff. It doesn't announce itself with a memo. It shows up as a budget line nobody can explain.

Pilots get funded on faith. Relearning organizations get funded on receipts.

I don't think the answer is to slow down deployment. Uber didn't stop using Claude Code. It capped spending and kept building, which is the right instinct. The better answer is to build the human-plus-agents stack with the meter running from day one. Which agent, doing which task, replacing or extending which person's work, at what cost, measured against what that outcome used to cost the old way. Most companies bought the agents first and are building the meter now. That's backward, and it's exactly what everyone does with new technology the first time through.

So take this into your next budget review, before the next invoice lands. If I asked you right now what your AI agents cost last month, and what they were worth, could you answer both halves, or only one? Uber's COO couldn't, in public, at one of the best-resourced companies in tech. You're not behind if you can't either. You're early. The tokenomics standards, the FinOps-for-AI vendors, and the sharper board questions are all arriving in the same twelve months. This is the year to close the gap, before someone else closes it for you. I'm easy to find.

Sources: Fortune, "Uber burned through its entire 2026 AI budget in four months. Now its COO is questioning whether it's worth it," May 26, 2026 · WRITER/Workplace Intelligence, second annual AI adoption survey, published April 7, 2026 (2,400 respondents) · TechCrunch, "The token bill comes due: Inside the industry scramble to manage AI's runaway costs," June 5, 2026 · FinOps Foundation, FinOps X 2026 Day 1 Keynote, June 9, 2026 · Paul David, "The Dynamo and the Computer," American Economic Review, 1990.