Uber's CTO said it plainly in April.
"I'm back to the drawing board. Because the budget I thought I would need is blown away already."
By March, Claude Code had spread to 84% of Uber's engineering org. Individual engineers were spending between $500 and $2,000 a month on tokens alone. The entire planned 2026 AI coding budget — gone before the second quarter started.
The same week, Microsoft began quietly cancelling Claude Code licenses inside the division that builds Windows, Office, Teams, and Surface. Engineers were told to migrate to GitHub Copilot CLI by June 30 — the last day of Microsoft's fiscal year. The official reason was "toolchain unification." The real reason was in the calendar.
When the company with the most negotiating leverage in the room walks away from a tool its own staff prefers, the signal isn't about strategy. It's about the bill.
This isn't a CFO story. It's a product story.
The coverage has framed this as a cost crisis. Gartner predicting 25% of planned AI spend will slip into 2027. Fortune reporting that AI tooling, at heavy usage, can cost more per task than the human it was supposed to replace. An MIT analysis suggesting AI automation only pencils out cheaper than human labour for roughly a quarter of the jobs people assumed it would.
Those numbers matter. But they're pointing at a symptom, not the cause.
The reason enterprise AI budgets are blowing up isn't that AI got expensive. It's that AI products were designed without cost as a constraint.
The PM optimized for capability. For impressiveness. For the demo moment where the AI does something genuinely remarkable and everyone in the room leans forward. Nobody asked: what does this cost per workflow? What happens when every engineer runs this for eight hours a day? What does the meter look like at the end of a quarter?
The result is what Uber got. What Microsoft got. What every company that handed out AI coding seats without a usage model got: an invoice that looked nothing like the forecast.
Because the forecast was modelled on seat licences. The billing is denominated in how much the model has to think. Agentic AI makes models think a lot. Sessions run for hours, spawn parallel threads, generate context volumes that bear no resemblance to the autocomplete interactions that shaped the original pricing structure.
The budget wasn't wrong. The product was wrong. Nobody built for the economics.
I've already been in this conversation
A few weeks ago, I was sitting with one of our largest healthcare clients. A big system out of Illinois. We've been building a referral management platform for them — the kind of workflow where delays have real consequences: patients waiting longer than they should, referrals falling through gaps between systems, coordinators manually chasing things that should never need to be chased.
We were walking through the phased rollout together. Phase 1, Phase 2, Phase 3 — cost projections at each stage.
Somewhere in Phase 2, they stopped me.
"This is more than we expected."
Not aggressive. Not a negotiation tactic. Just honest. The number on the screen was larger than the mental model they'd walked in with, and that gap was sitting in the room with us.
I've been in enough of these conversations to know what's behind that sentence. It's not that they don't see the value. A referral management workflow that actually works saves them far more than what we were quoting. It's that nobody had given them a framework for what AI automation costs across a real multi-phase rollout. So the first concrete number feels like a shock — even when it's fair.
That conversation happened before I read the Uber memo. After I read it, I understood it differently.
The question in a sales conversation used to be: does this work? Can the AI do the thing? Is the demo impressive?
That question hasn't gone away. But there's a second question sitting right behind it now — and it's the one that's killing deals in procurement: what does this actually cost us at scale?
Not the licence fee. The real cost. What does usage look like when 50 people run this daily across three departments? Is there a cap? Is there a predictable number I can put in front of a CFO who just watched Uber's CTO admit on the record that he had no idea what he was spending?
If you can't answer that question concretely, you lose deals you should be winning. Not because your product is worse. Because your buyer is scared — and they have receipts from companies that moved fast and got burned.
Three things that change on your roadmap
Cost-per-workflow becomes a product metric, not just an infrastructure concern. Every feature decision you make is also a cost decision. If you're building an agentic workflow that loops ten times when it could loop twice, that's not an engineering problem — it's a product problem. You designed it that way.
Containment beats capability. The enterprises pulling back from broad AI tools are not pulling back from AI. They're pulling back from open-ended, metered, unpredictable AI. What procurement will approve now is a specific tool, for a specific workflow, with a predictable cost-per-outcome they can model. That's a different product than "give everyone access to the most powerful AI available."
The ROI conversation moves earlier. You used to prove ROI after the pilot. Now you need a credible cost model before the pilot starts. Not because buyers don't trust you — because their finance teams have been burned enough times that they're asking the question before anyone signs anything.
The companies that win the next phase of enterprise AI aren't the ones with the most impressive demos. They're the ones with the most defensible cost-per-outcome story.
What I actually think this means
There's a version of this story that reads as bad news for AI startups. Budgets tightening. Enterprise procurement slowing. The experimental phase ending.
I read it differently.
The experimental phase was always going to end. What replaces it is something more useful: a buyer who finally knows what they want.
The CTO who pulled Claude Code licenses because the bill was unpredictable is now, for the first time, genuinely interested in a product that solves one specific workflow at a known cost per run. He wasn't interested in that conversation six months ago. He was busy distributing seats.
The Uber budget story and the Microsoft licence cancellation are painful for the companies involved. For anyone building focused AI agents that do a defined job at a predictable cost, they are the clearest market signal in a year.
The budget reckoning isn't the end of enterprise AI. It's the end of enterprise AI that nobody bothered to make economically legible.
That's not a crisis. That's an opening.
