Uber & Microsoft are questioning the AI ROI
This is mostly a non-issue for nontechnical businesses
In April, the CTO of Uber said that they’d blown through their entire AI budget for calendar year 2026.
Two weeks ago, Microsoft cancelled their developers’ Claude Code subscriptions because they were getting too expensive.
Uber’s COO, this week, said that they are finding it “harder to justify” the costs AI coding tools and that they have started limiting Claude Code subscriptions.
And then the bill showed up.
In Q1, Anthropic moved its enterprise customers to token-based pricing (usage pricing) for all Claude Code usage. Since then, the true cost of the AI tools has become known, and companies can’t justify the spend. Enterprises are no longer being subsidized by VCs and investors.
AI Coding Tools Are Not Normal Automation
Claude Code is not the same thing as automating invoice reading, email routing, receipt matching, or customer follow-up.
Those are normal business automations.
They have clear inputs.
They have clear outputs.
They run the same way every time.
Software engineering is different.
Software engineering is part science, part craft, and part “who decided to build it this way in 2017 and why is that person no longer here?”
When an engineer uses Claude Code, they have to explain the whole problem. They have to explain the goal, the edge cases, the codebase, the dependencies, and all the weird little traps that live inside the system.
Then the AI agent has to read the code, understand the structure, make the change, test the change, and often revise the change.
All of that uses lots and lots of tokens.
Tokens are the meter, and the meter is running.
Why Claude Code Gets Expensive So Fast
Imagine you wrote a 1,000-page book.
Now you want to add a new chapter in the middle.
You do not just say, “Write chapter 37.”
You explain what the chapter is supposed to do. You explain the characters, the plot, the tone, the setup, and the payoff.
Then the AI has to read the book, write the chapter, and reread enough of the book to make sure the new chapter does not wreck everything that came before or after it.
That is what is happening in software.
Except the book is a codebase.
And the book is changing every day.
And 4,000 engineers are all asking the AI to help edit different chapters at the same time.
This is how you blow through an AI budget.
This Is Not a Problem for Most Companies
For non-software companies, this is mostly a non-issue.
Most small and midsized businesses do not need Claude Code. They need AI automation.
That path is much simpler:
Document the process in painful detail.
Have a developer figure out which parts can be automated.
Build the automation.
Measure the cost per task.
In one business, we automate reading emails. We know the cost per email.
In another, we automate reading invoice. We know the cost per line-item.
The only variable is volume. How many emails? How many receipts?
That is manageable.
A task automation is not wandering around your whole codebase trying to reason through five years of engineering decisions. It is doing a job.
Same input.
Same output.
Same cost model.
Task automation is the fastest and best path to transformation for all non-software companies right now and for the next several years.
This Is a Big Problem for Software Companies
Enterprise software companies have a harder choice.
They can cap Claude Code usage and slow the productivity gain.
They can keep Claude Code unlimited and reduce headcount.
Or they can move to cheaper AI coding tools that are not as good, and accept a smaller improvement.
None of those are fun choices.
And this is where the AI coding story gets uncomfortable.
If two strong developers with unlimited Claude Code can outperform ten average developers without it, the spreadsheet starts doing spreadsheet things.
That does not mean every tech layoff is because of AI.
A lot of the current layoff wave is probably still the result of overhiring from 2019 to 2022 combined with current share prices taking a beating.
But AI is now part of the math because these AI coding tools are costing companies 4 and 5-figures per month per engineer.
And once the Claude Code bill gets large enough, the question changes from “Should we give every developer AI?” to “How many developers do we need if the best ones have AI?”
That is the real issue.
The non-software world will not run on Claude Code.
The non-software world needs boring, specific, well-built automations that do the task as well as (or better than) a human does at a known cost.
Software companies have a very different problem.
They are not just automating tasks.
They are automating parts of the people who build the product.