What Nobody Budgeted For
Design Ops
AI Design
Nvidia's VP of Applied Deep Learning told Axios last week that compute costs for his team now exceed what his employees cost. Uber's engineering teams burned through their entire AI tooling budget on Claude Code before the year hit midpoint.
Somewhere right now, a finance team is building a slide with "AI ROI" on the cover and a very uncomfortable number underneath it.
The part that gets underestimated often: AI doesn't replace professionals. It replaces parts of what professionals do - and you still need someone who knows the difference. Someone to write the prompts, build the workflows, inject the context, audit the output, and catch the mistakes an AI doesn't know it made. That work isn't free. The people who assumed otherwise are finding out through their Q2 budget reviews.
For example, a tool like Claude Design can generate a component. But you need to write the prompt, point it to the design system, review what came back, fix what it misread, and iterate. That process can easily take a mid-level or senior designer three times longer than doing it by hand, and burn a significant chunk of the week's token budget on a single task.
And the costs aren't stabilizing. Enterprise AI software fees rose between 20 and 37% last year, according to spending management firm Tropic. The subsidized pricing that introduced most teams to these tools looks a lot like how cloud computing started - generous until you're locked in.
The answer isn't to pull back. It's to treat AI like any other budget line: match the tool to the task, reserve expensive tokens for problems that actually need them, build workflows that make the costs predictable.
The teams getting this right aren't the heaviest users. They're the most intentional ones.