KorvaJoin the waitlist
Money

Only 11% Of Companies Get Their AI Budgets Right

A cheaper model can use far more computing and still fail, and the company pays anyway.

AI image
WSJ What’s News · Why Businesses Can’t Budget for AI

Only 11% of companies budgeted for artificial intelligence properly in a recent study cited by Wall Street Journal visual editor Stephanie Stam, below the prior year's share.

Agentic AI is software that works through a task in steps, choosing and checking as it goes. Each step uses tokens, which providers bill for.

In a spring study from Stanford, Carnegie Mellon, UC Berkeley, and Microsoft, cheaper models used far more tokens in 32% of cases.

In one YouTube frame analysis, the simpler Gemini 3 Flash took 952 steps and 10 times the tokens, then failed for $14. The stronger Gemini 3.1 Pro finished in 85 steps for $1. The business pays for Flash's $14 failure.

Searches for token maxing, the push for heavy AI use at work, peaked in June. With models changing every few weeks, precise forecasts quickly go stale.

Questions people ask

Why is AI harder to budget for than regular software?

Most software charges a set price per user. With AI, the charge depends on use, and use depends on which model handles a task and how many tries it needs. When models change every few weeks, past spending is a weak guide to future bills.

How can a cheaper model end up costing more?

A lower price per token can be wiped out by heavy token use. A simpler model may take hundreds of extra steps, retry, or consult other tools before failing. All of those steps are billed, so the total can exceed the cost of a stronger model that solves the task quickly.

What are businesses doing to control the cost?

Experts point to training staff to match the model to the task, saving strong models for hard work. Another step is setting daily token limits per employee.

More in Money