A new survey from Harness, a software delivery company, asked 700 finance and engineering leaders a simple question: how much of your AI spending actually pays off. The answer was that about one in every four dollars produces nothing measurable in return.
For a company spending 1 million dollars a month on AI, that works out to roughly 260,000 dollars a month with no clear benefit, according to the survey. That is real money leaking out of the business every month, and most companies cannot explain where it goes.
The core problem is that AI costs behave nothing like normal software costs. A company email subscription costs the same every month. AI tools charge based on how much people actually use them: how many questions get asked, how much text gets generated, how many small pieces of text, called tokens, get processed. Usage swings wildly month to month, and so do the bills.
Most companies also buy AI from three or more providers at once, each with its own pricing rules. One tool might charge per question, another per document, another per seat. Comparing them takes real effort, and most finance teams have not built the tools to do it. More than half of the companies surveyed said they are forecasting AI costs through guesswork rather than data, and more than 40 percent are still tracking spend by hand in spreadsheets.
Making things worse, some of the biggest AI costs hide in plain sight. AI coding assistants and writing copilots often get billed the same way as ordinary software licenses. They do not look like AI spend on a budget report, so nobody flags them for review, even though they can be one of the largest line items in the company.
The industry is starting to respond. Oracle and Amazon Web Services have both added new billing features designed to make AI usage easier to see and track. The Linux Foundation, the nonprofit that oversees major open-source software projects, launched a new group this year focused entirely on AI cost management, working alongside its existing group that handles general cloud spending. Even OpenAI, which earns more every time a business spends more on its models, published its own guidance this year telling business customers to build better tracking and governance before scaling up.
That detail matters. When the company selling you the product is telling you to watch your spending, it is worth taking seriously.
None of this means AI is a bad investment. It means most companies adopted it before building the basic financial habits that any other major purchase would require: someone in charge of the budget, a way to compare vendors, and a system to catch waste before it becomes routine. The businesses that fix this now will spend the next year competing on cost efficiency while others are still trying to figure out where their money went.