Your software bill used to be predictable. You added a person, you added a seat, and the number moved in a straight line. Large language model tools do not work that way. They bill by what you consume, so the same task can cost very little or a great deal more depending on which model you point it at and how your team uses it.
Most of the companies I work with do not have a CFO. They have a controller and a founder. So when AI spend starts drifting, the controller is the one who has to catch it, and that means building the coding structure, setting the budget, and running the review with the founder directly.
You do not need a new tool for this. The controls already live in your accounting system. Code every software and AI dollar to a vendor and a department, put a gate in front of new spend, and review budget against actual every month, by department and by vendor. The same discipline catches the dormant SaaS seats nobody cancelled after someone left, which is the slower-moving version of the same problem.
One more thing worth raising with whoever prepares your return before year end. How these costs get classified can change when you take the deduction, and if you are building AI features into your own product it can affect whether you qualify for the research and development credit.
That's the short version. On the Fusion CPA blog I laid out the whole playbook: why token spend is so hard to budget, which businesses feel it first, and the four steps to get the monitoring in place. Read the full article on the Fusion CPA blog