Every control finance built for software assumes an annual contract and a fixed seat count. AI has neither. Here is what to measure instead — and why spend alone will mislead you.
Software procurement was designed around a predictable object: a contract, a seat count, a renewal date once a year. You negotiated at the start, you reviewed at the end, and in between the number did not move.
AI tools break every part of that. Spend moves daily, driven by consumption rather than headcount. A team can triple its usage in a week without buying anything or telling anyone. And most of the controls finance relies on — renewal calendars, seat audits, annual reviews — simply never fire.
With seat-based software, the invoice is roughly the same every period, so nobody needs to watch it. With consumption billing the invoice is a report on a decision that was made three weeks ago and cannot now be changed.
By the time an overrun appears in accounts payable, the credits are spent. The only useful moment to know is while the month is still running.
Four things, in this order.
This is the mistake worth avoiding. A high number and a low number look identical without context, and they call for opposite responses.
A team spending heavily with high consumption is getting value — and cutting them is the wrong move, however alarming the figure looks in isolation. A team spending moderately with almost no usage is pure waste, and it will never draw attention on a spend report because the number is unremarkable.
Read the two together and the decision is usually obvious in seconds. Read spend alone and you will cut the wrong team.
A meaningful share of AI tooling is bought on personal and departmental cards, outside any procurement process, often by people acting entirely in good faith because the official route is slow.
Two things to say about that. First, it is a governance gap, not a discipline problem — people route around processes that are slower than the work. Second, the answer is not to monitor employees. It is to make the visible route fast enough to be worth using, and to read the finance data you already have, where every one of those purchases eventually appears.
If you are doing this manually, start narrow: list every AI tool you pay for, get each one’s consumption and its committed capacity, set a monthly expected spend per team, and check it mid-month rather than at invoice time.
Even that crude version beats what most companies have today, which is a number that arrives after the fact and an argument about who is responsible for it.