At some point in the last year, your AI spending stopped being one line item and became a whole page.
Maybe you have noticed it. The chatbot subscription that was a "one-time experiment" in January. The four seats on a tool where two people actually log in. The usage-based invoice that varies wildly month to month and nobody on the team can explain. The credit card statement where AI vendors are now the second-largest recurring category after software you actually remember buying.
None of this had a decision behind it. It accumulated.
That is why more businesses are putting a leash on their AI spend — not because AI is a fad, and not because it isn't working. Because spending that has no owner, no review date, and no attached job is not a strategy. It is a leak with a dashboard.
How the spend got loose
AI did not arrive through the purchasing process. It arrived through the browser.
Software used to come with a procurement path: a ticket, a quote, a sign-off, a rollout. AI came as a free trial that quietly started billing a card on file. The tool is good, the team uses it, and nobody ever makes the "keep or cut" decision out loud — the trial just graduates into a permanent line item by default.
Small businesses feel this more than anyone, because they have no procurement department to say no. There is usually one person who signs for everything, and that person is busy running the business. So the questions that enterprise finance would ask — who uses this, what does it replace, what would we do without it — never get asked at all.
Add the billing model and it gets worse. AI vendors charge three ways at once: per seat, per usage, per token. A tool can look cheap at $20 a seat and still produce a four-figure invoice from a busy month. You cannot predict the number, so you stop looking at it.
The forecast that saw this coming
This was not hard to see, and the people whose job is watching saw it early. In July 2024, Gartner predicted that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025. The reasons it named: poor data quality, inadequate risk controls, escalating costs, and unclear business value. Its analysts put it plainly — executives were impatient for returns while organizations struggled to prove and realize value.
That forecast was about big enterprises and big projects. The mechanism it describes is the same one that shows up in a small business as a confusing page of AI line items. Escalating costs. Unclear business value. Projects that clear the "it works" hurdle and then die on the "what is this worth" question.
The value is real. It is also uneven.
The honest version of this story is not "AI is a waste of money." In the same July 2024 research, Gartner reported survey respondents citing a 15.8 percent revenue increase, 15.2 percent cost savings, and 22.6 percent productivity improvement from their generative AI work. The numbers exist. The returns are real.
But Gartner's own analysts cautioned that the benefits are specific to company, use case, role, and workforce. What that means in practice: in the same company, one AI tool can be paying for itself ten times over while another quietly burns money. The average looks fine. The line items tell a different story.
That is the whole case for the leash. Not "cut AI." Not "AI doesn't work." The case is that some of your AI spend is earning its keep and some of it isn't, and you currently cannot tell which is which.
What the leash actually is
The leash is not a budget cut. It is visibility plus a decision rule. Five steps, none of them heroic:
1. Inventory the line items. Go through the credit card statement and the accounting software, line by line, for the last six months. List every AI subscription, every usage-based invoice, every seat count. Include the tools nobody remembers buying — especially those. This is an inventory, not a witch hunt. You cannot manage spend you cannot name.
2. Attach each spend to the job it does. For every tool, write down the actual work it does in your business: "drafts the first version of client emails," "transcribes meeting notes," "generates product images." If you cannot write down the job in one sentence, that is the answer. A subscription with no describable job is a donation.
3. Check usage, not seats. Log in and look. The pattern is almost always the same: you are paying for twenty seats and six people use the tool. Seat-based billing is the easiest money an AI vendor will ever make from you, because the invoice looks small and the unused seats are invisible. Cancel the seats. It takes ten minutes and it is the fastest saving in this whole exercise.
4. Set a review date — quarterly, not annual. AI changes faster than annual reviews can see. A tool that made sense in January can be obsolete by April, and a tool that looked useless in January can become essential by July. Pick a quarter-end, spend an hour going through the inventory, and make one keep/cut/downgrade decision per tool. The point is not perfection. The point is that the decision happens at all.
5. Write down the rule. One page: who is allowed to buy AI tools, what needs a second look before anyone signs up, and when the review happens. It does not need to be a policy document. It needs to be the difference between "we decide this" and "the trials decide for us."
What it costs you not to
Skip the leash and the costs are quiet, which is why they are easy to ignore. Renewal inertia — the tool that keeps billing because nobody cancelled it. The reconciliation surprise — the month where the usage invoice doubles and nobody can say why. The seat bloat — paying for people who do not use the tool.
But the deepest cost is not the money. It is the map. If you never connect spend to value, you can neither defend the good spend nor cut the bad spend intelligently. You will either keep paying for everything out of inertia or cut everything out of frustration — and both of those are decisions made in the dark.
There is also the fear underneath this whole topic, and it deserves naming: if I put a leash on AI, will I fall behind the competitors who are going all-in? The evidence says the opposite. Most of the abandoned projects Gartner described were not abandoned because the companies were too careful. They were abandoned because the value never got tied to a real job. The companies that fall behind are not the ones that ask what AI is worth. They are the ones that pay for a year and never ask at all.
Start by seeing
Every business that spends well on AI — and some of them do — can answer one question about every tool: what did this do for us, and what did it cost? Not in a fancy dashboard. In a sentence.
The leash is not a muzzle. It is a way of making sure the AI spend you have is the AI spend you want. Start with the credit card statement, the seat counts, and the tools nobody remembers buying. Look at the page. Then decide what actually earns its place on it.