Your AI Now Needs Its Own Accountant

Gartner just told its clients to stand up a new function. Not for building AI. For counting it.

The report is called "Driving AI ROI," and the argument underneath the framework diagrams is blunt: traditional IT financial management was built for deterministic systems, software that behaves predictably, and AI is not that software. Costs vary per query. Performance is probabilistic. A model can pass every technical benchmark and still bleed money. Gartner's prescription is an AI-specific financial management practice, a cross-functional center of excellence pulling in enterprise architecture, finance, and AI experts, whose job includes tracking whether the AI is worth what it costs, from the first pitch deck to the day someone finally turns it off.

We have been making a version of this argument in this column for a while. Your AI Subscription Is Becoming a Utility Bill was the first Budget Math. The Quiet Death of Per-Seat SaaS was the second. The through line was that AI broke the way companies buy software, and the bill would eventually force a reckoning. When one of the most influential IT research firms in the world tells its clients to stand up a financial management practice for it, the reckoning has a budget line.

The gap that explains the memo on your desk

Two numbers from Gartner's own surveys sit next to each other in the report, and they explain most of the AI behavior you are watching at work.

Ninety-one percent of corporate board members view AI as an opportunity to drive shareholder value. Seventy-five percent of the CIOs and technology leaders Gartner surveyed report that current AI implementation costs outweigh the realized benefits.

Read those again as a pair. One group sees the upside. The other is paying the invoice. Every confusing AI mandate, every transformation memo, every tool rollout with no success metric attached lives inside that gap. The board is funding the 91 percent story. The operators are living the 75 percent reality. Something has to close the distance, and Gartner's bet is that it will be financial discipline, not breakthroughs.

The fine print matters, and we will hold ourselves to it: both figures are sentiment surveys, fielded in spring 2025, with Gartner's own disclaimer that they reflect respondents, not the market. The 91 percent comes from 330 nonexecutive directors; the 75 percent from a 506-person sample of CIOs and other technology leaders. Treat the pair as a measured mood, not a law of physics. The mood is still the point.

Value drift, the term worth learning before Q4

The most useful thing in the report is a phrase: value drift. Gartner defines it as the state where a model remains technically accurate but becomes financially unviable, because contexts changed or compute costs rose. The AI still works. The math no longer does.

Variable software cost is not the new part. Anyone who has managed a cloud budget has watched a bill spike. But a cloud bill can spike while still holding a visible relationship to consumption: more usage, bigger invoice, and the argument stays about efficiency, not existence. AI can break that relationship at the transaction level. A model can spend more on retries and longer contexts while returning less, so cost rises as delivered value falls, and every dashboard still says the model is fine. That disconnect between what a request costs and what it returns is why the term earns its keep. Gartner's fix is unit economics: cost per successful outcome, tracked continuously, with reviews that retune or retire models when running them costs more than they return.

There is a second phrase worth filing next to it: shadow AI spend. The unsanctioned tools, the personal subscriptions, the API keys nobody tagged. Gartner lists it as a cost category to hunt down, which tells you the firm's clients already know it is there. And it carries a quiet consequence: if you cannot see that usage, your unit economics are not conservative. They are incomplete.

The prediction, handled with tongs

The report's headline forecast will travel fastest, so let's handle it carefully. Gartner projects that by 2028, 70 percent of AI initiatives will be decommissioned due to unmanaged cost explosions and value drift. The version you will see in vendor decks stops there. The sentence does not. The full claim applies to organizations lacking a dedicated AI financial management practice, which also makes it an argument for exactly the kind of engagement Gartner advises on. Worth knowing; not disqualifying. And to be precise about what it is: a Gartner strategic planning assumption, not a measured outcome. Nobody will be graded on that number until 2028, but the direction is hard to argue with. Projects with no owner, no unit economics, and no review cadence do not die loudly. They drift until someone in finance notices.

The one question to carry into your next AI meeting

You do not need a center of excellence to use any of this. You need one question.

What is this tool's cost per successful outcome?

The math is not exotic: everything the tool costs to run, the direct spend, the retries and failed attempts, the human time spent reviewing and correcting its output, divided by the number of workflows it verifiably completed.

Run it once and the purchase order stops being the price. A support copilot that costs 4,000 dollars a month is not a 4,000 dollar tool. Add the overage charges, the failed drafts, and the review time, and the all-in cost might reach 6,000 dollars. If the system reliably resolves 1,000 cases, the number to manage is 6 dollars per verified resolution, not the subscription line on the invoice. The figures are illustrative. The habit of counting that way is the point.

If nobody in the room can answer it, that is the finding. Not a crisis, a finding. It means the project is being managed on the 91 percent story instead of the 75 percent math, and the distance between those two numbers is where budgets go to disappear.

The accountants are arriving either way. Better to be the person who asked first.

We read 47 AI and tech newsletters so you don’t have to. getaiminute.com