How to tell AI spend from AI return

Gartner’s May 2026 forecast put worldwide AI spending at $2.59 trillion this year, up 47%, with infrastructure more than 45% of the pile. John-David Lovelock’s line is the one that belongs in a mid-market packet: so far the spend is driven by vendors and hyperscalers. Enterprises “have yet to really flex their spending potential,” and CIOs still struggle to show a business outcome.

McKinsey’s 2026 State of AI survey still had only 37% of respondents attributing any EBIT to AI — roughly unchanged — and about 6% as high performers (significant impact and at least 5% of EBIT). Eighty percent said AI improved their own productivity. Those two facts can both be true. A faster draft is not a cheaper close. This page is how you stop mixing them. It is not a promise that a tool will pay for itself.

Seat cost versus token cost

Two invoices get lumped as “AI.” They do not behave the same way.

McKinsey found about 20% of respondents already saying AI operating costs, including tokens, constrained use — while most still planned to spend more. If the only metric is seats, the organization will buy seats. If the only metric is tokens, it will run them without a finish line.

Add a third cost that never appears on the vendor invoice: human review after the draft, and rework when the draft invents a date, a customer, or a covenant. Time saved at the desk does not show up as margin if the review step and the headcount plan do not change.

Pilot versus production

A pilot has a demo. Production has a definition of done, an owner, and a number that can move. Gartner said in September that only 22% of organizations had successfully scaled AI across multiple business units. You do not need to be in the 22%. You need one artifact you can date.

Good production tests: monthly close commentary, support macros, bid/no-bid memos, invoice exceptions. Bad: “make the company more AI-native.” If you cannot name the artifact that is done, you are still in the pilot, whatever the slide says.

Write the current path: who starts it, how long it takes, where it fails, what “done” looks like. That is the baseline. No baseline, no after. Ninety days, one owner whose bonus includes cycle time or error rate. If nothing moved, cut the overlapping copilots and keep the model on the one path that has an owner.

A budget that grows 47% while the EBIT share stays flat is not a strategy. It is a bet that the spreadsheet will catch up.

Questions a board can ask

Directors do not need a model architecture briefing. They need a control framework that survives the next vendor cycle.

1. Which single workflow is in production, and what is the definition of done?
2. What did that workflow cost last quarter in seats, tokens, and review time — separately?
3. What number moved: cycle time, error rate, or headcount plan? “People feel faster” is not a number.
4. Who can turn the spend off, and who owns the data the model can see?
5. What happens if the vendor changes the price or the model next quarter — is there an exit, or only a renewal?
6. Are we in the 37% who can attribute any EBIT, or are we buying seats because the industry forecast is $2.59 trillion?

None of those questions imply a guaranteed return. They imply a company that can tell spend from a result. If the packet cannot answer them, the next copilot is a catalog item, not an operating change.

Sources

• Gartner AI spending forecast, May 19, 2026
• Gartner scaling survey, Sept. 1, 2026
• McKinsey, The State of AI: Global Survey 2026