The CFO of a PE-backed restaurant group should check every AI project against the last full fiscal year before the planned sale: list each project's cost and go-live month, keep in the budget those live before that year begins, and send the rest to the operating partner to decide whether that year carries its cost.
Researchers at the Atlanta and Richmond Federal Reserve Banks and Duke University surveyed senior financial executives, most of them CFOs, for an NBER working paper dated March 2026 and not peer reviewed. The CFO Survey's panel gave 603 responses and supplemental surveys 145, from November 11, 2025 to mid-January 2026; the median firm had 118 employees. The labor productivity gain executives attributed to AI in 2025 averaged 1.8 percent over the 699 who answered, while the revenue and headcount changes they attributed to AI imply 0.6 percent for 678 firms. The authors write the gap likely reflects delayed output realization and quality improvements that are not yet captured in measured revenues. The paper puts leisure and hospitality in low-skill services, which had smaller implied 2025 gains than high-skill services. In all four sectors, the 2025 reported gain approximately equals the 2026 implied one, which the authors say suggests a possible resolution to the productivity paradox: gains from 2025 investment in AI will increase sales revenue the following year, at least according to CFO expectations.
The paper finds much AI spending, particularly among smaller firms, takes the form of operating expenses rather than capitalized investment, and that large companies also have substantial operating expenditures, but also spend on hardware and internal development. Operating expenses come off EBITDA when booked. Where a buyer sets the group's enterprise value at a multiple of one fiscal year's EBITDA, a project going live inside that year puts its cost into that EBITDA and, if the CFOs' expected lag from investment also runs from go-live, its added sales into the next year.
Take a group whose sponsor plans a sale process in early 2028 on fiscal 2027 results, with the 2027 budget due at the December board meeting. In the two weeks before 2027 budget requests are due, the head of FP&A reads the contract and latest invoice of the 50 largest software and outside-service vendors by trailing 52-week spend, then every request in week three, when the requests are due, listing each AI project with its budget owner, operating cost by month in subscriptions, services and training, each marked one-time or recurring, and go-live month at the last location, actual or planned. Smaller vendors' current contracts and AI bundled into existing subscriptions stay in the 2027 cost base. At the end of week three the CFO sorts the list by go-live month. An AI phone-ordering line due to go live in November 2026 keeps its place in the 2027 budget. An AI catering-sales assistant planned for April 2027 goes in December, with its 2027 one-time cost, its 2027 recurring cost and any 2027 saving or added sales its owner names, to the operating partner, who either approves those costs or moves its start past the sale.
The group pays every AI cost from its own budget, and the list costs three weeks of the head of FP&A's time. An operating partner can ask any portfolio company with a planned sale year for the same list.
On savings, the paper finds little evidence that AI has had, or in the near-term will have, large effects on the total number of employees or on costs, with one moderate exception: large firms, those with 500 or more employees, 22 percent of the 722 of known size, expect to reduce headcount by 0.7% in 2026 due to using AI.
