A company should settle three things before buying an AI operating layer: which systems it touches, who owns it, and how finance measures payback.
In this post, an AI operating layer is software that sits between AI tools and the systems a company already runs. Those systems include accounting, sales records and payroll. The AI tools include agents, programs that carry out a task without a person approving each step. The AI operating layer sets which data each AI tool may read and what it may change. It also keeps a record of what each AI tool did.
Few firms using AI have changed how they manage and store data. A Census Bureau working paper from April 2026 drew on a Census survey of more than 117,000 US firms. The survey covered November 2025 to January 2026. Among firms using AI, about 7 to 8 percent had changed how they manage and store data in the past six months. The paper has not had the review that Census gives its own publications.
The same paper found that staff sometimes use AI for work tasks without their firm formally adopting it. The firm's survey respondent reported this staff use. The staff themselves were not asked.
A Federal Reserve note from April 2026 compared three public surveys of AI use. It says early workplace surveys found staff using AI "without the knowledge or approval of IT or management". The note adds that this unapproved use may still be "a substantive problem for organizations".
So the list of systems should name every system the AI operating layer may read or change. It should also name the AI tools staff already use, approved or not.
The owner should be one manager whose budget pays for the AI operating layer, such as the head of operations. That manager approves each new AI tool and each new system the AI tools may change. The IT team keeps the software running, and the owner answers for what the AI tools do.
A working paper from the National Bureau of Economic Research, dated March 2026, surveyed nearly 750 corporate executives. Most of them were chief financial officers. The main survey ran from November 11 to December 16, 2025 and drew 603 responses. The paper has not been peer reviewed.
One extra survey in the paper drew 183 valid responses about the type of AI spending. On average, firms with 500 or more employees in that survey expected 55 percent of their 2026 AI spending to go on running costs. The paper lists those costs as subscriptions, services and training. So finance should measure the AI operating layer each month, as a running cost.
On average, the executives said AI raised output per worker by 1.8 percent in 2025. Their own revenue and headcount estimates implied a smaller gain, 0.6 percent. The authors write that the gap is likely "reflecting a delay in revenue realizations", meaning sales that arrive later. So the payback measure should be a cost line that finance already reports each month, with revenue effects reported separately.
What to do first:
- Chief operating officer: by October 30, list each system the AI operating layer may read or change, and each AI tool staff already use.
- Chief executive: by October 30, name the one manager who owns the AI operating layer and pays for it from that manager's budget.
- Chief financial officer: from the November month-end close, report the monthly cost of the AI operating layer next to the cost line it should reduce.
The chief operating officer's list starts with tools staff already use. The Federal Reserve note also cites a survey of individuals. It found about 41 percent of the workforce using generative AI, the kind that drafts text, for work in November 2025.
