Last week Fieldwork's founder joined the ACG Utah CFO Roundtable in Salt Lake City, on a panel with CFOs from consumer products, construction, and a contractor roll-up, in front of a room of finance leaders, bankers, auditors, and investors. The agenda had seven topics. The Q&A had one: AI. The same four questions came from nearly every seat, and the answers below are the ones that survived cross-examination.

How should a finance team govern AI spend?

Treat it as a new spend category that behaves like a utility bill but gets signed like software. Track token spend by department, team, individual, and model type, and give every department its own AI line that rolls up to the board.

The model-type cut is the one most teams skip. A frontier model can cost ten times the workhorse model, so a new top-tier release should not mean org-wide access until someone brings a use case that justifies the price. Visibility by model is what lets a CFO make that call with data instead of vibes.

Resist hard spending caps for now. A cap is a blunt instrument that taxes power users exactly when the company wants more of them. Govern with visibility, model tiering, and manager conversations. And name the unsolved problem honestly: spend data shows who is using AI, not how well. Until usage quality is measurable, the most leveraged employee and the least efficient one look identical on the dashboard.

How do you measure AI ROI beyond productivity claims?

Hours saved is not ROI. Every vendor pitches time savings; unless those hours become headcount avoided, revenue, or cash, that is a feature, not a return.

The traditional framework is intact. Every investment has to drive one of three outcomes: more revenue, less cost, or better productivity and quality of work. The first two are easy to measure and easy to approve. Almost all AI spend lives in the third, which is exactly why it needs the most discipline, not the least. Pick the proxy metric before the pilot, not after, and walk away when it does not move. When productivity is measurable, measure it: one software team saw output per engineer roughly triple after operationalizing AI across the function. Same framework; it just took work to find the right output metric.

Some of this is tuition while the measurement catches up. Fund it deliberately, keep the board aligned, and prune hard once the data shows up.

Build, buy, or partner?

The default rule: if someone has built and is selling the exact solution, buy it. Specialization scale is real, and when a company builds, it owns the pager forever.

Two conditions apply. The product has to work, and the math has to work. The math is where the market is struggling. AI vendors price at the value they believe they create, not at their cost, so five figures to automate a few hours a week for one mid-level employee is now a common quote. That is a subscription to a demo. When that happens, pilot until the vendor proves the math, or build a scrappy internal version to keep the vendor honest. Either way, keep contracts short. Both the products and the prices will look different in twelve months.

How does the organization change?

Titles change before headcount does. When agentic tools got real this year, employees across functions started asking whether they were now supposed to be building their own automations. The right answer was no: keep doing the job you were hired for, and centralize the automation work in a small team modeled on business operations, with a backlog, ROI vetting, and a context layer over company data built before any agent ships.

Pair that with a role-by-role map of what AI use is expected and which tools are approved, so nobody is guessing at scope. The board scoreboard for all of it is revenue per employee. That is the number that tells you whether the AI story is real or theater.

The room did not want a vision of AI. It wanted a budget line, a metric, and an org chart. That is a finance problem, and that is the good news. For the measurement side of that problem, see the note on AI-era benchmarks.