CFO Daily · Sunday, October 4, 2026

Koko: New number on the AI-ROI fight: KPMG surveyed over a thousand finance leaders and found seventy five percent now use AI regularly, up from under a third in twenty twenty four. And seventy one percent say it's at least meeting ROI expectations.

Sam: Which sounds great until you remember McKinsey's thirty seven percent EBIT-attribution number from this same window. Two credible surveys, same question, very different answers.

Koko: Because they're not asking the same question. KPMG's bar is 'meeting expectations.' McKinsey's bar is 'can you point to EBIT.' Those are miles apart.

Sam: And KPMG also found AI performs best on judgment-heavy work, forecasting, decision support. Not the rote automation most pilots started with.

Koko: Meanwhile Supabase just raised a hundred fifty million and bought Turso because agents now generate seventy percent of its new databases. Infrastructure cost that nobody's line-itemed.

Sam: So here's the sentence for the CEO: there are now two legitimate answers to 'is AI paying off,' and which one your team quotes is itself a decision finance needs to own.

Koko: Let's actually sit in why this gap matters operationally. If a CFO walks into a board meeting and says seventy one percent ROI, and a director read the McKinsey piece and says thirty seven percent EBIT, that's an awkward five minutes.

Sam: It's worse than awkward, it's a credibility problem. You look like you cherry-picked the flattering survey.

Koko: Right, so the fix isn't picking a side. It's defining your own metric before anyone asks. What does 'ROI' mean at your company, tied to which line.

Sam: KPMG's own data actually points to the fix. Leaders using agentic AI with strong governance report outcomes three to six times better than peers. That's not a measurement artifact, that's a real operating difference.

Koko: So governance isn't the brake on ROI, it's the thing that produces it.

Sam: That's the part I'd push back on slightly. Governance correlates with better outcomes, sure. But it could just as easily be that the companies disciplined enough to govern AI well are also the companies disciplined enough to deploy it in the right use cases to begin with.

Koko: Fair, correlation versus causation. But either way, the instruction to FP&A is the same: don't scale a pilot until you can show the EBIT link, full stop.

Sam: And flag which survey definition you're using every time you cite a number upward. CFOs are going to get asked which one by the audit committee before year end.

Koko: Second thread: Supabase's raise is a tell. Four million new databases a month, AI tooling driving most of it, and they needed fresh capital just to handle the load.

Sam: That's a cost center finance hasn't named yet. Compute is in the AI vendor contract. The data layer underneath each agent, usually isn't.

Koko: So when procurement renews an agent contract, the question is literally: what percentage of this spend is compute versus data-layer overhead.

Sam: And layer on the CNBC piece — forty two billion dollars of European data-center investment already delayed or cancelled because of community pushback. Denmark, Scotland, Seoul, all pushing back on siting.

Koko: Which means the multi-year compute pricing in your AI vendor contract is sitting on assumptions about capacity that might not hold.

Sam: Right, that's a treasury and procurement problem as much as an FP&A one. If you're locking a three-year AI infrastructure deal right now, stress-test the vendor's delivery assumptions against actual sited capacity, not their roadmap slide.

Koko: Where I'd actually push back on you: is this a new risk, or just the old vendor-concentration risk wearing a new jacket?

Sam: It's old risk, new scale. Power and water constraints didn't used to gate a software renewal. Now they do.

Koko: Quick counterpoint to all the budget anxiety: Chatham Financial's still the clean proof point, trade validation down from thirty minutes to under four, fully auditable.

Sam: And EY cutting audit deficiencies from twenty eight percent to under five percent with AI-assisted review, per The Accounting Podcast. Accountant pay rising for AI-exposed roles too.

Koko: So it does work. It just works where someone engineered the governance in first, not bolted it on after.

Koko: So here's the call. Controller: before your next AI spend update goes to the board, write down your own definition of ROI, tied to a specific P&L line, and use it consistently. Stop borrowing whichever survey number is flattering that week.

Sam: FP&A: get the data-layer and compute cost split out of every agent vendor line item before the next renewal. If the vendor can't break it out, that's your answer.

Koko: And treasury, procurement, anyone signing multi-year AI infrastructure deals: pressure-test the delivery promise against actual sited capacity, not the roadmap. Forty two billion in Europe alone is already stalled.

Sam: Audit committee's question tomorrow is simple: which number are we quoting, and can we defend the definition behind it.

Koko: Full write-up's at koko knows dot A I. One thing worth watching: whether 2027 budget cycles shift AI dollars from more tools toward actual governance infrastructure.

Sam: That's the real tell. See you tomorrow.