CFO Daily · Tuesday, September 15, 2026

Hi, this is Koko from Koko Knows.

Let's get right to it, because there's a number out today that puts hard math behind something you've all been feeling for weeks. Accenture surveyed seven hundred fifty executives and found that enterprises spent roughly two and a half billion dollars on AI tokens last year, and eighty percent of that spend, four out of every five dollars, can't be tied to a business outcome. Not "underperforming." Not "hard to measure." Literally no quantified link. That's the sharpest signal I've seen all week, and it changes the conversation from "is AI adoption flat" to "why can nobody tell me what we're actually buying."

And here's the second piece that explains the first. Deloitte and Teradata data out today shows seventy-eight percent of enterprises are running at least one agent pilot, but only fourteen percent have scaled one org-wide. That's an eighty-nine percent pilot-to-production failure rate. Put those two numbers together and the diagnosis writes itself: pilots keep multiplying because nothing forces a dollar-to-outcome link before anyone tries to scale them. It's not a model quality problem. It's a governance and measurement problem, and it's sitting squarely in your lane.

So here's what I'd take into Q4 budget reviews. Stop treating agent funding requests as pilot demos and start treating them as scaling business cases. Require the unit economics up front: what does this cost per task, what outcome does it produce, and what's the kill criteria if it doesn't scale. IBM is making a related point today that I think is right — token spend needs the same FinOps discipline you'd apply to cloud spend a decade ago. Budgets per workflow, per-task monitoring, model-sizing rules that match compute to task complexity, not just throwing the biggest model at everything. Pick one workflow this quarter and pilot that discipline before agentic tools multiply across your close process and FP&A stack, because once they're embedded in the close, retrofitting governance gets a lot harder.

There's a third thread worth flagging for your capital allocation committee, and it's more structural. BCG is out with a number that should get attention in any infrastructure-heavy AI conversation: six trillion dollars in planned US buildout capex through the end of the decade, for data centers, AI compute, and grid capacity, running into a projected shortfall of two million skilled tradespeople by 2030. That's not a funding problem, it's an execution problem. If your firm has capital committed to AI infrastructure, this is the moment to stress-test your ROI timeline against build delays, not just against whether the capital is available. Availability of dollars was never the constraint. Availability of electricians and pipefitters might be.

And layered on top of all that, you probably saw AI stocks slide after several major CEOs jointly called for a slowdown in model development. I wouldn't call that a reason to reverse any capex commitments, but it is a volatility signal worth putting in front of your board alongside the capex conviction story from earlier this week. Consistency of message matters here, and right now the market's getting mixed signals from the very people driving the buildout.

If I had to leave you with one governance move, it's this: put your breach rate and your pilot-to-scale rate on the same dashboard. There's data circulating that shows almost ninety percent of organizations suffered a GenAI-related breach this year, up sharply from last year, even as confidence in preventing unauthorized access went up over the same period. That gap between rising confidence and rising incidents is the same pattern as the gap between pilot enthusiasm and production failure. Deployment speed is outrunning verification at every layer, and that's a board-level governance gap, not a technical one.

Practically, for your next FP&A cycle, I'd do three things. First, make token spend a budgeted, outcome-tagged line item instead of an open-ended cost bucket buried in cloud infrastructure. Second, put a scale-or-kill gate in front of every agent pilot cohort, with unit economics attached, before it gets a production budget line. Third, ask your architecture and integration teams whether your ERP core stays stable while these new agent layers sit on top of it, because a clean core with a composable edge is turning into the safe pattern enterprises are converging on, and it's a lot easier to govern and audit than agents wired directly into fragile legacy systems.

That's the rundown. Four in five AI dollars still unaccounted for, nine in ten pilots stuck in place, and a capex story that's more fragile on execution than it looks on the spreadsheet. Require the outcome link before you fund the next tranche, and you'll be ahead of most of your peers.

That's it for today. Thanks for listening, this has been Koko Knows.