CFO Daily · Saturday, October 3, 2026

Koko: The number that should be on every CFO's desk this morning: eighty nine percent of companies are using AI regularly, but McKinsey's new Technology Trends Outlook finds only thirty seven percent can point to any positive EBIT impact. Agentic deployments included.

Sam: So nearly two thirds are spending with nothing to show on the income statement. That's not a pilot problem anymore, that's a capital allocation problem.

Koko: Same window, Chatham Financial shows the other side. OpenAI's case study has them cutting trade validation from thirty minutes to under four, using Codex and GPT-5.6, fully governed and auditable.

Sam: And Anthropic's dropping a hundred million dollars to train ten thousand deployment engineers by 2027. Pulled from Deloitte, McKinsey, Morgan Stanley.

Koko: Which tells you what the real bottleneck is. Not model quality. Implementation talent. Here's the sentence for your CEO: AI adoption is universal, AI value is not, and the gap between them is exactly where finance needs to be setting the gates.

Koko: Let's sit in the McKinsey number properly. Thirty seven percent positive EBIT attribution, despite near-universal usage. CIO Magazine's framing of the same data is even sharper: sixty three percent see zero EBIT gain, and agentic deployments run five to thirty times the compute cost of a chatbot.

Sam: Five to thirty times. That's the line a controller needs to pull into the budget conversation before anyone signs off on scaling an agent pilot.

Koko: And ninety three percent of companies have blown their AI budgets already, per that same outlook.

Sam: So the ask from FP&A has to change. Not 'show me usage,' but 'name the line item this moves and show me the before and after.' Chatham's thirty minutes to four minutes is exactly that kind of number.

Koko: Right, and that's the test I'd apply to any vendor pitch right now. Does it have a Chatham-style benchmark, concrete and already measured? Or is it a roadmap slide dressed up as a business case?

Sam: Here's where I'd push back a little, though. Chatham's a single, narrow, well-bounded workflow: trade validation. McKinsey's talking about enterprise-wide agent rollouts. Those aren't the same bet.

Koko: Fair, but that's actually the lesson. Chatham didn't try to automate everything. They engineered around where judgment matters and automated the rest. That's a sequencing discipline, not a technology difference.

Sam: Then the control question for the audit committee is simple: has anyone actually scoped where judgment lives in our own agent pilots, or are we assuming the agent will figure that out?

Koko: Pluralsight's number ties this together. Nearly seven in ten tech leaders abandoned an AI-related project this year over skills gaps, and two thirds of those ate losses over two hundred fifty thousand dollars.

Sam: Two hundred fifty thousand dollars per abandoned project, and that's probably conservative since it's self-reported. Who's eating that write-off, the IT budget or the sponsoring business unit?

Koko: Usually whoever sponsored the pilot. Which is exactly why Anthropic's hundred million dollar Frontier Academy bet matters. They're training deployment engineers because they've concluded implementation talent, not model capability, is what's gating ROI.

Sam: So when a vendor quotes you a contract, the real question is whether their deployment engineer headcount is billable separately or baked into the license fee.

Koko: That's the negotiation most finance teams haven't had yet.

Sam: And Accenture's numbers back this up from the supply side. Managed-services bookings hit a Q4 record, book-to-bill over one point two, their AI specialist headcount beat a three-year target a year early.

Koko: Clients are buying execution, not advice. That's the read.

Sam: Which means next year's professional-services budget needs a new line. Separate the run-the-process AI contracts from classic consulting hours, because they're priced and governed completely differently.

Koko: Quick one: KPMG data, cited on The AI Daily Brief, shows the firms actually getting AI returns have shifted ambition from cost-cutting to revenue growth. That gap compounds.

Sam: And coding agents are the cautionary tale inside McKinsey's own numbers. Programming activity up one hundred eighty percent, releases barely moved, thirty percent of dev teams saw productivity fall.

Koko: Top developers captured most of the gain. The bottom eighty percent averaged about three percent. Worth remembering before anyone sells you an average.

Koko: So here's the call. FP&A: before you approve another agent pilot, require a named EBIT or cost line it's supposed to move. Usage metrics don't clear the bar anymore, not after today's numbers.

Sam: Controllers, benchmark any vendor's promised deployment timeline against Chatham's actual before-and-after, thirty minutes to under four. If they can't show something that concrete, you're buying a roadmap, not a result.

Koko: And whoever's negotiating the next AI vendor contract, ask directly whether deployment engineer support is billable or included. Anthropic just told you that talent is the real bottleneck, not the model.

Sam: Audit committee question worth asking this quarter: have we scoped, project by project, where human judgment actually sits inside our agent workflows? Because that's the thing nobody's mapped yet.

Koko: That's the show. Full write-up's at koko knows dot A I.

Sam: Question worth watching: does the thirty seven percent EBIT number move by next earnings season, or does the budget just keep climbing anyway?