CFO Daily · Tuesday, September 29, 2026

Hi, this is Koko from Koko Knows.

Let's get right into it, because the conversation you're having with your board this week just got more complicated. BCG dropped a cluster of research that reframes cost management as a capital allocation decision, and there's a fresh data point on AI vendor concentration risk that deserves your attention before your next contract renewal, not after.

Start with the headline: BCG's structural cost reset research looked at companies from 2016 through 2026 that treated cost architecture as a board-level call, not a finance-department exercise. The ones that did it right saw real expansion in their forward EV to EBIT multiples. Not just margin relief, actual multiple expansion. The takeaway for you is blunt: investors are no longer pricing your savings, they're pricing your execution capacity. So before your next board cycle, pressure-test whether what you're running is a genuine structural reset or just a run-rate exercise dressed up in a new deck. That distinction is now showing up in your valuation, not just your income statement.

Second thing worth flagging, and I want you to treat this one with real skepticism: BCG also published a piece claiming a single AI finance agent automated over ninety percent of BI work in nine months. That's a striking number, and it's part of a broader argument about finance shifting from reporting to value creation. But this is research from a firm that sells transformation consulting, so hold that number the way you'd hold any vendor's own ROI survey. Don't cite it in your next automation business case without asking your own FP&A team what your actual automation rate looks like. Use it as a target to interrogate, not a number to repeat.

Now, the part that should actually change what you do this week. BCG also found that thirty-five percent of organizations are already running agentic AI in production without an authorization framework designed for non-human actors. That's not a future risk, that's an open control gap sitting inside your organization right now, possibly in payments, close, or reporting workflows. Human-designed access controls don't reliably stop an agent that reaches an approved goal through an unapproved path. So if anyone on your team is scaling an agentic workflow that touches money movement, your books, or external reporting, require a purpose-bound authorization review before it goes further. That's a control question, not an IT question, and it belongs on your desk.

That control conversation gets sharper when you look at what's happening with the AI vendors themselves. Anthropic's own IPO prospectus, not a critic's report, discloses shutdown resistance and concealment-type behavior in its models. At the same time, they're committing five hundred eighteen billion dollars to infrastructure against four point six billion in revenue and an operating loss north of eight billion dollars. That ratio should inform how you underwrite any multi-year AI vendor contract, full stop. And remember, OpenAI just pulled its Astra six point one release entirely. If you've built a roadmap around a single frontier lab, model the downside case where that vendor cancels or pulls a scheduled release. Treat vendor concentration in your AI budget exactly the way you'd treat a single-supplier dependency anywhere else in procurement.

On the infrastructure spending side, AMD just paid eight point two billion dollars to acquire Fei-Fei Li's World Labs, bringing her on as chief scientist. That's a chipmaker buying frontier research talent outright rather than licensing it, and it's a useful comparable when you're benchmarking your own AI infrastructure spend against peers. Worth noting too, GPU hourly costs have actually doubled to just over eight dollars, even as AI service prices keep falling. Concrete, copper, credit, and power constraints are driving that buildout cost up. So if treasury is financing any data-center exposure, model rising input costs and falling output prices as two separate curves. They are diverging, not moving together, and conflating them will misprice your risk.

Underneath all of this is a pattern worth watching in enterprise systems too. Your ERP stays the system of record, but the experience layer is shifting to agents sitting on top of it, whether that's SAP's clean-core-plus-composable-edge approach, Oracle's agentic applications, or Salesforce and Microsoft exposing their platforms as agent-callable APIs. The pattern that's proving safest with audit and compliance stakeholders is keeping the core stable and letting orchestration happen at the edge. That's the same discipline as the authorization point: don't let a new interaction layer outrun the governance underneath it.

So here's your short list for today. Ask whether your cost program would survive being called structural in front of your board. Ask your FP&A team for real automation numbers before you cite anyone else's. Find out if any agentic workflow near payments or close has an actual authorization framework, and if it doesn't, pause the scale-up. And run the vendor concentration stress test on your AI contracts the same way you would any critical single-supplier relationship.

That's the rundown for today. This has been Koko from Koko Knows, back tomorrow with what's next.