CFO Daily · Tuesday, September 22, 2026

Hi, this is Koko from Koko Knows.

Let's get right into it, because the sharpest signal today is one that should make every CFO pause: a new report out this morning finds multi-agent AI systems are now executing supply-chain actions directly inside ERP systems. Freight re-routing, dock allocation, safety-stock rebalancing, and at targeted checkpoints, the human approval step has simply been removed. This isn't a pilot or a dashboard demo. This is agents making operational calls and the system just... proceeding. I'm flagging this first because it's the exact template you're already wrestling with for close and payments automation. If you're piloting agentic finance workflows anywhere in your stack, the homework today is to go map precisely where approval gates disappear in that supply-chain deployment, before you scale anything similar in finance. Not after.

Second thing on my radar: OpenAI shipped a new agentic platform, called V7, built specifically with finance and insurance in mind. The pitch is "institutional memory" — solving the problem where an agent can't tell which fund report is current, or how the same entity maps across five different systems. That's the exact failure mode behind a lot of reconciliation errors. So if you're evaluating agent vendors right now, don't ask them for a productivity number first. Ask them directly how they handle cross-system entity resolution. That's where the errors actually originate, and it's a much better filter than a benchmark slide.

Which brings me to BCG's new harness-engineering framework, making the rounds this week. They're citing a bank deployment with over thirty percent productivity gains, and the headline claim is that model choice is secondary — governance structure is what determines whether these systems scale or fall apart. I'd treat the specific number as one advisory firm's case study, not proof. But the sequencing logic is sound and worth stealing for your own pilots: build the control layer first, then scale. Controls before scale, not scale then patch controls in. That ordering is the single most useful checklist item out of any of today's reporting.

Now, a capital markets note that FP&A should have on the radar. AWS just committed to over three million additional Nvidia GPUs through twenty twenty-nine, on top of what they already run. Neither AWS nor Nvidia is disclosing the cost. At the same time, Nvidia's carrying a hundred-and-five-billion-dollar residual-value guarantee tied to its Ohio campus build. That's compute capex compounding fast, locked in well ahead of confirmed demand. If your compute costs get passed through from any vendor in this chain, don't model a single number — build wide sensitivity bands, and treat those guarantees as contingent liabilities in your vendor risk assessments, not just growth signals.

On the capital markets side more broadly: Oura's IPO filing at up to two-point-two billion dollars, running alongside Anthropic's hundred-billion-dollar run rate, tells you the IPO window is reopening and investors are still pricing AI-adjacent growth stories generously. If you've got a liquidity event on the horizon, treasury and IR should be watching this window closely. It may not stay open long.

Here's the throughline I want to leave you with. Across every one of these stories, agent authority is expanding into real operational execution faster than oversight structures are catching up. That's not a hypothetical risk anymore, it's what's happening in supply chain right now, and EY's own survey this week confirmed the governance gap is real and unresolved across the enterprise. So here's the one rule I'd operationalize today: before you approve any agent's move from advisory to execution in a finance workflow, require a documented control checkpoint as the gating criterion. Not a vendor benchmark, not a productivity claim, a control checkpoint you can show your auditors. That's the difference between adopting this responsibly and finding out the hard way where your approval gate used to be.

That's the rundown. I'm Koko, and this has been Koko Knows. I'll be back tomorrow with what's changed next.