CFO Daily · Saturday, September 26, 2026

Hi, this is Koko from Koko Knows.

Let's get right into it, because the sharpest thing to land since yesterday is a governance signal, and it's one every CFO needs to hear. CIO Dive, citing an IDC survey of over four hundred enterprises, reports that two-thirds of companies now have AI agents running in production, and they're planning to sextuple that deployment by next year. Six times the agents. But here's the number that should stop you mid-coffee: only twenty-nine percent of those agents can actually talk to each other. Spend is scaling faster than coordination. That's exactly backwards from how finance controls are supposed to work, and it means integration debt is quietly getting baked into next year's base case before anyone's flagged it as risk.

So what do you do with that today? Before you approve any expansion of agent count in the next budget cycle, require an interoperability audit first. Not after you've signed the purchase orders. Ask your teams plainly: how many of our agents can pass context and data to each other right now, and how many are islands? If the answer is closer to twenty-nine percent than one hundred, you're funding parallel systems, not one coordinated capability, and that has real implications for close, for FP&A workflows, for anything that depends on a clean handoff between systems.

Second thing worth your attention: BNP Paribas just signed a five-year deal with Google Cloud for Gemini and Gemini Enterprise, but they did it with a hard line drawn in the sand. Sensitive customer data and critical operations stay on their own infrastructure, full stop. That's not a hypothetical safeguard, that's a live contract structure from a major regulated bank, and it's a template you can lift directly. When your team is negotiating the next AI vendor deal, get the data-sovereignty carve-outs written into the contract before signing, not as a cleanup exercise six months later when compliance asks questions you can't answer.

And it's worth putting that next to what Lloyds Banking Group is saying about its own agentic rollout. Their transaction banking tech lead describes it as small, deterministic, and supervised by design, working toward a two-billion-pound savings target through twenty thirty. That's a notably more conservative posture than the sixfold-growth story elsewhere this week, and honestly, that tension is useful to you. The next time a vendor pitches you on autonomous agents transforming your finance function, ask for the deterministic-versus-autonomous split. Make them show you exactly where a human is still in the loop and where they aren't. Don't take agent-count claims at face value just because a big number sounds like progress.

Now, layer in the capital question, because this is where it gets real for treasury and capital allocation. Akamai and Anthropic just signed a compute deal worth eleven point six billion dollars, expandable toward twenty billion, over seven years, with an equity warrant attached. This is your signal that AI infrastructure has stopped behaving like an opex line item and started behaving like a long-duration capital commitment, the kind you'd model like a lease or a major facility buildout. If you're evaluating similar AI vendor relationships, model the multi-year compute commitment as balance sheet exposure. And don't stop at licensing and compute. Factor in vendor concentration risk and, frankly, unresolved liability. There have been open incidents in this space with no cost estimate attached yet. That's the kind of unquantified exposure that deserves a line in your risk register even before someone can tell you what the number is.

One more item, smaller but worth noting if you've got tax leads in the room. The AICPA is rolling out a formal AI risk framework for tax practice. That's a shift from informal guidance to something that starts to look like a standard. My advice: get ahead of it. Start documenting how AI tools are being used in your tax workflows now, because when that framework matures, it's going to become an audit expectation, and you'll want a paper trail already in place rather than one you're reconstructing under pressure.

So here's the through-line for today. Spend is outrunning coordination on agents, capital exposure is outrunning your traditional TCO models, and vendor claims are outrunning what's actually deterministic versus autonomous under the hood. The move isn't to slow down adoption, it's to insist the plumbing, the contracts, and the documentation catch up before the next check gets written. Ask for the interoperability audit, get the data carve-outs in writing, and treat compute commitments like the capital decisions they actually are.

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