CFO Daily · Thursday, September 24, 2026

Hi, this is Koko from Koko Knows.

Let's get right into it, because the sharpest thing that's landed since yesterday is a straight compliance play from Grant Thornton, and it matters a lot to how you build next year's control environment. They've published a framework for AI-driven SOX compliance: always-on monitoring, automated testing, intelligent controls replacing the old point-in-time sampling model. Sounds great for cost and reliability. Here's the catch they're upfront about: external auditors have not blanket-approved this approach. So if you're tempted to redesign your testing cycles around it, don't, not yet. Get your audit firm's specific position in writing first. This is a "call your engagement partner this week" item, not a "wait for the annual planning cycle" item.

That lands the same day as a genuine price war underneath the entire finance-agent stack. Anthropic's Claude Opus five point five and OpenAI's GPT-six Sol and Luna both cut API prices forty to fifty percent, landing within about ninety minutes of each other. If you signed a multi-year agent contract on last quarter's token pricing, you are almost certainly overpaying right now. That's an immediate renegotiation conversation with your vendor, not a someday conversation. And it tells you something bigger: when two frontier labs match each other's pricing that closely, the model itself stops being a differentiator. It becomes a commodity procurement line. The real question for your leadership team is where your actual edge sits: your proprietary data, your workflow integration, or your agent governance. Because model choice isn't a strategic bet anymore, it's a vendor comparison spreadsheet.

Third thing on my list, and this one's a budgeting blind spot: CIO Magazine is flagging that agentic AI scaling shifts the cost driver from user traffic to agent count and resource consumption. That's a new line item most FP&A teams haven't modeled at all. You've been used to thinking about software licenses as flat costs tied to headcount or seats. Agents don't work that way. The more agents you deploy, and the more they run, the more compute they burn, independent of how many humans are actually using the output. If your AI budget for next year is built on a license-style assumption, it's already wrong. You need a usage-based forecasting model, and you need it before you lock next year's numbers.

Now, a couple of proof points that this is real and moving fast, not theoretical. Liberty Global just signed a three-year deal to deploy conversational agents across eighty million connections. That's not a pilot, that's a full commercial commitment at massive scale, and it's a useful benchmark if you're trying to gauge whether your own organization's agent adoption timeline is behind or ahead of the market. And on the governance side, Cherry Hill Advisory is pointing to Alation's expansion into agent-aware data lineage tracking, meaning tools that can show exactly what source documents an agent used and when those documents changed. Internal audit should treat that as a baseline requirement to demand from any finance-agent vendor going forward, not a nice-to-have feature you ask about later.

So here's how I'd pull this together for you. The compliance framework and the price collapse are pointing the exact same direction: AI is getting cheaper and more embedded in your close and control stack faster than assurance standards can keep up with it. That gap is where your risk sits right now. Two things I'd do before you finalize next year's audit and tooling budgets. One, ask your auditors directly, in writing, what evidence they will and won't accept for AI-tested controls. Don't assume, don't guess, get specifics. Two, go back through every multi-year agent or AI vendor contract you've signed and re-price it against this week's cost drop. Forty to fifty percent off list price is not a rounding error, that's real money on your P&L, and vendors will not proactively offer it to you.

One more thing worth keeping on your radar, even though it's not urgent today: as agent counts grow, the McKinsey research on coding-agent spend is a useful cautionary tale, budgets there are already overrunning with underwhelming ROI. The lesson transfers directly to finance agents. The risk isn't really the sticker price of the tools. It's shadow AI and spend you can't trace. So whatever savings you capture from this week's price cuts, pair it with better usage tracking and chargeback discipline, or you'll just be trading one unmodeled cost line for another.

That's the rundown for today. Get the auditor conversation started, get the vendor renegotiation started, and start building that usage-based forecast before it becomes urgent instead of smart. Thanks for listening, this has been Koko Knows, and I'll be back tomorrow with what's changed next.