CFO Daily · Friday, September 18, 2026

Hi, this is Koko from Koko Knows.

Let's get right into it, because the biggest news today is a direct response to something I flagged you about earlier this week. OpenAI just shipped an upgrade to its Admin Console that links ChatGPT and Codex usage to actual task-level outcomes and ROI, not just raw token spend. Remember that Accenture finding that eighty percent of AI spend can't be attributed to any business outcome? This is the first vendor-side answer to that exact problem. OpenAI's even running an illustrative sales case claiming two hundred forty-five percent first-year ROI. Now, take that number with the grain of salt it deserves, it's a vendor case study, but the capability itself matters. If your AI vendor can now show outcome tracking by work category, that becomes your new floor for what you accept in a renewal conversation. My advice: stop letting vendors hand you an aggregate usage dashboard and call it reporting. Starting today, make task-level ROI reporting a contractual condition, not a nice-to-have feature you hope shows up eventually.

Second thing worth your attention: there's fresh research out of Berkeley, cited by Tom Tunguz, showing that the software wrapper around an AI model, what people are calling the "harness," can cut the cost of an equivalent answer by seventy-one percent with zero loss in accuracy. Let that sink in. That's not a model upgrade, that's an architecture decision, and it's sitting right there for the taking. If your FP&A team is running vendor bake-offs right now and only comparing model benchmarks or per-seat pricing, you're missing the lever that actually moves the cost curve. Get your team asking vendors directly: what does your orchestration layer look like, and can you prove a cost-per-answer number, not just a price list.

Third, and this one's more of a signal than an action item today: a former OpenAI researcher has built a new model at a company called TypeSafe AI that skips the verbose, human-readable output entirely for machine-to-machine communication. Think about all the token bloat baked into automated workflows just because the model is talking like it's writing you an email. As you build out agentic close and FP&A processes that run at volume, this points to a coming category of tooling optimized for cost-per-decision rather than cost-per-seat. I'd set aside a small pilot budget line to test this as it matures, separate from your general AI seat spend, because the economics could look very different.

Now, here's where I want to connect some dots for you. There's a stat from Auditoria that should make every CFO pause: sixty-six and a half percent of finance teams are increasing their AI investment, but only twenty-one percent are seeing measurable results. That's not a technology problem, that's a capital allocation problem, and you'd never accept those numbers on any other capital project. Before you approve the next AI budget line, hold it to the same evidence standard you'd demand anywhere else in the P&L. And while you're at it, stress-test your workforce assumptions too. Census data is now showing AI-exposed new graduates already taking a thirteen percent hit in starting pay and a five-point drop in employment probability. If your workforce planning assumes AI substitution savings that haven't materialized yet, that's a real risk sitting in your model.

One more thing on the radar, less urgent but worth tracking: Nokia and Microsoft just partnered on an agentic data foundation for telecom operators, cutting data integration time from weeks to minutes. This is another data point that agentic tooling is landing in core operations first, not just office productivity tools. Expect vendor-integration and data-governance questions to hit your shared infrastructure budgets before they ever touch your close process, so if you're not already looped into those procurement conversations, get looped in now.

If I zoom out, here's the throughline I'd leave you with today. The market just handed finance leaders an actual toolkit, outcome-linked dashboards, harness-level cost efficiency, machine-native models for high-volume workflows, and a growing library of ERP vendors building governed, auditable agent layers on top of trusted systems of record. That last point matters too, keep your ERP as your system of record and let the agentic layer sit on top of it, not replace it. The pattern the safest enterprises are using is clean core plus composable edge, decoupled extensions that don't destabilize your audit trail.

So today, three things. Ask your AI vendors for outcome-based ROI reporting before your next renewal. Push your teams to scrutinize harness architecture, not just model pricing, in any bake-off. And hold your AI budget line to the same proof-of-results bar as any other capital request on your desk.

That's your rundown. This has been Koko Knows, I'll be back tomorrow with what's new.