Last week, OpenAI released GPT-6 Astra, and within days my feed filled with people switching back to ChatGPT from Claude, so I opened my ChatGPT, set it loose on an agentic workflow, gave it a real task, and let it run.

Twenty minutes later, I checked back and it had used up my entire daily token limit, then asked for permission to spend more money on this task. A hundred dollars and two hours later, the task is still not complete.

Not that I care too much about $100. What actually stayed with me was different: how quickly the tokens disappeared, how little visibility I had into what my experiment would cost, and that after completing it, I didn’t gain any additional clarity.

That kind of test is happening everywhere right now, which is part of why finance teams are calling this year's AI bill a genuinely different problem than last year's. Below, I lay out the three things converging into that problem, and the two-stage sequence you actually have for managing it once a capability is live.

That test is one small example of something bigger. Three things are converging right now, and none of them are new on their own, but together they add up to what a lot of finance teams are quietly calling a budget disaster.

Last week's edition told you about the rollout leads who could report every seat activated and every login, and nothing about what any of it produced; that's the third piece showing up in real conversations, not just mine.

Put those three together, and the pricing confusion turns out to be the smaller problem. Capability outran the organization's ability to manage it, and the bill kept climbing underneath that gap.

As the CFO, you don't have a lot of direct control over any of this. You don't set what a vendor charges next month, and you don't set how much more capable the next model release is. What you do control is how a capability moves through two stages, in order, before it becomes a permanent line in the budget.

The first stage is enablement. Every new capability starts here: a defined team size, not the whole company; a budget ceiling that caps what it can cost while you're finding out if it's worth it, and enablement metrics tracked instead of return: how many people are actually using it, how often, and what's getting in the way.

The second stage is the business case. Once a capability has proven itself in enablement, the job is to push it toward a real use case with a real cost attached: what it costs to run, what it's supposed to produce, and someone who owns that number.

A capability that never makes this move either earns a permanent, unmonitored place in the budget, or should have been turned off already.

The framework above tells you what order to move a capability through; the subscriber section turns that order into a working file.

Both stages include a worked example in the attached spreadsheet, based on real Claude Team pricing, so you get a model to edit rather than a concept to reconstruct yourself.

Those twenty minutes with GPT-6 Astra taught me something more useful than a cost lesson: how quickly capability outpaces the plan to use it well. Pick a mode for whatever you enable next, write down the end date, and you're already ahead of most of the rollouts I saw this month.

Thanks for reading, and I'll see you next Tuesday.

This newsletter is for you, and we want to make it as valuable as possible. Please reply to this email with your questions, comments, or topics you'd like to see covered in future issues. Your input shapes our content!

Want to dive deeper into balanced AI adoption for your finance team? Or do you want to hire an AI-powered CFO? Book a consultation!

Did you find this newsletter helpful? Forward it to a colleague who might benefit!

Until next Tuesday, keep balancing!

Anna Tiomina
AI-Powered CFO