I spoke with several AI implementation leads this month at different companies, including one running a rollout in an organization of more than 100,000 people. Each of them had adoption metrics: seats activated, weekly active users, and the share of the company using the tools. Not one had a metric for what the spend returned, or what it cost per unit of work. When I put my CFO hat on and asked what they expected in return, the answer, more than once, was a version of "that is not my problem."

I keep coming back to that, because the budget responsibility here does not go away; it lands on finance.

The rollout leads are measured on adoption; the business units drive usage; IT holds the contracts; and CFOs are expected to manage the spend and demonstrate its return.

Want to work with me directly? Reply to this email for a 1:1 session or a custom program for your organization.

Budgeting and AI are my regular fall topics. This is what finance teams are focused on, so every year I look at the budgeting process through an AI lens. In 2024, I wrote about using AI to streamline your budgeting process.

In 2025, I suggested budgeting for licenses and training. Back then, most AI tools were billed at a flat per-user fee, and it wasn’t too high. At that price, enablement was the right approach: buy the seats, write a short policy, train people, let them learn, and do not expect a return yet. The spend was small enough that nobody had to defend it.

That will not survive the 2027 budget.

The assumption underneath last year's approach was that this would get cheaper. Models keep improving, and the price per token keeps falling, so the bill should follow. But this is not what happened. The price per token fell, but the bills went up anyway.

Gartner published the mechanism in August: inference cost per agentic workflow will rise more than fivefold through 2028. The cheaper the token, the more sense it makes to route work to a reasoning model. A reasoning model costs at least five times as much as a basic exchange, and people then build workflows that run dozens of those exchanges where they used to ask one question. The unit got cheaper, and the number of units grew much faster.

The vendors moved the same way. Slack folded two enterprise tiers priced at $20 and $32 into a single one at $45, and Salesforce sells Agentforce as consumption credits, Adobe's Firefly credits reset monthly without rolling over, and Microsoft charges a $200 monthly platform fee for Copilot Studio on top of the per-seat licence. The AI capability comes with a price increase, a new meter, or both.

Zylo's 2026 index found 78 percent of IT leaders hit with unexpected charges tied to consumption or AI pricing. One Fortune 500 company budgeted $500,000 for Microsoft Copilot and finished the year at $820,000, on token overages and tier thresholds nobody had modeled.

Forecasting works when you have a baseline and a driver: take last year's consumption, apply a rate, adjust for headcount or volume, and defend the number. AI consumption gives you neither. The metric did not exist two years ago, so there is no history to extrapolate from, and it is not tied to anything else you already forecast. It is not electricity, where the load is known and stable, nor is it water, where consumption per head barely moves.

What your organization consumes in 2027 depends on what people learn to do with these tools between now and then.

Then the unit itself moves underneath you. Token prices change, and what a token buys changes with them, often inside a couple of months, so the model you priced the budget on in September is not the model your team is using in March.

Flexera found that only 31 percent of organizations have accurate visibility into what they currently spend on AI software, so for most teams even the starting number is an estimate.

I went through the billing mechanics earlier this year, if you want a deeper dive.

So it turns out that budgeting AI for 2027 and beyond is a hard task, and there is no simple answer. In September I am running a series looking at it from three angles, each one aimed at a different part of the budget problem.

Throughout September, I will share the frameworks I am using with clients and the questions that come up as we work through them.

Nobody has fully figured this out yet, so I will be clear about where the answers run out and where we are still testing.

Before you can work on any of those three questions, you have to know what you are already paying for, and that is harder than it sounds. So here is a task for this week. Put together a list of the AI subscriptions your company is already paying for, with who owns each one and when it renews. Most people find the list takes longer than expected and still comes back incomplete, which is itself worth knowing.

For paid subscribers, I am sharing a prompt that gets you through that exercise faster and further. It maps your stack by capability rather than by vendor, which matters because invoices list companies, not capabilities, so anything you buy twice remains invisible until you reorganize the view. You can run it this week.

One more thing before you go. I would like to hear from you on this one more than usual. If your AI renewal has already come back this year, tell me what happened to the number. If you have tried anything since, tell me what you tried and whether it held.

Everything you share stays between us.

Reply to this email, or send me a note on LinkedIn.

Until next Tuesday.

Anna

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