🚀 A CFO’s Guide to Taming the AI Chimera | The AI CFO

Last week, I took a four hour flight on a plane with no WiFi and went into what I call "Airplane Mode". I spent the entire time spilling thoughts into an Apple Note, capturing whatever came to my mind.

One realization hit me like a lightning bolt: somewhere between Opus 4.5 and Fable, I lost the plot. I was seduced by the siren song of vibecoding. I spent more of my time operating as a 1x vibecoder instead of a 10x CFO.

This admission stung, because I've been deep down the vibecoding rabbit hole since the launch of Claude Code. But it was also clarifying: now that I've gone on the hero's journey and come back a vibecoder, I'm excited to bring what I learned back to finance.

A CFO's Guide to Taming the AI Chimera

Q2 tech company earnings calls followed a familiar pattern. The CEO kicked off by waxing poetic about the massive ROI AI is providing for their business. Then they'd hand it over to the CFO, who laid out the hard facts: the company's AI spend is way up, but they're having a hard time quantifying the actual business impact.

➡️ In April, Uber's CTO had disclosed that the company burned through its entire 2026 AI budget in Q1 alone. By May, COO Andrew Macdonald admitted that he couldn't connect the company's rising AI spend to any measurable improvement in their product.

➡️ Microsoft tells a similar story. Earlier this year, they began canceling most of their direct Claude Code licenses, just six months after opening up access company-wide.

➡️ More recently, Meta announced that they are planning to sell their excess computing power to other companies to help cover the cost of their massive data center buildout.

The finance world is at a loss for dealing with AI spend because it doesn't behave like any line item a CFO has managed before. It's three kinds of spend fused into one, wearing a single disguise: the appearance of software, the management demands of a workforce, and the pricing model of a utility.

It's a completely different beast, and we need a new playbook in order to manage it effectively as AI becomes a bigger part of every company's spend profile.

 The 'own goal' of tokenmaxxing

To understand how we arrived at this point, you have to rewind to late 2025. LLMs took a genuine leap forward with the Opus 4.5 model cohort, and agentic tools went mainstream almost overnight. Companies raced to show they were adopting AI, and the tactic of choice was internal token leaderboards.

Microsoft, Amazon, and Salesforce all created their own versions, ranking employees by usage. The most notorious was Meta's "Claudeonomics," which ranked 85,000 employees by tokens consumed and was taken down within days of leaking publicly.

But at the same time, two other factors were at play:

1. Agentic workflows are MUCH more token-hungry than earlier single-turn versions. According to Gartner, agentic models require 5 to 30 times more tokens per task than a standard chatbot.
2. AI companies subtly shifted their pricing models from subscription-based billing to metered billing, eliminating the massive subsidies they'd been providing customers.

This combination of factors caused tokenmaxxing to slide into spendmaxxing. Agentic AI software spending is projected to hit $207B in 2026, up 139% from $86.4B the year before.

Suddenly, executives & investors looking at these trends realized that every dollar of AI spend needs to justify itself, else token bills will spiral out of control.

The CFO's instinct (and why it might be wrong)

I have zero doubt that AI is providing value for early adopters; research from Ramp shows that companies leading in AI are growing faster than their peers. So the CFO's responsibility is to ensure that their company is spending AI well.

But you can't manage AI spend like a traditional software subscription, because it's a brand new beast.

The Chimera

AI spend is a Chimera. It doesn't fall neatly into the existing software category, but instead looks like three traditional types of spend blended into one.

AI spend looks and feels like a software expense at first glance. It shows up as a vendor line in your P&L and you 'subscribe' like any other SaaS tool. But there are a few major differences:

1. Variable Costs: Per-token pricing means that AI is billed on a meter, like electricity or any other utility. The cost of an Anthropic/OpenAI 'seat' swings wildly depending on what tasks you point it at, and which model you use for those tasks.
2. Variable Returns: Two 'identical' $10,000 Claude invoices can vary widely in the value they deliver, in the same way two similarly-paid employees can. The ROI an AI workflow delivers is a function of how well you train it, direct it, and hold it accountable, same as any employee. (I wrote about this in a recent AI CFO essay). This is the part most companies have yet to figure out. It's also why the tokenmaxxing leaderboards went so wrong: rewarding tokens burned is rewarding activity over outcomes.

AI is a new type of resource: priced like a utility, managed like a workforce, and presented to customers in a software package.

Once we see AI spend for what it is, it's clear that the old playbooks no longer apply. So what should this new playbook look like?

Taming the Chimera

The CFO's job is to 'tame the chimera': create visibility into what's being spent, understand where it's working and where it isn't, and allocate AI budget accordingly.

Borrowing from the AI spend playbook our friends at Ramp are following:

➜ See it. One of the golden rules of finance is that what gets measured, gets managed. We need to start measuring AI spend with the same level of granularity as headcount spend is measured today.

➜ Understand it. Once you understand the raw dollar spend, the next step is connecting that spend to real outcomes. This step is critical for a CFO of an AI-forward company; implementing hard caps or cutting spend indiscriminately is short-sighted.

➜ Control it. Once you have visibility, you can treat AI budgeting like any other scarce resource. Partner with team leaders during budgeting to have each one make the business case for their spend, so the dollars flow to the highest-leverage uses instead of getting spread evenly across the org.

Each of these deserves its own deep dive. Over the next few issues, I'll dig into each one so you have a concrete framework for understanding what your AI spend is actually doing.

For now, the CFOs who start here will be ahead of the conversation when the board comes asking.

📡 On My Radar

• On becoming an UnCFO: Unconventional AI is one of the coolest companies you haven't heard of (and an airCFO client 😎). This essay from their CFO, Ali, is a great reflection on how the CFO's role will shift in the age of AI.
• On supercharging finance with AI: CJ Gustafson's latest episode with the CFOs of Opendoor & Datadog provided some great examples of the mindset required to lead an AI-powered finance team.
• On getting Funded: Andrew Rea just announced TaxWire's Series A and I got to sit down with him for the full retrospective on all three rounds. He was more honest about the process than most founders are willing to be. Check out the latest Funded episode.

That's all for this edition of The AI CFO - we'll be back in your inbox soon with more musings on the future of AI-powered finance & operations. Please reply to this email directly with any feedback/suggestions/just to say hi!

Cheers,

Alex Wittenberg, CEO @ airCFO