Sarah Friar, OpenAI's CFO, said something recently that I'm stealing with pride: enterprise customers have moved from "tokenmaxxing" to "cost per unit of intelligence."

Tokenmaxxing is a great word for a bad habit. It's when a team measures AI success by how much they used it. More prompts, more integrations, more seats. Activity as the scoreboard.

Cost per unit of intelligence is a different question entirely. It's not "how much AI did we use." It's "what did we actually get for it."

Friar broke that down into four questions she asks about any piece of AI-generated work:

1. Did it complete work that mattered?
2. What did it cost, including employee time, review, and rework?
3. Was the result good enough to use?
4. Did it help us move faster or make a better decision?

I love this framework because it's the first AI measurement I've seen that accounts for the invisible cost. Every leader I talk to can tell me what their AI subscription costs. Almost none of them can tell me what the review and rework cost. That's the number that actually determines whether AI is paying for itself.

Here's where I'd push it further, because "did it help" and "was it good enough" are still soft calls until you attach them to something you can point to.

For question 1 (did it complete work that mattered), don't just ask if the task got done. Ask if it was a task that was already on someone's list, competing for their time, with a real cost to not doing it. AI finishing something nobody needed is activity, not value.

For question 2 (what did it cost), time the review. Not the AI generation, the human part. If a first draft takes AI 90 seconds and your team spends 45 minutes fixing it, that's the real cost of the unit. Most teams have never timed this once.

For question 3 (good enough to use), separate "good enough to send" from "good enough after I fixed it." Those are different outcomes and they get reported as the same win.

For question 4 (faster or better decision), pick one number you'd have tracked anyway. Time to close a deal. Time to first draft. Number of options you actually considered before deciding. If you can't name that number before you start, you won't be able to answer the question after.

The pattern underneath all four is the same one I see in every client engagement that actually sticks: measure the work, not the tool. A subscription is easy to justify. A workflow that holds up under its own numbers is a different conversation, and it's the one worth having.

Alex