Most enterprise AI arrives priced per seat and per unit of consumption, often both on the same contract. Per seat is settled the day you sign. You know the count, and as more people use it, cost per active seat looks better every month. Per token moves, and it moves upward with usage — which is the thing everyone else in the building is being congratulated for - “Tokenmaxxing”.

Both numbers are correct. Both divide the cost by something you bought.

When the contact center costs more, you know why: more cases came in. Spend over cases closed gives you eleven dollars a case this year against thirteen last year. The unit underneath is work the business delivered.

Seats and tokens tell you how much access you bought and how much compute got consumed. Neither says what came out.

And nobody can tell you what the tokens were spent on. A token is not a unit anyone in the business recognizes. Not a case, not a transaction, not an hour. It measures something the model did. So the bill goes up, and the reason it went up isn't legible to the people being asked to approve it.

Which leaves you holding two bad options.

Cap the consumption, and you hear about it inside a week. People say they can't do their work. Finance becomes the department slowing AI down — I've heard operators put it almost word for word: the finger points at Finance the second they try to rein in spend and reduce user tokens.

Leave it uncapped, and eventually the board asks what the spend bought. The honest answer is that a few thousand people are each having a slightly better day. A cleaner email, a faster first draft, a meeting that got to the point sooner. Real value. Not a line you can present.

From outside the seat this reads like a measurement failure. It isn't. The structure of the deployment decided which unit the value would arrive in, and that decision usually gets made in IT, sometimes by the CEO, rarely with finance in the room.

And it wasn't a bad decision. Giving everyone access is worth doing. It just delivers its value in invisible “quality” units and its cost in highly visible financial units.

IT is looking at a success. Access went out, usage is climbing, and adoption is the number they're accountable for. It is a good number. But that is also the “success” that is causing your bill to rise.

Operations is looking at something it can't account for. Teams are busy, people say the tool helps, yet throughput hasn't moved. Their reporting doesn't explain it, because the tool never entered a process their reporting covers.

The board is looking at three updates that don't reconcile, and it turns to the one person whose update is written in dollars.

Nobody in that picture is wrong and nobody is hiding anything. Each seat is holding a different true number about the same program. Yours is the only one denominated in money, which is why the mismatch surfaces in your chair first — and why the one lever you have, capping consumption, breaks IT's number and Operations' people at the same time.

I had a call with a CIO who had bought an AI tool and given everyone access. Usage was fine. Nothing had moved in the numbers they report — cases closed, revenue, sales, none of it — and the bill kept climbing. What I told them came in two parts.

The first is training, which most companies skip on the way to buying the next thing. People take the first answer a model hands them when the tool rewards a second and third pass — cheap to fix, and real return you will never see on a page.

The second is where you get something tangible you can put in front of a board: build AI into the workflows you were already measuring. A workflow already has a denominator. You were reporting on it last year and the year before, so the number you already report tells you whether AI helped.

Broad AI access buys a better organization. Embedding AI into workflows returns tangible improvements you can share.

The question I'd open with: name one process where AI is sitting inside the workflow, and tell me what that process's number was before, and what it is now.

Expect a usage answer back — seats active, prompts run, hours people say they saved. Nobody's being evasive; that's the only answer available when the tool hasn't entered a measured process yet. It's a location, not a dodge. It tells you where things stand.

I wouldn't spend much of this year trying to produce a single company-wide AI return figure. That effort usually produces a slide rather than a decision. Begin evaluating one workflow at at time, and let the company-level story get built out of pieces that are each individually true.

One thing this week: ask that question about one process that was already being reported on before AI showed up. If nobody can produce the two numbers, the tool hasn't been put where the numbers already live. That one is fixable.

— Isaac