You Can See What AI Costs. You Can't See What It's Worth.

Usage-based AI billing put a per-person price tag on knowledge work. But it shows only one side of the ledger — and the reflex to cap the side you can see costs more than the spend.

A few days ago Sam Altman described (1:01 mark of the interview here) a complaint he now hears on repeat: “My company spent my entire 2026 budget in Q1. Can you make this more efficient?” He continued by describing the shift: “at the beginning of this year, an issue that never came up. People were totally happy with the amount they were spending, to all of a sudden, a huge issue.” What’s telling isn’t the overspend. It’s that the same spend went, in his words, from a non-issue to a crisis almost overnight. The bill didn’t change character. The demand to justify it arrived.

Plenty of software is billed by consumption — APIs, data warehouses, and messaging platforms have worked this way for years. What’s new is the shape of the bill. You can see, to the dollar, what each employee’s AI use costs. You cannot see — not even roughly — what it produced. The cost is legible, per-person, assignable. The value is none of those things. It’s a one-sided ledger: the cost column filled in, the value column blank.

Now add the anchor. That per-person cost is measured against a baseline of zero, because a year ago it was zero. So every dollar reads as additional — and additional spend has to justify itself. That justification is exactly what can’t be produced, because the value is invisible. The spend stays under suspicion — not because it’s wasteful, but because no one can show that it isn’t. That’s the real engine behind the cutting: not that AI is expensive, but that its cost is the only half of the equation anyone can defend.

In situations like this, companies reach for the obvious lever: they cap it. Often uniformly — every seat gets the same monthly ceiling. The uniform cap isn’t the last resort; it’s the first reach. When you can’t tell which employee’s $500 is leverage and which is noise, “everyone gets the same” is the one allocation nobody can blame you for — which optimizes for the allocator’s blamelessness, not the company’s return. You’d never ration sales travel or R&D evenly across every head regardless of return. Uniform allocation is fairness standing in for information, and it taxes your best users to subsidize your most timid. Uber ran the whole arc in months (TechCrunch): it pushed staff to use AI as much as possible, ranked them on leaderboards, then — after burning its annual budget in four — capped every employee at $1,500 a month. (Across-the-board cuts, efficiency mandates, and forced-adoption quotas are the same move in different clothes.)

There is an obvious parallel to draw. Metered, consumption-based, variable by team — that was cloud a decade ago, and we built a whole discipline for it. FinOps. You tag every dollar to the workload that consumed it, then judge it against what that workload is worth. The companies that panicked and throttled cloud uniformly strangled their own products; the ones that tagged and allocated won. Just do that for AI.

It doesn’t transfer, for one reason: cloud cost attaches to projects; AI cost attaches to people.

A cloud workload rolls up to a product, and a product has revenue — a price. So when you tag cloud spend, the tag lands on something already denominated in dollars of output. Tag AI spend with the same rigor and the tag lands on a person — and a person’s output isn’t priced the way a product’s revenue is: not cleanly, not per unit of work, not in a number you can set next to the cost. The tagging was never the hard part. The hard part is that the only stable place the spend sits is the one entity in the building without a price.

So if you can’t make the value legible — and you probably can’t, not per-person, not soon — the answer isn’t to measure better. It’s to change what you compare against.

The whole problem is the zero anchor. Stop comparing AI spend to nothing, and start comparing it to the real alternative — which, whenever a team has needed to do more, has always been people. Say your six-person team needs more output. You can hire a seventh — call it $150K fully loaded — or spend to make the six you have faster. To win that comparison, AI doesn’t have to replace anyone; it only has to make the team enough better that the seventh chair can wait. Load both sides honestly — AI plus the time to wield it, against a salary plus management, onboarding, and the lumpiness of a permanent head — and at $500 a person it isn’t close. Even at the rate that triggered Uber’s cap — $1,500 a month — six people run less than one hire. The point was never that AI is cheap. It’s that “cheaper than the alternative” is answerable. “What is it worth” isn’t.

That reframe doesn’t require solving measurement; it moves the question to an honest altitude. From there, two rules. Judge the spend high — at the team level, against the alternative cost of additional capacity. Allocate it sharp — never uniformly, because the value is uneven even though you can’t see how. Most companies do the exact opposite: judging at the individual level against zero, rationing in a flat line across everyone. Judging where it’s hopeless; cutting where it does the most damage.

Which leaves the question many people are thinking yet few say out loud: if the value of AI really is invisible, is it even there?

You already know the answer, because you’ve been paying it for years. You don’t meter your analyst’s output task by task. You bet on a new hire months before a single review tells you whether you were right. Human value is exactly this invisible — we’ve made our peace with paying for it on judgment. AI doesn’t fail a test that people pass; it fails a test we never gave people.

So the asymmetry was never that AI’s value is uniquely invisible. It’s that AI’s cost is uniquely visible. The meter is the new thing in the room — not some special doubt about what AI is worth. Its worth is as real, and as unprovable, as that of the people working beside it. AI just had the bad luck to arrive with a price tag attached.