AI adoption stalled at half of US business. Now they're putting their own staff on a meter.

Half of American business now pays for AI. The quarter that number crossed fifty percent, companies started rationing it.

Ramp’s spend index, built from corporate-card and bill-pay data across more than 50,000 businesses, put paid AI adoption at 50.4 percent in March 2026, up from 35 percent a year earlier. Then the climb stalled, holding at 50.6 percent in April. In the same window, the Wall Street Journal reported the other side of it: companies capping how much AI their own staff can use, because the cost is climbing faster than the budget.

Record adoption and rationing in the same quarter don’t belong to the same story. You don’t put a tool that’s paying for itself on a quota. Rationing is what a company reaches for when it committed to AI before it could predict the cost, and the only lever left is the people using it.

What a quota admits

A cap on your own staff is a confession. The company can see the total and not the inside of it. It knows the bill is too high and can’t say which work drove it, so it throttles everyone rather than the spend that actually hurt. That isn’t cost control, it’s cost avoidance, and it comes out of the same return the company bought AI to get.

The spend isn’t mysterious. It only looks that way because almost nobody on the finance side is watching the unit it bills on.

The bill is tokens

Strip it back and an AI bill is one quantity: tokens, billed by the million. You bought AI per seat, a fixed line on the budget. It bills per token, a number that moves with every request and that no seat count predicts. That mismatch is the whole surprise. A license is a flat fee. What the license generates is not.

Every overrun resolves to the same arithmetic, tokens times rate. Which leaves two ways to bring a bill down, and only two. Push fewer tokens through the model, or push the tokens through a cheaper model. Both are within your control, and a staff quota touches neither one.

Context is the part that rides on every call

The expensive habit isn’t how often the agent runs. It’s how much context rides along each time it does.

The standing instructions loaded at the start of a session get billed again on every turn that follows. A standing instruction set that has swollen because nobody pruned it becomes a tax on every request for the rest of the run. The discipline that controls this inverts the instinct: keep the always-on context short, and load the rest only when a task actually calls for it. Reference material parked in the window is paid for on every call whether the model reads it that turn or not.

Microsoft published a clean version of this from its own production agent, the one that triages Azure incidents. Its single biggest source of failures was metrics analysis: the agent kept pulling raw metric data into the context window to reason over it directly. The fix was to stop. Let the model decide what to compute, run real code to do the computation, and return only the answer. Failures on that task went to zero. And once the agent was no longer paying what its engineers called the token tax on raw data, it could analyze time ranges an order of magnitude longer for the same spend.

That pattern holds well past incident response. The teams that keep their bills down don’t hand the model the raw pile and hope. They put the bulk data behind a tool and feed back only the result. The other half is precision: an instruction exact enough to land on the first attempt, because a vague request gets paid for twice, once to produce the wrong thing and again to correct it.

The context you do resend carries a discount most teams leave sitting there. Cached input is published at a fraction of fresh input, around a tenth on current price sheets. A workload that holds its cache runs at a sliver of one that rebuilds context from nothing every call. The surprise bills almost always come from paying full rate for context that could have been cached, or never sent at all.

The model is most of the bill

The second lever is which model does the work, and the gap between rungs is enormous. The flagship model most teams reach for by default sits at the top of the price ladder, around $25 per million output tokens. The lite tiers run under fifty cents for the same volume. That’s roughly sixty to one for the identical job.

The catch is that you can’t just point the work at the cheap model and pocket the difference. A smaller model holds quality only when the task is defined tightly enough that it doesn’t need the larger one’s room to guess. Hand it a loose prompt and a fuzzy definition of done, and the quality drops, the work gets retried, and you’ve paid for the cheap model twice on your way back to the expensive one. That failure is what convinces a team it needs the flagship for everything. It doesn’t. It needs each job specified well enough that the smallest model that can do it actually does. The work of capturing the sixty-to-one isn’t picking a model from a dropdown. It’s the prompt and the spec underneath it.

Capping the wrong number

Rationing will spread through the rest of this year, and it will keep failing, because it caps the wrong number. A quota on access protects the budget by discarding the output the company adopted AI to produce. The companies that come out of this ahead won’t be the ones that used AI least. They’ll be the ones that learned to read the token line: trimming the context that rides on every call, and putting each job on the smallest model that can actually carry it. Anyone still rationing staff a year from now is doing it because they never looked inside the bill.