In April 2026, the median American business spent $2,246 a month on AI services. The average business spent $140,842. Both figures come from Ramp’s analysis of transaction data across more than 70,000 U.S. companies.
That spread, roughly 60:1, is unusual for an established expense category. Utility bills and telecom contracts do not distribute that way, because consumption in those categories stays broadly proportional to the size of the organization. When an average sits that far above the median, it indicates that a small number of firms are operating on a completely different scale from everyone else, and that most of them arrived there without deciding to.
Most businesses are currently near the median and are budgeting on the assumption that they will stay there. That assumption deserves more scrutiny than it usually receives.
Why the gap exists
AI reaches the P&L through two pricing models that behave nothing alike, and most organizations record both of them as software.
Subscription seats do behave like software. The cost is bounded by headcount, procurement approves it once, and the monthly figure is predictable enough to forecast a year out. Token-based usage works on entirely different terms. It is metered consumption billed by volume, and its ceiling is set by how much work gets sent to the model, rather than by how many people are employed. An organization can add no new licenses at all and still see its AI bill triple over a quarter.
This is why companies tend to move up the spend curve without any corresponding procurement decision. The increase usually traces back to an automation running more frequently than anyone modeled during design, or a process retrying on failure, or an integration passing an entire document to a frontier model when a single section would have answered the question. None of those register as a “purchase.” They register as a larger number on an invoice that still sits in the software line.
The same pattern appears when you normalize for company size. Ramp puts the median at roughly $46 per employee per month, with the middle half of companies spread between $3 and $352. A hundredfold difference in per-head cost between two firms of similar size in the same industry has very little to do with what their vendors charge and almost everything to do with how each of them uses the technology.
The pricing you are budgeting against is not fixed
There is a second reason to address this sooner rather than later. Current pricing for frontier models reflects a period of heavy subsidization, funded by losses the model builders have been reasonably open about. Those same companies are moving toward profitability and, in several cases, toward public markets, which makes it unlikely that today’s rates represent a stable long-term baseline. Organizations building their AI budgets around current unit economics are, in effect, forecasting against someone else’s willingness to absorb losses.
None of this warrants alarm, and none of it argues for slowing adoption. The practical implication is narrower: bringing consumption under control is considerably easier during a period of forgiving pricing than after that period ends.
Four questions worth answering before the next invoice
Who is using AI, and for what work?
This is a question about workloads rather than software seats or licenses. Most organizations can name their vendors but cannot describe the tasks those vendors are being used for, which reduces any subsequent cost conversation to guesswork. A usable inventory identifies which teams are using which tools, for which categories of work, and at what frequency. It also needs to account for automated processes that run without a person present, since those often represent a disproportionate share of consumption – precisely because nobody experiences them as usage.
Is each task running on an appropriate model?
Model selection is the largest single cost lever available to most organizations, and it is routinely left untouched. Frontier reasoning models cost several times what smaller models cost, and a substantial portion of the work sent to them consists of classification, extraction, summarization, and formatting, all of which smaller models handle at a fraction of the price with no measurable difference in output quality. Where a model exposes a reasoning effort or extended thinking setting, that control carries direct budget implications and should be matched to the difficulty of the task rather than left at its maximum by default.
Are requests written well enough to avoid rework?
An imprecise first prompt generally costs twice, once for the response that missed the mark, and again for the correction, which makes effort spent on the initial request a cost measure as much as a quality one. Output length works the same way. A user who needed three sentences and received three paragraphs has paid for the difference, and stating the expected length of a response is among the simplest reductions available to any organization.
Can spend be attributed to a function?
A vendor invoice records what was paid. It does not indicate which department incurred the cost, which workflow generated it, or what the organization received in return. Without that attribution, productive spend and waste are indistinguishable, and there is no basis on which to defend the budget for work that is delivering. Some AI work supports revenue directly and a great deal of it produces cost avoidance or recovered capacity instead. Both are legitimate outcomes, but neither can be evaluated while the spend sits in a single undifferentiated line.
What managing it actually involves
Very little of this requires specialized tooling to begin. Most organizations can make meaningful progress by separating AI from general software on the chart of accounts, tagging spend to the department or function that generated it, setting consumption thresholds with alerting on any account that supports automated workflows, and then reviewing those figures on the same cadence as any other variable operating expense. The providers expose usage data at a level of detail that supports this, and spend management platforms have started ingesting it directly.
The harder part is organizational rather than technical, because someone has to own the number. In most companies, AI spend currently belongs to whoever signed the first contract, someone who is rarely the person best positioned to judge whether the consumption behind it is reasonable.
A familiar problem in an unfamiliar category
For most of the past two decades, MSI’s teams have worked with organizations that could state their copier costs to the dollar, while having no visibility at all into what their printing cost them. The equipment sat on a contract, so the expense felt known. The consumption behind it stayed invisible until someone metered the fleet and attributed volume back to departments, at which point the waste tended to surface quickly and in places nobody had suspected.
AI occupies a similar position today, without the meters. The subscription is visible while the consumption is not. Companies currently sitting near the median are generally there because their usage has not yet scaled, rather than because they have found some structural advantage. As adoption broadens across departments and more work shifts into automated processes, that usage will scale. The question is whether anyone will be tracking it when it does.
If you are uncertain which direction your organization is heading, that is a discussion worth having before the next budget cycle rather than during it.
AI for business, done responsibly.