The median company in Ramp’s July figures spent $11.95 per employee on AI for the year.
That is about a pound a month per person. Not a budget. A rounding error.
If you have spent the past two years assuming everyone else got on with this and you are the one running late, that number is the correction. Most organisations are doing almost nothing.
The spread is roughly 600 to 1
Ramp publishes a monthly AI Index built from what its own customers actually spend. The July breakdown:
- Top 1% of businesses: a median of $7,400 per employee
- Top 10%: $650 per employee
- Median firm: $11.95 per employee
Per employee, per year. Those units matter and they are easy to lose.
Almost every conversation about the AI gap is conducted at the $7,400 end. That is where the case studies come from, and the conference talks, and the posts telling you what you should already have done. It is a real group of companies and they are spending seriously.
They are also the top one per cent.
The honest comparator is the middle, not the leaders
The number that describes your position is $11.95, and it is worth understanding why the median is the right one to look at rather than an average.
A handful of firms spending $7,400 a head will drag any average a long way upward. The median ignores them by construction: it is the company sitting in the middle of the list. So when you picture the organisation you are competing with for talent, or for clients, or for the same contract, the picture to hold is the one spending about a pound a month per person on this.
The gap is not between you and the average. It is between the average and a small group of outliers spending a thousand times more.
Wide adoption, almost no depth
Here is where it gets more interesting than a simple “nobody is doing it” story, because that is not what the data says either.
Ramp reports that 43.5% of US businesses in its sample paid Anthropic something in July, up 1.1 points on the month. OpenAI reached 39.7%, up 0.23 points. So roughly four in ten have a paid relationship with each of the two largest suppliers.
Nearly everybody has bought something. Hardly anybody is spending. That pairing is a more useful finding than either half on its own, and it describes something most of us would recognise: the licences exist, a few people use them, and nothing about how the work gets done has changed.
One more detail worth having. Growth in first-time buyers has slowed at both of the two largest suppliers, and Ramp is explicit that this is not people defecting to cheaper or Chinese models. First-time buyers still go to the American vendors. The extra spend is coming from companies that were already spending heavily, going deeper.
That is a concentration story rather than a saturation story. It is the honest version of “the gap is widening”: not everyone racing ahead, but a few pulling away from a flat field.
The market was offered the best model and mostly said no
The second finding in the index is about price, and it answers a question I get asked often.
Anthropic released Fable 5, the most capable model in its range, during the period Ramp is measuring. In its first month it took 6% of the tokens businesses bought from Anthropic and 11.4% of the money, because it costs more per token than everything else on the menu.
Ramp’s own reading:
“with Fable 5, we’ve found a new upper bound for how much businesses are willing to spend on AI. Here, more performance is not worth the price tag.”
Businesses were shown the best thing available and mostly declined it on cost. Not because they could not afford it, but because the cheaper model did the job.
If you have been holding off because you assumed getting value out of this means buying the most expensive thing on the market, the market has already tested that for you and reached the opposite conclusion.
What this is not
Ramp is a corporate card and spend management company. The index is built from its own customers’ card and token spend, so this is company data with a commercial interest attached, and it is not neutral research. Attribute it to Ramp, and treat it as a large well-placed sample rather than a survey of the economy.
Three further limits. The sample is US businesses, so do not read these as global or regional figures. Ramp says the token-level chart “skews slightly more tech-y than our typical AI Index sample”. And Fable 5’s low share is one month of data on a model that was brand new in the period, which shows a price ceiling rather than a verdict on the model.
None of that changes the shape. A 600 to 1 spread does not come from sampling noise.
The question to take into the week
Work out your own figure. Total annual AI spend, divided by headcount.
Most people cannot answer that in under a week, which is itself the finding. The spend is scattered across individual subscriptions, a team licence somebody expensed, and a pilot that never got cancelled. Nobody owns the number, so nobody manages it.
Whatever it comes to, you now have something to put it against. Under $12 and you are the median, which is a more comfortable place to be standing than most of the commentary suggests. Over $650 and you are in the top ten per cent, in which case the question is not whether you are spending enough. It is whether you can show what the spending bought.
On Thursday I want to look at that second question through a company that spent $40 million building its own model, and published where the money actually went. Almost none of it went where you would expect.
Source: Ramp AI Index, August 2026, published 12 August 2026. All figures are Ramp’s, drawn from its own customers’ spend, and describe US businesses.