The AI Index Built From Receipts

Every AI valuation rests on one assumption. Businesses will keep paying more for AI every year. Almost everything I read about that assumption comes from people selling AI. Labs report their own revenue. Consultants survey executives about what they plan to buy. The buyers rarely speak.

Last week I found a dataset where they do. The Ramp AI Index records which American businesses paid for AI each month and how much they paid, straight from the receipts. It is free and it updates every month. Here is what I learned.

Disclosure up front. UpRound holds indirect exposure to Ramp through USVC. Everything below is public.

Ramp is a corporate card and spend management company. More than 85,000 American businesses run their spending through it. Its economics lab turns those payments into the index, and Ara Kharazian, Ramp’s lead economist, writes it up every month.

The rule is simple. If a business paid for an AI product in a given month, it counts as an adopter that month. Most adoption numbers come from surveys, where a manager is asked whether the company uses AI. The index skips the question and checks whether the company paid.

When Ramp launched the index in spring 2025, the sample was about 30,000 businesses using its cards and bill pay. It has nearly tripled since. The St. Louis Fed now carries the series in FRED, its public economic database, with history back to January 2023.

The sample has limits. Ramp’s customers lean toward startups and technology companies. Large enterprises that run spending through other platforms are missing. Free tools and personal accounts never appear as a payment, so the index undercounts true adoption. And it measures what Ramp’s customers pay. It does not measure any AI company’s revenue.

The index started with one question, whether a business pays for AI at all. In June 2026 it began measuring how much businesses spend per employee. It now also tracks AI tokens every week. A token is the unit AI models bill by, roughly a word or part of one. The token data comes from a smaller Ramp product that Ramp says is not representative of all its customers, though it points the same way.

More than half of businesses pay for AI. In October, 56.8% of businesses on Ramp paid for at least one AI product. A year earlier it was about 44%.

Anthropic now leads OpenAI. In October, 44.8% of businesses paid Anthropic and 40.6% paid OpenAI. Many businesses pay both, which is why the two figures add up to more than 56.8%. OpenAI led from the start of the index until April this year, when Anthropic passed it for the first time.

A small group does most of the spending. The median business that pays for AI spends $16.32 per employee per month, about the cost of one AI subscription. The top 1% spend a median of $10.2K. Ramp says that group drives 80% of spending on OpenAI and Anthropic. Its spending peaked in July, dipped through August and September, then hit a new high in October. When the heaviest buyers pause, the whole index feels it.

Usage is growing four times faster than the bill. In the four weeks to October 23, businesses used 37% more tokens and spent 9% more on them. Prices for top models are falling fast, and Kharazian attributes most of the summer’s spending slowdown to those price cuts. Demand is not falling. Each unit of it is getting cheaper.

Can the biggest spenders hold the new high? The top 1% reached $10,199 per employee in October, above July’s $9,873. Holding there means the summer was a pause. Slipping back means the heaviest buyers are near their ceiling.

Does usage keep outrunning price cuts? Right now it does, so bills still rise. The week prices fall faster than usage grows, AI revenue forecasts get harder to defend.

Who wins the businesses that pay nothing yet? About 43% of firms on Ramp still pay for no AI. They are the next wave of customers, and the receipts will show which provider they choose.

I will be reading this index every month from here. If you invest in AI, you should too.

Bashar,
Managing Member of UpRound

Have a question about anything in this issue? Reply to this email. It comes straight to us.

Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.

It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.

Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.

Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.

Try Wispr Flow free