What does Ramp’s token spend really say about AI at work?

If you only read vendor keynotes, AI adoption sounds like a wave that lifts every desk. Ramp’s AI Index — built from card and bill-pay data across more than 70,000 US businesses — is blunter. It measures who pays, and for tokens, which maker gets the money.

For the week of 21–27 September 2026, Ramp’s token-spend leaderboard (maker view, volume mode) is almost a closed shop: Anthropic 51.0% of share (index 75.7) and OpenAI 44.5% (index 66.1). Together that is roughly 95.5% of attributed maker spend in the panel. xAI sits at 2.5%, Google at 1.6%. Everyone else — Cursor, Moonshot, DeepSeek, Meta — is a rounding error in this window.

Adoption is no longer a yes/no question

Ramp’s own research arc matters here. Early letters tracked the share of firms with any paid AI transaction. By late 2025 that line had climbed toward half of businesses on the platform — 46.6% in the December reading that also showed Census survey figures lagging badly. By mid-2026 the lab was explicit: binary adoption is approaching saturation among paying firms, so the useful metric is intensity — spend per employee, subscriptions versus agents versus tokens.

That is the real-world picture. Lots of companies have a receipt. Far fewer have rewritten how work moves.

What the duopoly actually means

A 95% two-maker share does not mean “everyone uses Claude and ChatGPT the same way”. It means that, among businesses routing metered API/token spend through Ramp’s Token Spend Management sample, the economic centre of gravity is still frontier US labs. Open-source and Chinese makers show up in the long tail; Ramp’s September letter put open-source routing at about 6.4% of AI-spending firms and 3.6% of firms overall — rising, but not yet the plot twist investors keep waiting for.

Two caveats, both from Ramp’s methodology notes. First, free tools and personal accounts never appear as corporate transactions, so paid adoption understates true workplace use. Second, the token board is a TSM sample — directionally similar to broader AI spend, not a census of every dollar.

What to take into Monday

• Treat “we use AI” as table stakes. Ask for tokens, seats, and which workflows actually changed.
• Expect model choice debates to stay loud while the invoice remains a duopoly.
• Read underestimation as a risk signal: what finance cannot see, security and compliance often cannot govern.

Ramp is not a press release. It is a cash register. And the cash register says American knowledge work is paying two labs for most of its metered intelligence — while the harder transformation work is still unevenly done.

Sources: Ramp AI Index (token spend) · January 2026 adoption letter · September 2026 letter