The Ramp Benchmarks Tell You Where Your AI Money Went
You can tell where your AI money went. Do you know what it bought?
Ramp published something worth reading this year. Using real payment data from thousands of businesses, they mapped what AI token spend actually looks like by company size, model tier, and usage depth. The median monthly spend is $2,246. The 75th percentile is $14,843. The 90th percentile is $73,030. Token usage grew 1,001% between January 2025 and April 2026.
These are real numbers. Actual invoice data. But what happens when we account for token waste?
What the benchmarks look like at 40% to 60% token waste
The previous two posts in this series argued that token waste is a behavior problem, not an infrastructure problem. Unenabled teams regen outputs. They dump context. They don’t decompose tasks. They route everything to the frontier model because nobody told them not to.
When GitHub’s own engineering team audited their production agentic workflows, they found enough waste to cut effective token costs by 37% to 62% across their most-used automations, without changing what the workflows produced. Here’s what Ramp’s benchmarks look like when you apply that range.
At 40% waste:
| Spend tier | Monthly total | Wasted | Productive |
|---|---|---|---|
| Median | $2,246 | $898 | $1,348 |
| 75th percentile | $14,843 | $5,937 | $8,906 |
| 90th percentile | $73,030 | $29,212 | $43,818 |
At 60% waste:
| Spend tier | Monthly total | Wasted | Productive |
|---|---|---|---|
| Median | $2,246 | $1,348 | $898 |
| 75th percentile | $14,843 | $8,906 | $5,937 |
| 90th percentile | $73,030 | $43,818 | $29,212 |
The organization at the 90th percentile spending $73,030 a month may be getting $29,212 in productive token value. The other $43,818 is regen loops, context dumps, and unconstrained agent calls.
Annualize that. At the 90th percentile with 60% waste, you’re leaving roughly $525,000 a year on the table. Not in efficiency gains you haven’t captured yet, but in spend that produced nothing.
What ROI does that actually require?
If your waste rate is 40%, your productive tokens need to generate a return on your entire bill, not just the portion that worked. If you’re targeting a 3x ROI on your AI investment, the productive portion has to carry 5x to cover the waste. At 60% waste, that number is 7.5x.
| Waste rate | Target ROI | Required return from productive spend |
|---|---|---|
| 0% (no waste) | 3x | 3x |
| 40% waste | 3x | 5x |
| 60% waste | 3x | 7.5x |
Most AI business cases are built against the nominal spend number. They’re not accounting for waste. That means the projects that look viable on paper may require returns that were never realistic, and the ones that looked marginal may already be underwater. This is the math that isn’t in the Ramp report. It isn’t a criticism. It’s a different question.
Ramp tells you where the money went, not what it bought
Ramp’s product gives you spend by vendor, by model, by team, and by month. That’s spend visibility. It’s necessary. And it stops short of the question that actually determines whether your AI investments are working: what did those tokens produce?
Not operationally in aggregate. Specifically. Did the workflow that ran 4.2 million tokens last month change how fast your team delivers? Did it eliminate a review step? Did it move a revenue metric? Did it reduce rework, compress cycle time, or produce a measurable output that didn’t exist before? Was the automation or intelligence effective, or just expensive?
Without answers to those questions, you have an invoice. You don’t have an investment thesis.
This is the gap Clairet is built to close. Not spend tracking: outcome tracking. Every AI initiative mapped to a baseline, an owner, and a success metric before it goes live. Every project measured against the results it promised, not the tokens it consumed. Knowing where your AI money went is step one. Knowing what it returned, operationally and financially, is what makes the next investment decision a decision, not a guess.
Hype is not a strategy. Find insight and direction with Clairet. Request a demo.
Sources: Ramp, “How Much Do AI Tokens Cost Businesses? 2026 Spending Benchmarks,” June 8, 2026 (April 2026 data); GitHub, “Improving token efficiency in GitHub Agentic Workflows,” May 7, 2026.