Your AI Budget Hit a Ceiling. 12% of What's Under It Is Waste.

A controller at AngelList read a weekly briefing, forwarded it to engineering, and found the company had been losing $10,000 a month to a setting nobody in finance had heard of.

“Ramp’s Token Spend Management briefing surfaced ‘prompt caching,’ not something on my radar as a Controller. I routed it to engineering immediately and we found we’d been losing $10,000 a month.”

Prompt caching is the discount you get for not re-sending the same instructions every time. If your AI tool sends the same brand guidelines, the same product list and the same tone rules with every single request, you pay full freight on all of it, every time. Turn caching on and the repeated part gets billed at a fraction. It’s a checkbox, roughly, and it was worth six figures a year at one mid-sized company.

AngelList is not an outlier. Ramp, the corporate card company that sees what 1,300 businesses pay for across more than 100 trillion tokens a month, says it finds “potential savings worth 12% of monthly AI spend for the average business”.

An eighth of the average AI bill is recoverable. And it matters more this month than last, because the other half of Ramp’s data says the bill has stopped growing.

The ceiling

Ramp publishes a monthly index built from what its customers charge rather than what they tell a survey. The August edition is titled “Cracks in the AI thesis”, and the crack is this.

Anthropic released Fable 5, which Ramp describes as the best AI model to ever hit the market. After a month it accounted for 6% of Anthropic’s tokens and 11.4% of spending, losing to OpenAI’s cheaper GPT-5.6 Sol despite being the stronger model. Ramp economist Ara Kharazian drew the conclusion:

“we’ve found a new upper bound for how much businesses are willing to spend on AI.”

The rest of the index says the same thing more quietly. Anthropic sits at 43.5% of businesses, up 1.1 points in the month. OpenAI is at 39.7%, up 0.23. Both grew, and adoption for both has slowed, with growth increasingly coming from companies already spending rather than new ones starting. Meanwhile 6.1% of AI-using businesses now route through model-serving platforms, the cheap tier that hosts open-weight models anyone can download and run.

Better model, more money, was the assumption the last two years ran on. Businesses just declined the upgrade. We argued the quality half of this in July, when Anthropic’s benchmark jump didn’t reach the marketing copy. This is the budget half, with payment records behind it.

A correction worth making, since we nearly repeated it. Several outlets covered this data as OpenAI closing the gap on Anthropic, and TechCrunch’s headline says exactly that. On the month Ramp published, Anthropic extended its lead: +1.1 points against +0.23. The horse race isn’t the story and the direction in the coverage runs the wrong way.

What a fixed budget does to a marketing team

If your AI spend is capped and your usage keeps climbing, the only lever left is efficiency. That used to be a platform team’s problem. Now it’s a line item on your budget with your name next to it.

Three of Ramp’s examples are things a marketing team does without thinking:

A single user caused a $1,600 weekend spend spike by switching to a new model. One team spent more than $3,000 on Anthropic’s Opus, the expensive tier. And turning on a faster processing mode produced a 6x cost increase in a single week.

None of those are mistakes exactly. They’re someone picking the better option in a dropdown, which is what you would want them to do, in a system where the better option costs six times more and nothing on screen says so.

That’s the shape of the problem. Marketing teams were handed tools priced by consumption and interfaces designed to hide it. It’s the same blind spot we found in the gap between AI cost visibility and reported ROI, one layer further down.

The part where you check who’s selling

Ramp sells the software that finds the 12%. The number is a marketing claim from the company whose product captures it, and it should be read that way.

Two things make it worth using anyway. The measurement comes from payment records rather than a survey, so it’s counting money that moved. And Ramp discloses its own limits, which most coverage dropped: the index notes its sample “skews slightly more tech-y than our typical AI Index sample,” and that actual Fable adoption “is likely even lower than what we have estimated here.” A vendor volunteering that its headline number might understate its own case is doing something unusual.

Treat the 12% as directional, not as your number. Your number is whatever you find when you look.

What to do about it

Nothing here requires buying anything.

1. Find out what you spent on AI last month. Not the seat licenses. The consumption: the API bills, the credits, the per-task charges buried in tools you already pay for. Most marketing teams cannot produce this figure in under a day, which is the finding.
2. Ask whoever wired up your AI tools one question: is prompt caching on? If your team runs the same brief, brand rules or product catalog through a model repeatedly, this is the AngelList $10,000.
3. Check which model your tools default to, and whether anyone chose it. A default set at signup twelve months ago is a price decision nobody made.
4. Put a number on one workflow before you scale it. Cost per output, at current settings. Without it you’re scaling an unknown unit cost, and the ceiling in Ramp’s data says nobody is going to raise your budget to cover the surprise.

The last two years of AI budgeting assumed the line goes up and the capability justifies it. Payment data now says buyers stopped believing the second half. The teams that keep their AI programs through the next budget cycle will be the ones who can show what a unit of output costs and prove it went down. At monday.com that budget moved from salaries into the compute bill, and the CFO said so on the record.