What Ramp reported

On 6 November 2026, TechCrunch reported new data from Ramp, the corporate card and expense company, on what its customers pay for AI. Companies are using more AI than ever, but they are not spending more on it. Token spend grew 1.5% across August and September, after growing 42% in July alone. Ramp’s lead economist, Ara Kharazian, said almost the entire spend comes from venture-backed, high-growth tech companies, and that the era of hypergrowth may be over. The story landed the same week Bloomberg reported Anthropic’s expected IPO valuation had fallen from US$2 trillion to US$1.2 to 1.3 trillion.

The claim, and what the numbers actually say

The claim in the air is that AI demand is cooling. Investors priced in hypergrowth, the spend line has gone flat, so the party must be ending.

I read the same data and see something else. Usage went up. Spend stayed flat. That only happens when each unit of work costs less. The article gives the reasons. OpenAI cut API prices in half in September. Anthropic made significant cuts too, and said in a filing that those cuts took about US$600 million off its revenue growth. And buyers got better at controlling cost: budgets, defaults that stop staff picking the expensive model, routine jobs routed to cheaper ones. October spend was already up 8.9% again.

So what changed is the price and the discipline. What did not change is the direction. Businesses that tried AI kept using it, and used more of it.

One more line in the article matters more than the growth rate. Almost all of this spend comes from venture-backed tech companies. The median firm on Ramp spends US$12 to US$15 per employee a month. If you run a shop, a clinic or a marketing team, this data is not about you. Your spend was never part of that curve.

Buyers are finding out how much tool they need

Here is what I think the flat line looks like from the inside.

A year ago the pitch was a different AI tool for each job. One to write, one to make images, one to take meeting notes, one to answer customers, and now a personal agent on top of all of those. I asked the same question each time a new one launched: do I need a different agent for each job, or is this just a pile of tools to optimise my life and work? I use Claude for almost everything, and I find one tool used well more efficient and more effective than a stack of them.

What Ramp calls cost controls describes companies arriving at the same answer at scale. A budget per team, a sensible default model, routine work sent to the cheap option. That is a company deciding how much tool it needs, instead of buying every tool that exists. To be fair to the data, it does not say how many subscriptions got cancelled. It does say the spend stopped growing while the work grew.

Measure workflows, not spend

Take a marketing team of five. They might hold a writing seat, an image seat, a research seat, a transcription seat and a general chat assistant, each on its own monthly plan. The spend looks like adoption. Ask a harder question: how many of the team’s jobs run end to end on any of them? Brief to draft to check to publish, with the campaign report at the end. If the answer is none, the stack is five ways of getting a first draft, and the people still carry the work between tools.

The unit of work is the workflow, not the prompt. One tool that a team has set up to run a whole process, with the context it needs and the checks it has to pass, does more than five tools each answering one question well. It also costs less, which is what Ramp is now measuring.

Same for a clinic. A booking enquiry on WhatsApp, the diary checked, three slots offered, the deposit taken, the reminder sent. One workflow, end to end. That is the measure I hold AeroChat to as well: a customer service workflow described once in the operator’s own words and run through to the action, with the order checked or the booking made before the customer hears back.

What to do with your own AI budget

Count your AI subscriptions and seats. Then count the workflows that run end to end on each one, from the trigger to the finished result, with a person only checking. Keep the tools that carry whole workflows. Give those tools more of the work before buying anything new. Set a budget and a default model, the way Ramp’s bigger customers now do, so nobody learns the price of a model from the invoice. And when the next tool launches, ask what it changes in your business on a normal day before you add it to the pile.

Source: TechCrunch: Ramp data shows AI spending stalling just as Anthropic nears its IPO