Is Business AI Spending Flat? What Ramp's Data Shows
Ramp data shows AI spend up just 1.65% in October, 56.8% adoption, and frontier models near 70% of dollars. Here is why it matters for buyers and AI vendors.
For Executives, Finance and Procurement

Are businesses spending more on AI?
Not much more, at least in the dollars Ramp tracks. Ramp's November 11th update says AI spending rose 1.65% in October, which is close to flat. Its October report described almost no growth in token spend across August and September, so spending has been essentially flat for three months. Over the same period, token volumes (the units of text models process) kept rising. Businesses are using more AI and paying about the same.
TL;DR
- Spending is flat while usage grows. Ramp reports 1.65% growth in October after almost none in August and September, even as token volumes rose month over month.
- Adoption is still partial. Ramp's sample put adoption at 56.8% of businesses, so plenty of companies have not started.
- Cheaper models explain much of the gap. Companies move routine work to lower-priced models, so 5x the tokens can cost the same money.
- Frontier models still take most of the dollars. Ramp's November report puts them at almost 70% of spending, with the share rising slightly in October.
- Support teams are the new growth area. Customer success and support went from under 1% of classified spending in June to 13% in October.
- Buyers have negotiating power, and so will their customers. Infosys said on its October earnings call that clients want AI productivity gains passed through in pricing.
Why can token use rise while spend stays flat?
Because the price per token is not fixed, and buyers are choosing it. A token is a small chunk of text a model reads or writes. Vendors bill by the million tokens, and the price varies a lot between a top model and a cheaper one.
Here is the arithmetic one analyst used to illustrate it:
| Scenario | Average price per million tokens | Tokens used | Monthly bill |
|---|---|---|---|
| Before | $5 | 1 billion | $5,000 |
| After moving routine work to cheaper models | $1 | 5 billion | $5,000 |
Token consumption can grow fivefold while spending does not increase at all, or even goes down. The business got far more AI work done. The model maker got no extra revenue from it. This does not even require switching vendors. A company can move down to a cheaper model from the same lab.
These are illustrative numbers, not Ramp's. But they show how a usage chart and a spending chart can tell different stories.
What is task saturation?
Task saturation means several models can already do a given job well enough that paying for more intelligence adds little or nothing to the result. Models keep improving. The work a company asks of them often does not need the improvement.
Think of a strategy document. You would expect to want the strongest reasoning available. But people who have spent hundreds of dollars in API costs on frontier models for that kind of task have sometimes looked at the output and wondered whether the bill was justified. The model was impressive. It just did not improve the decision enough to matter.
This is not a loss of interest in AI. It is buyers figuring out which AI they are willing to pay for.
How are companies deciding which model to buy?
Increasingly, with evals. An eval is a test of the work you actually need done, judged against your own standard for an acceptable result. Instead of assuming the newest model is a required upgrade, a team runs its real tasks through a cheaper model and checks the output.
The pattern reported from enterprise conversations: pick a relatively cheap model, test it, and iterate toward the cheapest model that passes. When a new model arrives, run the same tests again.
This gives procurement evidence instead of a hunch. If a cheaper model passes the tests for the work being bought, it is a viable substitute. It does not need to be equally smart in every situation. It only needs to meet your requirements. That shifts negotiating power toward the buyer.
Ramp's October report also describes customers negotiating enterprise contracts from favorable positions, based on accounts from businesses that had signed or reviewed those deals. So even a company that stays with one provider can use eval results to negotiate the price down.
Where is AI spending actually growing?
Support. Ramp's latest report identifies customer success and support as a growing source of spending, going from less than 1% of classified spending in June to 13% in October.
One reading is that these use cases have become easy for ordinary companies to adopt. The best practices are known, and the work is easy to understand. Businesses a little behind the leading edge feel confident going after them.
Support also shows how a mixed-model setup works. A customer wants an accurate answer fast, and the business wants reliable resolution at acceptable cost. A cheaper model can handle most of the volume, while a stronger model handles the hard cases. A router, software that sends each request to the right model, does the sorting, and routers are now commonly available.
Coding follows the same logic. Arguing over which model is smarter goes nowhere. Evals give that debate evidence. Does the model produce the right change? Does it work in the codebase? How much human review and repair does it need? If a cheaper model meets the standard at lower total cost, use it. If it fails, the failure shows whether the expensive option is worth it, or whether two or three iterations on the cheap model still comes out cheaper.
Do frontier models still matter?
Yes. Ramp's November report says frontier models accounted for almost 70% of spending, with their share rising slightly in October. Some tasks demand the extra capability and justify the cost.
But a big share of dollars is not a big share of work. As more tasks saturate, more of the total work runs on cheaper models, even if the premium tier still collects most of the money.
Lower prices can also create new demand. A task that was not worth doing at $1 may be worth doing at 10 cents. This is the old Jevons idea: cheaper resources can increase total consumption. Token demand can keep growing even if revenue per token falls. The open question is whether usage grows fast enough, or provider costs fall fast enough, to keep the economics healthy for the labs.
What does the Infosys comment add?
It shows the same pressure from the other side. On its October earnings call, Infosys described AI revenue growing very quickly while customers demanded that AI productivity benefits be passed through in pricing. The source transcript cuts off before the full discussion of its margins, so I will not characterize the financial result.
The point is simple. If you buy AI and get cheaper work done, your own customers will expect to share in the savings. Any company selling services built on AI should expect the same question from its clients that it asks of its model vendors.
What does this mean for you?
If you buy AI for your company: you have negotiating power. You can get more capable tools while being stricter about what you pay for.
- List your recurring AI tasks (support replies, summaries, code review, drafting).
- For each, collect 20 to 50 real examples and write down what "good enough" looks like.
- Run them through a cheaper model and compare against your current one.
- Move the task to the cheapest model that passes. Re-test when new models ship.
- Use the results in contract talks, even if you keep the same vendor.
If you are a manager or team lead: expect questions about which model your team uses. Come with test results, not opinions.
If you sell services that use AI: expect clients to ask for a share of the productivity gain. Plan pricing around outcomes you deliver, not the hours or tokens saved.
If you invest in or work for an AI company: better models and rising usage are not enough on their own. Research, model development and infrastructure still have to be paid for, so watch revenue, not token charts.
Keep the limits in mind. This is one company's sample (Ramp's, at 56.8% adoption), three months of data, and an interpretation of it. Flat spend does not mean spend will never grow.
Can you build your own AI spend and eval tracker?
You can, and a small internal tool is often more useful than another subscription. A simple app could log each AI workflow, the model it runs on, the monthly token cost, and the pass or fail result of your test cases, so you can see where a cheaper model would do. Remy builds full-stack business apps from a plain-English description, with a database and sign-in included, so a finance or operations lead can describe that tracker and get a working version their team can log into. You can then adjust it as your evals change.
Frequently Asked Questions
What did Ramp's data say about AI spending in October?
Ramp's November 11th update said AI spending rose 1.65% in October, essentially flat. Adoption reached 56.8% of businesses in its sample. Its October report had already described almost no token-spend growth in August and September.
Why would companies use more AI but spend the same?
They move routine work to cheaper models. In the illustrative example above, dropping the average price from $5 to $1 per million tokens lets a company process five times the tokens for the same $5,000.
Are frontier models losing out?
Not in dollars. Ramp's November report says frontier models made up almost 70% of spending, with their share rising slightly in October. The change is that many everyday tasks no longer need them.
What is an eval, and do I need one?
An eval is a test built from your own real work, judged against your own standard for acceptable results. If you are choosing between models or negotiating a price, it is the most direct way to show whether a cheaper option is good enough.
Why does the Infosys comment matter?
Infosys said customers wanted AI productivity benefits passed through in pricing while its AI revenue grew rapidly. It shows that the price pressure on model makers also reaches the companies that build on them.
How we source, write, and correct our articles. Editorial standards
Related articles

How to Review AI Coding Tool Usage Before Renewal

Is the AI Bubble Real? Nvidia's Avalanche Worry Explained

Do You Need the Most Expensive AI Model? Task Saturation
