AI spend benchmark
Four ratios, from numbers you already have. What share of revenue AI takes, what share of R&D, what it costs per employee, and whether the line is growing faster than the business is.
AI spend as a share of revenue
4.0%
On these growth rates it is 5.5% next year, up 1.5 points.
- Share of R&D spend
- 13.3%
- AI spend per employee
- $6,000 a year
- Next year's AI spend
- $1,080,000
- Verdict
- Growing faster than revenue
Your own numbers, not a peer median. No credible public benchmark for AI spend exists yet, so this page does not quote one.
What is a reasonable amount to spend on AI?
There is no honest peer benchmark to answer that with yet. The public figures are self-reported, a year old, and rarely say whether they count model APIs alone or seats and GPU compute as well, so a board decision made against one of them is a decision made against noise. The four ratios that hold up are internal: what share of revenue AI takes, what share of R&D, what it costs per employee, and whether it is growing faster than the business. The last one is the question a board actually asks, because a cost growing at twice the rate of revenue stops being a cost line and becomes a margin question.
The four numbers, and what each one is for.
- Share of revenue
- The headline. One line, absorbed in three seconds, and the figure that gets compared year over year forever once you have said it once. Say it before somebody derives it.
- Share of R&D
- The context. It answers whether AI is a new budget or a reallocation of one that already existed, which is usually the first challenge from a technical board member.
- Cost per employee
- The sanity check. It travels well next to the other per-head numbers a board already knows, and it makes per-seat AI tools visible next to model spend.
- The trajectory
- The one that matters. A share of revenue is a snapshot. Whether it rises or falls next year is the decision, and it is the only one of the four that implies an action.
A ratio you cannot break down will be broken down for you.
Every question that follows the headline number is an attribution question, and a single total cannot answer any of them.
About benchmarking AI spend
What is a normal amount to spend on AI?
Nobody credibly knows yet, and this page will not pretend otherwise. The public survey data is thin, self-reported, a year behind, and rarely defines whether it counts model APIs only or seats and GPU compute too. A number quoted against that is worse than no number, because a board will act on it.
So this tool benchmarks you against the four things you can actually verify: your own revenue, your R&D budget, your headcount and your growth rate. Those are the comparisons a board asks for anyway.
Which of the four ratios matters most?
The trajectory, every time. A share of revenue is a snapshot and boards absorb it in one line. What they react to is whether the line is growing faster than the business, because that is what turns a cost into a margin question.
If your AI spend is growing at twice your revenue growth, the ratio doubles roughly every three years, and someone will eventually ask what the ceiling is. Better to have the answer before the question.
Should AI spend sit in R&D or in cost of goods sold?
It depends on what the spend buys, and most companies have both. Model calls that serve a customer-facing feature belong in cost of goods sold, because they scale with usage and they sit directly against gross margin. Model calls that build the product, plus the AI tools your engineers hold seats for, belong in R&D.
Companies that never split the two report one number that means nothing, and discover the gross margin question a year late. Splitting it is the single most useful thing a finance team can do with an AI bill.
What should we include in the AI spend figure?
Everything you pay because of AI: model provider APIs, models running inside your cloud bill through Bedrock or Vertex or Azure, per-seat AI tools, and the GPU compute behind anything you host yourself. Most companies have four categories and count one, which is why the total on this page is usually a surprise.
Spend visibility covers how each one gets found.
The number is higher than we expected. Is that bad?
Not by itself. A high share of revenue attached to a feature customers pay for is an investment working. The same share attached to spend nobody can allocate to a team or a product is the problem, and the difference between the two is attribution, not size.
The useful next question is never how much, it is what for, and which of it could buy the same output for less.