Menlo Ventures, a venture capital firm, published its fourth State of Enterprise AI report on 3 November. It is the most detailed public estimate of where business money for AI is going this year. The headline is size: companies are now spending at a rate of $176.8 billion a year on using AI, 5.4 times the figure a year earlier. The more useful part for a small builder sits further down, in the table that splits that money by job.
That table shows the money piling into tools that do one well-defined job. Coding alone takes $34 billion. Marketing tools take $1.2 billion, sales $1.4 billion and creator tools $1.5 billion. Those three are small, young categories, and that is where an indie builder still has room.
One disclosure before the numbers. Menlo says it is an investor in Anthropic and OpenRouter, both of which appear in the report. Read the vendor figures with that in mind. The category figures below do not depend on any single vendor.
The data behind the trend
First, what the big number means. Menlo counts "deployment" spend: money companies pay to use AI, as opposed to money spent training models. The figure is an annual run rate in September, meaning the monthly spend multiplied out to a year. Campus Technology points out that it should not be read as money already spent this year. Menlo built the totals from a survey of nearly 400 US enterprise buyers run by Morning Consult in September, plus its own model of public filings and other reports. So these are estimates, and they describe large US companies. They still tell you which jobs buyers are willing to pay for.
Second, the number of buyers barely moved while the bill grew. The share of enterprises paying for AI rose from 55% to 65%. Spending per company grew much faster: the average company spent 4.7 times what it spent last year, against a plan of 1.9 times. The median company, the one in the middle, spent 2.2 times. A small group of heavy users pulled the average up.
Third, the split by job. Of the $77.3 billion spent on AI applications, coding takes $34.0 billion and workplace chatbots $25.0 billion, together 76%. Everything else is small:
Category | 2026 spend | Growth on 2025 | Share held by AI-native startups |
|---|---|---|---|
Coding | $34.0B | 7.7x | 96% |
Customer success | $2.7B | 4.3x | 42% |
Creators | $1.5B | 5.0x | not split out |
Sales | $1.4B | 3.6x | 54% |
Marketing | $1.2B | 1.8x | 46% |
All figures in the table come from the "Coding Leads AI Application Spend" and "By Category" charts in Menlo's report.
Fourth, who gets the money. AI-native companies, meaning firms built around AI from the start, now take 77% of application spend, up from 63% a year ago. In coding they take 96%. That includes the AI labs' own coding tools, but Menlo also reports that the independent coding companies grew 2.9 times to $10.1 billion. Large labs and small specialists grew in the same market at the same time.
Fifth, who is doing the work. On OpenRouter, a service that lets developers reach hundreds of AI models through one account, agents now produce 80% of all tokens, about seven times the usage that comes from people typing. A token is a small chunk of text the model reads or writes, and it is how AI usage is billed. An agent is software that takes a goal and carries out several steps on its own. Menlo notes that OpenRouter's users lean towards individual developers and startups, and Campus Technology adds that it measures one platform, not all AI use. Treat it as a sign of direction.
Why it is happening
Menlo's explanation is simple. Money goes where agents can do a job from start to finish and someone can check the result. The report names three conditions: a place where the agent can act, complete context it can learn from, and results it can verify.
Coding meets all three. The code sits in a repository the agent can edit. The project's files and documentation tell it how things are done. Tests tell it, and the buyer, whether the work is right. When a buyer can see that the work is right, they keep paying, and they pay more as the agent takes on more of it. That is why spend per company grew faster than the number of companies buying.
Customer support meets most of the conditions too. Support playbooks write the steps down and a closed ticket is a clear result, and the category grew 4.3 times to $2.7 billion. Menlo says marketing grew just 1.8 times to $1.2 billion because its conditions are harder to meet. "Write better marketing" has no test suite. Context lives in people's heads, and results show up weeks later, mixed with everything else the business did.
There is a second reason the money is narrow. Menlo found that agent spend grows with the amount of work done, so it is no longer capped by how many staff use the tool. A tool that finishes one task many times a day earns more than a general assistant opened twice a week. Buyers can also justify a narrow tool more easily, because they can point at the job it does.
What it means for indie builders, creators and small businesses
For indie builders, the opening is in the small categories. Marketing, sales and creator tools together come to about $4.1 billion, by our sum of Menlo's figures. That is small next to coding, and the largest companies have put their effort where the money already is. In marketing, established software companies still hold 54% of spend. Sales is close to an even split, with startups on 54%. Neither category has a settled winner. The report's own reasoning also tells you why these categories are small today: the work is hard to check. If you make one marketing or sales job checkable, you remove the reason buyers hold back.
For creators, the creator category is the fastest-growing of these three, up five times on last year according to Menlo's table. The named companies are video, image and music generators. A creator who builds a narrow tool for other creators, around a job they already do every week, is selling into a category where buyers have shown they will pay.
For small businesses, the lesson is about what to buy. The tools winning enterprise budgets do one job you can measure. Before paying for a general AI subscription, ask which single task in your week has a clear "done" state, and look for a tool that does exactly that.
On money, an estimate. A narrow tool that bills per job, rather than per seat, follows the way enterprise spend is moving. We covered the pricing side in Unlimited seats, pay per outcome.
On risk. The big numbers come from a venture firm that invests in several of the companies it ranks, and they describe large US companies. The categories could change fast: Menlo itself says incumbents gained ground in nearly half the categories it tracks this year. Treat the table as a map of where buyers are paying today, and test demand with your own customers before you build.
A worked example: one narrow marketing job, priced per run
Take a hypothetical solo builder, Priya, who knows small online shops. She wants to build something in the marketing category. A general "AI marketing assistant" fails all three of Menlo's conditions, so she picks one job: rewriting product listings on Shopify stores that have more than 100 products and weak descriptions.
She checks the job against the three conditions. The agent can act: it reads and writes listings through the Shopify connection the store owner approves. The context is complete: the store's best-selling listings, brand notes and a list of banned claims (no medical promises, no fake scarcity) go into a short rules file. The result can be checked twice. First, every rewritten listing must pass a checklist before it is saved: length, the product's main search phrase in the first sentence, no banned words. Second, after four weeks, she compares click-through and add-to-cart rates on rewritten listings against a set of listings she left alone. That comparison is the proof she sells with.
Now the costs, all estimates. A batch of 200 listings might use around 400,000 input tokens and 200,000 output tokens, including the checklist pass. At the $4 per million input and $20 per million output tokens that Menlo lists for Claude Opus 5.5, that is about $5.60 per batch. A cheaper model would cost less. Hosting and the Shopify app listing add a small fixed cost each month.
Pricing, also an estimate: $49 a month for up to 200 listing rewrites, then $0.20 per extra listing. A store owner compares that with paying a copywriter, which is easy to do because the job is the same job. Twenty stores bring in $980 a month at a model cost of roughly $112. The time cost is the real investment: perhaps two or three weekends to build the first version, and a month of running it free for five stores to collect before-and-after numbers.
Priya does not build a second feature until the first one has a proof number. Every extra job dilutes the one thing her customers can check. This is the same idea as building one agent per job and running it across every client.
What to do about it
Use this decision rule before you build anything in marketing, sales or creator tools. Write the job in one sentence. Then answer three questions:
Can the tool act directly in a system your customer already uses (their store, inbox, CRM, editing app or scheduler)?
Can the context it needs fit in a short written file, such as examples, rules and banned phrases?
Can you or the customer check the result with a rule on the day, and with a number within four weeks?
Three yeses: build it. Two yeses: narrow the job until the missing one becomes a yes. One or none: pick a different job.
Then follow these steps:
List five jobs you already do for clients or your own business every week. Run each through the three questions and keep the best one.
Write the pass/fail checklist first. If you cannot write it, the job is not narrow enough.
Run it by hand or with a simple script for three customers, free, for four weeks. Keep a control group of work you did not touch, so you have a comparison to show.
Price per job done, with a monthly floor. It matches how buyers think about the work, and your price grows as the tool does more of it.
Sell with the number. "Rewritten listings got X% more add-to-carts than untouched ones" is the claim a small business owner can repeat to themselves when the invoice arrives.
One last thing on distribution. Being the tool that does one job also makes you easier to find, in search and in AI answers. Our earlier trend piece on where creators should put their search effort in 2026 covers that side.

