The Enterprise AI 100 One hundred companies the enterprise actually buys from.
Our annual list of the companies building what enterprises run on. Not the largest AI companies, not the best funded, and not a list of the most talked about. One hundred that clear a published bar, scored on a published method, with the arithmetic printed next to every name.
Seventy-five of the hundred are North American
This market is more concentrated by geography than by category.
2026 to 27 edition · compiled 19 Oct 2026
A list that prints its own arithmetic
Most lists of AI companies are a judgement presented as a fact. This one is a judgement with the working shown, so you can disagree with it precisely rather than generally.
AI-native only
A company qualifies only where AI is the core of what the customer buys. Enterprise software vendors who have layered AI on a system of record do not qualify, however good the layer. The enterprise buys the suite and receives the AI, and that is not an AI purchase. Fifteen very large companies were rejected on this test alone, and they are named in the edition.
A materiality bar, published
A company qualifies on capital or on evidence. Either it has raised a billion dollars or more, or it is publicly listed, or it publishes four or more named customers that are themselves large enterprises. A company with no published enterprise customers and no scale does not appear, whatever it is worth on paper.
Genuinely global, with one exclusion
The list is deliberately not United States only, and the edition carries companies from twelve countries across North America, Europe, Asia-Pacific, South Asia and the Middle East. China is excluded, and the reason is printed rather than implied: deployment inside Chinese enterprises cannot be verified from outside, and a list that cannot verify should not include.
The List ranks. The Index never does
These are two different instruments and we keep them apart on purpose. The List is an editorial ranking of companies against a published method. The Index is a record of verified customer claims that gives no score and no grade to anyone. Nothing a company does in the Index moves it on the List, and nothing it pays for moves it either.
GRADE, and what each letter is worth
Five dimensions, weighted, with the weights printed. Every company in the list carries its five scores next to its total, so you can see which dimension carried it and argue with that rather than with the ranking.
Gravity
Operating scale, as a published band rather than a funding number. A hyperscaler raises nothing and a Series B raises a great deal relative to itself, so the two are not comparable on capital alone.
Reach
How many industries the company actually sells into. Cross-industry, several industries, or a single vertical. Depth in one vertical is not penalised, it is scored elsewhere.
Application depth
Whether the product sits in the core workflow, is embedded in a workflow, or is an optional tool alongside the work. The further in, the harder it is to remove.
Durability
Years operating, divided by eight and capped. The lightest weight on the list, because age is evidence of survival rather than of relevance.
Enterprise fit
Whether the enterprise is the buyer, one buyer among several, or a builder inside an enterprise. Below a floor here, the company is out of the list entirely.
What GRADE is not. It is not a quality score, a buying recommendation or a measure of how well anything works. It describes position in the enterprise market, nothing more. A company can score ninety and be wrong for you, and a company at the bottom of this list can be exactly right. Use it to understand the shape of the market, then use the register to check what any of them has actually deployed.
The Enterprise AI 100, 2026 to 27
Sort any column. Filter by region, segment or ownership, or search a company, a city or one of the named customers behind them. Capital position is written as a position rather than a bare number, because a public company's figure and a private company's figure are not comparable.
| #↑ | Company↓ | Headquarters↓ | Founded↓ | Capital position↓ | GRADE↓ |
|---|
The customers under each company are large enterprises it names publicly on its own site, captured on the compilation date. They are evidence that a company sells to enterprises, not evidence that any particular deployment is running. For that the register is the instrument, and the two should not be confused.
The market is more concentrated by geography than by category
Twelve countries, five regions, and three quarters of the list in one of them. This is the finding most worth arguing with, and the one we expect to change most between editions.
Eighteen companies across five countries, weighted to the United Kingdom and Germany, and concentrated in applications rather than in models or infrastructure.
Six companies across South Asia and Asia-Pacific, and one in the Middle East. We expect this to be the fastest moving part of the list and we would rather print the number than round it up.
Not because there is nothing there. Because deployment inside Chinese enterprises cannot be verified from outside, and a list that cannot verify should not include.
What the hundred are actually selling
Segment counts across the edition. The applications layer is where the companies are, and the infrastructure layer is where the money is.
Seven companies build the models the other ninety-three depend on. That dependency is the most important structural fact in the list and the one least visible from any individual company's material. Read the applications layer knowing that almost all of it rests on a handful of model providers and a handful of clouds.
Read the full 2026 to 27 edition
Every company with its full entry, the rejected names and why, the method in full, and the findings drawn off the cohort itself. Free to read. No vendor paid to appear and no vendor saw it before publication.
What people ask about the list
No, and this is the one rule we will not move on. Nobody pays to be considered, to be included, or to be placed. We do take payment for verification in the Index, which is a separate instrument, and a company being indexed has no effect on whether it appears here or where. If that distinction ever stops being true we will stop publishing the list.
Because GRADE scores position in the enterprise market rather than prominence in the discourse. A company embedded in a core workflow across many industries will outscore a better known one that sits alongside the work. Every company carries its five dimension scores, so you can see which one did the work and decide whether you agree.
Because of the AI-native test. Fifteen very large companies were considered and rejected because what the enterprise buys from them is a system of record, a content library or hardware, with AI layered on top. They are named in the edition with the reason attached, rather than quietly left out.
No, and that is deliberate. The List is ranked on what companies do, their capability and the capital behind them. Verified deployments are the Index's business, and mixing the two would turn a record into a ranking. If you want to know what a specific company has actually deployed, that is a register question, not a list question.
A large enterprise the company names publicly on its own site. We count them, we list them, and we do not treat them as proof of anything beyond a commercial relationship the company is willing to put its name to. The depth of that relationship is a separate question and the register is where it gets answered.
Because we cannot verify deployment inside Chinese enterprises from outside, and every other company on this list is held to a standard of published, checkable evidence. Including companies we could not hold to that standard would make the list weaker, not broader. We state the exclusion in print rather than leaving a reader to notice the absence.
Annually. This is the 2026 to 27 edition, compiled in October 2026. The edition number and compilation date sit on every page, and both should travel with any number you quote from it. Acquisitions and listings that happen mid-cycle are carried into the next edition rather than patched into this one.
Write to us and we will look at the next edition against the same two bars: AI-native, and material on capital or on published enterprise customers. There is no fee, no submission product and no expedited route. The most useful thing you can do is publish your enterprise customers on your own site, because that is the evidence the materiality bar reads.
The list tells you who. The register tells you what is running.
Use the hundred to understand the shape of the market, then check what any of them has actually deployed before it goes into a shortlist.