Introducing the Arion Enterprise AI Atlas: a map with a method
Every few weeks another enterprise AI landscape makes the rounds. Hundreds of logos in tidy boxes, arranged by category, impressive at a glance and close to useless the moment you try to make a decision with it. The maps rarely tell you the one thing that matters most: whether a product is an AI company or an established platform that added an AI feature. Those are two different purchases, with two different risk profiles, and a wall of logos treats them as the same thing.
We wanted a map we could actually use in advisory work. So we built one, and today we are opening it to everyone. The Arion Enterprise AI Atlas is a living, sourced map of the enterprise AI market, and every product on it is classified by a single published method. You can explore it now at enterpriseaiatlas.ai.
The distinction that runs through everything: Native or Embedded
The Atlas sorts the market with one structural question. Is the AI the product, or is the AI inside a product that already existed?
AI Native solutions depend on their models. The model is not a feature; it is the reason the product exists. Turn the models off and the product stops working or loses its reason to be. Native solutions run the length of the stack, from compute and foundation models through agent platforms to AI-first applications.
AI Embedded solutions are established enterprise products that added AI. The host platform predates its AI capability and runs without it. The AI augments workflows, data, and users that were already there. A copilot inside a CRM, agents on a workflow platform, generative features in a collaboration suite: all embedded.
This is not a cosmetic label. It changes how you buy. Adopting your incumbent's embedded AI means less integration, familiar data and governance, and value that rides infrastructure you already own. Adopting a native product means buying a new capability on its own merits, with its own security review, its own data path, and its own place in your architecture. Neither is better in the abstract. But knowing which one you are looking at is where a sound decision starts.
There is a third term worth naming, because it is where most enterprise value actually gets realized: AI Enhanced, the deployments you build by composing native tools with your existing systems. The Atlas does not map it as a panel, because it is something buyers and integrators assemble rather than a population of products to catalog. But it is the reason the map matters. The point of knowing what is native and what is embedded is to combine them well.
How every product is classified
Positioning is not classification. Plenty of products market themselves as AI-first; the method ignores the marketing and looks at the architecture. Each product is run through five tests, applied in order.
The decisive one is dependency: does the product deliver its core value if you remove the models? If no, it is native. If yes, it is embedded. When that answer is clear, it settles the classification. When it is genuinely ambiguous, four supporting tests break the tie by weight of evidence: architecture (do models sit in the primary execution path or an assistive one), origin (was the product built around models or did AI arrive later), commercial (is AI the headline you pay for or an add-on to an existing license), and interface (is the main interaction model-mediated or a conventional UI with AI assists). Origin and interface are signals, not verdicts, which keeps the method from lazily tagging every startup native and every incumbent embedded.
On top of the classification, every product carries an agentic level from 0 to 3: none, assistive, agentic, and autonomous. Level is an attribute, not a category, so it cuts across the whole map. A product earns the agentic tag at Level 2, where it plans and executes multi-step work with human checkpoints, and above.
Two design choices keep the map honest. The unit is the product, not the vendor, so a large vendor appears many times and can sit on both panels, once for its AI-first platform and again for the copilot inside its suite. And every placement carries a written rationale, a confidence rating, and at least two cited sources, so any entry can be checked rather than taken on faith.
What is on the map today
As of this week the Atlas holds 348 products from 300 vendors, split 195 native and 153 embedded, across 46 category cells on two panels, backed by 663 cited sources. It is curated rather than exhaustive on purpose. The value is judgment, not census, so a product earns a place only when it is enterprise-grade, generally available, showing real traction, and actively shipping. Consulting firms, hardware-only offerings, and products still in stealth are out of scope.
The map is revised continuously as the market moves, not reprinted once a year. New products, agents, and acquisitions are classified as they ship, and what changed is published rather than quietly edited.
Built to be audited, not admired
The reason to trust a map is that you can argue with it. The full method is published, so any classification can be tested against the same criteria everyone else is held to. Vendors who believe a product is placed wrong can challenge it with evidence against the five tests, and the decision and its reasoning are recorded. This is the same standard we hold ourselves to in advisory work: analyst-grade research, human-centric, with no technology agenda of our own.
There is also a companion; Atlas Horizon. Atlas Horizon is a watchlist of the emerging vendors we are tracking toward the map, the products that do not yet clear the inclusion bar but are worth watching.
Explore it
The Atlas is open to browse today at enterpriseaiatlas.ai. Filter by panel, category, agentic level, and vendor, open any product to see how it was classified and what sources back the call, and tell us what we missed. If you are trying to turn an AI strategy into a deployed digital workforce, this is the map we use to help clients do exactly that, and now it is yours to use too.
Frequently asked questions
What is the difference between native and embedded AI?
Native AI means the product's core value depends on its models. Remove the models and there is no product, such as a foundation model or an AI coding agent. Embedded AI means an established platform that predates the AI, and runs without it, has AI added on top, such as a CRM or analytics suite with a copilot. The deciding question is dependency: does the product still deliver its core value with the AI switched off?
How does the Atlas classify each product?
Every product is run through five tests in order. Dependency is decisive: does the product deliver its core value without the models? The other four, architecture, origin, commercial model, and interface, resolve ambiguous cases by weight of evidence. Each product also gets an agentic level from 0 to 3 and at least two cited sources, and every placement records a written rationale and a confidence rating.
How many products does the Atlas cover?
As of August 18, 2026 the Atlas covers 348 products from 300 vendors, split 195 native and 153 embedded, across 46 category cells, backed by 663 cited sources. It is curated rather than exhaustive, and revised continuously as the market moves.
Can a vendor challenge how a product is classified?
Yes. Placement is editorial and method-driven, never purchased. A vendor that believes a product is classified incorrectly can submit evidence against the five tests, and the decision and its reasoning are recorded. Classifications are re-verified before each edition and when major events like acquisitions or re-architectures occur.
