Most enterprise AI guesses because it has nothing to ground against. Atlas is what it reads from instead — canonical metrics, glossary terms, entity relationships, and the permission rules Guardian protects. Written down once, owned by the people who actually know.
Defined once · Owned by your team · Read on every request
Every business already has these definitions somewhere. They sit in spreadsheets, in people's heads, in a wiki page nobody has opened in a year. Atlas is where they finally live in one place, kept by the people who own them, where the AI can read them.
Verified formulas live in Atlas, and the AI uses them rather than guessing from column names. When finance changes the formula, the answer to "what's our gross margin?" changes everywhere at once.
The AI knows what churn means in your business, and whether customer, account, and buyer name the same thing or three different ones. Entity relationships are written down, so the join is something Atlas already knows rather than something the model invents.
Atlas defines field-level visibility: who can read which columns, under what conditions, with what masking. Guardian protects it on every request. The AI cannot hand back what the user isn't cleared to see, and a summary or an inferred figure doesn't slip past it either.
Analysts edit the semantic layer directly in Context Studio: glossary terms, metric definitions, and playbook content, alongside browsing and scoping the model. The definitions live in one place, and every answer reads from them.
Two responsibilities, two components. Atlas writes down what your business means and who can see what. Guardian checks each call against those definitions as it happens, then logs the decision.
Metric formulas, glossary terms, entity edges, and permission semantics. Analysts and data governance edit them on purpose, then version and own them. Defining something is a deliberate act, done once and reused, not redone for every call.
For each call, Guardian reads who is asking, looks up what Atlas says they're allowed to see, and checks the model's answer against the canonical definitions. It runs constantly, and it runs the same way every time.
Mixing definition and protection is what produces shadow AI: rules that live nowhere and get applied differently each time. Keep them apart and the same question starts giving the same answer.
The user gets an answer or a refusal, with a citation, replayable end to end. Atlas decides what is true and who may see it. Guardian makes sure that is what actually happens.
Atlas is a layer of meaning over your data, not a copy of it. The boundary is deliberate. Definitions live here; the data itself stays exactly where it already is.
Atlas sits at the end of the path. The user sees one thread. Underneath, the platform sees a question, a policy decision, and an answer grounded in Atlas rather than guessed at.
Studio is the surface where the user sits. Guardian is the gate the call passes through, with the answer and the decision both kept on the record. Atlas is the layer underneath, the meaning the AI reads instead of guessing. No one of the three does much on its own. See the platform architecture → for the full diagram.