Your business, made legible to the AI.
Most enterprise AI hallucinates because it has nothing to ground against. Jubi Atlas is what the AI reads from instead of guessing: canonical metrics, glossary terms, entity relationships, and the permission rules Guardian enforces on every request.
Three things every business already knows. Written down once.
Every enterprise has these definitions somewhere — in spreadsheets, in heads, in deprecated wiki pages. Atlas is where they live as a single source the AI reads from, with the people who own them.
One definition of revenue. Not eleven. Verified formulas live in Atlas; the AI uses them, not its best guess from column names. When finance changes the formula, the answer to "what's our gross margin?" changes everywhere at once.
finance.gross_margin_quarterly
The AI knows what churn means in your business, and whether customer, account, and buyer are the same thing or three different ones. Entity relationships are explicit, so the AI doesn't have to guess which join is right.
customer ↔ account (1:N)
"churn" = no_orders_60d
Atlas defines field-level visibility: who can read which columns, under what conditions, with what masking. Guardian enforces it on every request. The AI cannot return what the user isn't allowed to see, whether through retrieval, summary, or inference.
customer.email · pii
visible_to: account_owner
Atlas defines. Guardian enforces.
Two responsibilities, two components. Atlas writes down what your business means and who can see what. Guardian inspects every request against those definitions in real time.
request →
The split matters. Definition is a slow act — analysts editing Atlas with intent. Enforcement is per-request and high-frequency. Mixing the two is what produces shadow AI: rules that live nowhere, enforced inconsistently. See the security model for Guardian's full per-request flow.
Two postures. Atlas is the truth in both.
Atlas grounds the AI either way — third-party copilot or Jubi-built agent. The difference is enforcement: external AIs can call Atlas as a tool; Jubi agents are required to.
Guardian sits on the routes a third-party AI takes when it touches your data. Atlas is exposed as a tool the model can call. The model may ground in Atlas; it isn't forced to.
Useful when the model lives outside your perimeter (ChatGPT, Copilot, vendor-embedded assistants) and you can't enforce a path. You still get identity binding, tool-call inspection, and audit at the boundary.
For agents your teams build on the Jubi platform, Guardian owns the full request path. Atlas grounding is enforced at the output gate — answers without an Atlas-anchored citation don't reach the user.
This is where the strongest claim holds: same question, same answer, replayable end to end, with field-level permission semantics applied per token of output.
Analysts curate. Workspace by workspace.
Atlas is curated, not generated. Analysts own each definition; AI can help draft and propose, but humans decide what's true. The unit of curation is the workspace: a scope, a set of definitions, and a small team who own them.
Scope a workspace
An analyst picks the tables, dashboards, and collections that belong to a domain — finance, ops, customer success. The workspace is a permissioned slice of the warehouse.
Annotate the data model
In Context Studio, the analyst marks which fields are PII, which columns mean what, and which dashboards are verified. The catalog tracks coverage so it's visible what's annotated and what isn't.
Define glossary terms and metrics
Term definitions go in the glossary. Metric formulas are written once, named, and versioned. The AI uses these names in answers — the user can click through to the definition.
Write playbooks for repeated analyses
Common questions become playbooks: a named, parameterised analysis the AI can run on a schedule or on demand. Playbooks are reviewed and validated by the analyst before they go wide.
Atlas references it all
The workspace publishes its definitions into Atlas. Guardian uses the published version on every request. When the analyst updates a metric formula, the next answer reflects the change.
The editing UI for the semantic layer (glossary terms, metric definitions, playbook content) is in active build. Today, Context Studio supports browsing and scoping; the in-product editing experience is being added in the next release.
The shape of the truth.
Three small examples — a metric, a glossary term, and a permission rule — to make Atlas concrete. The schema is more elaborate in practice; this is the spirit.
name finance.gross_margin_quarterly definition (revenue - cogs) / revenue grain quarter owner finance-analytics verified 2026-04-15 source card #1842 (Metabase)
term churn scope customer-success definition customer with no order in 60d aliases churned, lapsed not cancelled (different concept) owner cs-ops
field customer.email class pii visible_to account_owner, support masked_for all_other summary forbidden (no inference path) enforced guardian · output gate
from customer to account cardinality 1 : N join customer.id = account.customer_id notes a customer may hold multiple accounts owner data-eng