AI-assisted semantic modeling (Coming this fall)
The end of
hand-made metrics.
Build your semantic model by hand in Design Center today, or have CoModel* agents draft it from the context you already have. Your team approves every definition.
A semantic model defines your business metrics and how to compute them from your tables. When someone prompts an AI tool for revenue, the model specifies the rows, joins and aggregation. AtScale writes models in SML, its open modeling language, and stores them in Git with the rest of your code. ACE computes metrics from the model, and Ava uses it to work out which metric a question means.
THE PROBLEM
Definitions drift faster than a team can fix them.
Hand-built models fall behind when the business changes.
A small team usually builds the semantic model by hand, one table at a time, over several weeks. After it ships, product lines get renamed, regions split and finance changes how it recognizes revenue. Unless someone updates the model each time, its definitions go out of date, and AI tools keep returning answers based on the old ones. An agent running a thousand queries a day won’t notice the drift.
HOW MODELING WORKS
Your context goes in,
and a semantic model comes out.
AtScale builds the model from what you already have: warehouse schemas, catalogs, documentation and earlier models. A modeler can build it by hand in Design Center, or CoModel* agents can draft it. The finished model is saved as SML in your Git repository after review.
AtScale CoModel : Humans in the loop
Context in
Ontologies
Data catalogs
Data warehouses
Models you already built
CoModel Agents
Reads the schema
Reads the query history
Reads the catalog language
Drafts the model in SML
Watches for drift
Semantic model out
Metrics & dimensions
SML in your repo
Approved by a person
Computed by ACE
Every pass makes it smarter
MODEL BY HAND, TODAY
Design Center gives modelers a visual canvas and a code editor.
Modelers drag fact tables and dimensions onto a canvas, draw the relationships and define each metric. The text editor shows the same model as SML, so engineers can edit it as code. A finished model goes through a pull request and deploys from dev to QA to production.
MODEL WITH AI, THIS FALL
CoModel* agents draft the model for your team to review.
CoModel is AI-assisted semantic modeling. It runs inside a coding agent such as Claude Code or Codex, with a data engineer directing the work. CoModel reads your warehouse schemas, catalogs, documentation and existing models, then drafts metrics and dimensions in SML. It asks the engineer about business decisions it can't make on its own, and it submits each change as an SML diff for review.
THE PROOF
An auto-generated model doubled AI accuracy on BIRD-Interact.
For BIRD-Interact, the public benchmark built from realistic enterprise questions, AtScale generated the semantic models automatically, with no hand edits. Those models alone made an LLM twice as accurate. CoModel builds on the same approach and adds human review of each definition.
FAQs
How do I know who changed a metric definition, and when?
Every definition lives in your repository in SML, a text format, so a change has an author, a date, a diff, and an undo. Your git history becomes the audit trail: no more Thursday-afternoon edits with no record of who touched what or why the board deck suddenly disagrees with the dashboard.
Can an AI agent see data a person isn't allowed to see?
No. The query runs as the person asking, so the row and column rules already in your warehouse apply to the agent too; there’s no separate permission system for AI to slip through. Without that, an agent pointed at a semantic model can reach every metric in it, including ones an analyst couldn’t open on their own, which is exactly the risk Governance closes.
What's a "blast radius," and how do I see one before a change ships?
A blast radius is the list of metrics and reports a definition change actually moves. Before a change merges, AtScale computes the old definition beside the new one and names what shifts, so a one-line edit to how revenue is defined shows its four-report impact in the pull request instead of in a board meeting after the fact.
Does shipping metric changes through review slow the team down?
No more than any other production change would. A change to a definition arrives as a pull request, gets reviewed with its blast radius attached, and promotes or reverts the way code does. That review step is what catches a breaking change before it ships, not after finance notices the board deck is wrong.
How does Governance relate to ACE, Ava, CoModel, and Canary?
ACE computes the metric, Ava works out which metric a question means, CoModel drafts and maintains the model, and Canary scores the answers (Canary coming this fall). Those four decide what the answer is. Governance decides whether it ships: which version of a definition is live, who approved it, and which rows the person asking is allowed to see.