Grok and AtScale help your AI stop guessing what business metrics are and how they’re computed

AtScale joined the SpaceX Super Intelligence Partner Network, so now Grok Build gets the metric definitions your finance team already signed off on.

Executive summary: AtScale is now a Marketplace Partner in the SpaceX Super Intelligence Partner Network. The AtScale plugin for Grok Build, available today, gives Grok the metric definitions a company has already approved and computes each answer on that company’s own data platform. Developers stop rebuilding business logic, and business teams get the same numbers Finance reports.

Key takeaways:

1. Grok Build reads your approved metrics through AtScale’s MCP server, inside each user’s existing permissions.
2. AtScale Ava resolves which metric a question means before ACE computes it on the platform you already run, so your data doesn’t move.
3. On the 317 BIRD-Interact questions that have a right answer, an AI-generated AtScale semantic layer took Grok 4.3 to 62.5% accuracy, and every model AtScale tested improved, by as much as 345% (preliminary results).

Ask any frontier model how many customers churned last quarter and you’ll get an answer in seconds. Whether it’s right depends on rules that aren’t in your database.

Does churn count logos or dollars? Does a downgrade count? When does a customer stop being active, and which accounts does Finance leave out? Those rules live in the finance close, a Power BI model, a dbt project and somebody’s spreadsheet. They rarely agree. A model that can’t see them picks one based on a column name. The chart looks finished, so nobody checks whether the model guessed.

Today AtScale joins the SpaceX Super Intelligence Partner Network as a Marketplace Partner, and the AtScale plugin for Grok Build is available now. With the plugin, Grok reasons over your data while AtScale supplies the definitions and computes the number the business already approved.

Smart models still get the number wrong

You can measure how often models get it wrong. BIRD-Interact is a public benchmark that runs business questions against real enterprise databases. AtScale ran seven models from six vendors on it, with and without an AtScale semantic layer. Grok 4.3 had the highest score of any of them without a semantic layer. Even so, every model answered more questions correctly once a semantic layer sat underneath it, by as much as 345%. For Grok, an AI-generated AtScale semantic model took it to 62.5% accuracy on the 317 questions that have a right answer (preliminary results).

Since every enterprise can license the same models, the advantage goes to the company whose AI reads the right definitions.

How Grok and AtScale work together

1. Grok Build connects through AtScale’s MCP server. It reads metrics, dimensions and relationships, and it only sees what the user’s existing permissions allow.
2. AtScale Ava resolves the question first. When a person’s question could mean more than one metric, Ava asks which one they meant and saves the answer as a rule. When an agent calls, Ava points it at the right certified metric, because an agent has no one to ask.
3. ACE, the AI Computation Engine, computes the number. It runs on the data platform you already use (Snowflake, Databricks, Google BigQuery, Amazon Redshift, PostgreSQL or InterSystems IRIS), so your data doesn’t move and the agent doesn’t write the metric SQL. The same question gets the same SQL and the same number, which is what lets a Grok-built report go in front of a board.

Say an analyst asks Grok for churn by plan. The model holds three churn metrics, so Ava asks which one they mean. The analyst picks logo churn, the one in the board deck, and Ava saves that as a rule so the next person isn’t asked again. When a Grok Build agent makes the same request, Ava skips the question and points the call straight at the certified metric. Either way, ACE computes it and returns the definition, its owner and the exclusions it applied. (Illustrative example.)

“When Grok adds a revenue metric to an application, it should use the same definition Finance uses in the quarterly report,” said Jay Schuren, President of AtScale. “Our work with SpaceXSI brings that business context directly into the development process. Developers can move quickly, and business teams can trust that the applications they build use the metrics they’ve agreed on.”

What changes for developers and business teams

• Developers stop rebuilding business logic. Metric definitions come from one model instead of being recoded in every app, dashboard and agent.
• Business teams get the number they agreed on. The figure in a Grok-built app matches the one Finance published.
• Each right answer costs less. ACE serves repeat questions from aggregates and cache, so cost per right answer fell about 21% across the seven models AtScale tested (preliminary).
• Access stays where it is. You sign in as yourself, the MCP tools are read-only, and warehouse credentials never leave AtScale.
• Every answer is inspectable. You can read the query ACE ran and the definition it computed from.
• One setup works beyond Grok. The same MCP server works in Cursor, Claude, ChatGPT and Google Gemini, and every client requests the same metric definitions.

“Enterprise teams want agents that can work with their data and understand how their business operates,” said Toni Adams, Head of Partnerships at SpaceXSI. “AtScale brings established business definitions and governed computation to Grok, helping customers build useful applications and extend agents into everyday business workflows.”

What’s next

AtScale and SpaceXSI plan to bring trusted business context to more of the development loop, starting with joint go-to-market for shared customers. The aim is for enterprises to use Grok on their most important numbers and prove those numbers are right.

Get started

Already running AtScale? Install the AtScale plugin for Grok Build from the marketplace listing and connect it to your own metric definitions. Setup takes about ten minutes. Want the full walkthrough? Read AtScale and Grok: The Comprehensive Guide.

Benchmark figures are from AtScale’s own preliminary BIRD-Interact runs in October 2026. They cover seven models from six vendors, each run with and without an AtScale semantic layer, with Grok 4.3 scored as the mean of three runs. The BIRD team has not verified these results, and they don’t compare with BIRD leaderboard results.