Watch AI agents get the right answer from a semantic layer, through MCP

Watch an AI agent query AtScale’s semantic layer through MCP. See how CoModel, Ava, and ACE replace text-to-SQL guesses with computed, auditable answers.

Why do AI agents give wrong answers with my business data?

Most AI agents answer data questions with text-to-SQL. The model reads your table names, guesses at the joins and the business definitions, and writes its own query. LLMs are probabilistic, so that guess can change every time you ask, which is useful for writing and code and a liability when the output is a revenue number.

AtScale is a semantic layer that sits between your data and your AI agents. Agents connect through MCP and ask for a metric by name, and AtScale computes the number from your governed definitions, so the same question returns the same answer every time.

In this tour, Ryan Siegler from AtScale’s developer relations team connects Claude to the semantic layer through MCP, asks an open-ended question, and then opens up the queries and the model behind the answer.

What you’ll see

1. Ava (Ask, Verify, Answer) resolves the question. When a term like “revenue” maps to more than one definition, Ava asks which one you mean before any query runs.
2. ACE computes the answer. The AI Computation Engine builds the query from the model and runs it where your data already lives, using aggregates to keep responses fast.
3. CoModel can update and build the model. AI-assisted semantic modeling drafts metrics, dimensions, and hierarchies from your context. Nothing publishes until you approve the diff, and every change is versioned in git.

Works with your AI tools: Claude, ChatGPT and Codex, and Grok, on Snowflake, Databricks, and BigQuery.

How can I ensure my AI outputs are defensible and auditable?

On the BIRD-Interact public benchmark, every model we tested scored higher with an AtScale semantic layer. Claude Sonnet 5 went from 25.9% to 56.6% on 410 business questions.