AtScale for AI: Technical Overview

Why LLM-generated analytics returns wrong numbers, and the AtScale component that fixes each source of error.

On the BIRD-Interact business questions that have a correct answer, Claude Sonnet 5 got 46% right writing its own SQL. With an auto-generated AtScale semantic layer and no manual tuning, it got 93%. This white paper explains the architecture behind that gap. It traces AI errors on enterprise data to seven failure points, from ambiguous prompts and dirty data to answers no one ever tests, and maps each one to the AtScale component that addresses it.

What you’ll learn:

1. The seven failure points behind wrong AI answers on business data, and why hand-written skills, catalogs, ontologies and plain text-to-SQL all leave joins, grain and metric logic in the model’s hands.
2. How ACE, the AI Computation Engine, compiles each request into the same SQL every time, including multi-pass ratios, semi-additive balances and fiscal calendars, and pushes it down to your own data platform.
3. Why agents see AtScale as one wide Postgres table over MCP, so they pick a metric and a filter instead of writing joins, under their own identity with row- and column-level security applied.
4. How semantic memory stores certified query shapes and reuses them, raising accuracy by about 8 points across three benchmark runs from a cold start.
5. How CoModel drafts the semantic model from the catalogs, documentation and BI models you already have, and how Canary tests it against gold-standard queries, names the metric behind each miss and sends it back to CoModel as a fix.