Your AI Doesn't Need More Tokens. It Needs Context.

Mosaic sits between the AI and the database, providing the business context that raw schema alone can't. The cost savings depend on which deployment path you choose.

With Mosaic MCP, the LLM still writes the SQL, but it works from compressed business context instead of your raw schema. Rather than receiving the full schema on every query, the model gets only the definitions relevant to the question being asked. In the benchmark, schema context dropped from 4,428 to 2,073 tokens per query, a 53% reduction. SQL token count dropped 24%. Together, estimated per-query cost came down 37%.

Strategy AI Agents take this further. They remove the LLM from query generation entirely. The LLM handles the natural language question while Mosaic does the rest. It looks up the relevant business definitions, generates the SQL, executes the query, and returns the result. Eliminating schema context entirely is what drives the ~98% reduction against the PostgreSQL baseline.

The business context that compresses token usage also prevents a more dangerous failure: a number that looks correct but is not. When direct text-to-SQL has to infer join logic from raw schema alone, it can choose a path that looks structurally valid but applies the wrong business logic. The database still returns a number, and nothing flags that the result is wrong. Across the benchmark’s 17 questions, direct PostgreSQL missed two. Mosaic answered all 17 correctly: 100% versus 88.2%. The difference comes from architecture. Mosaic defines canonical join paths, approved relationships, and fan-out protections once in the semantic model, then applies them consistently to every query. That makes accuracy a property of the architecture, not something you have to prompt your way toward on every query.

AI analytics pilots are often misleading for exactly the reasons this benchmark illustrates. A pilot starts with a small schema and manageable questions, so direct querying looks both accurate and cheap. Neither advantage survives production, where schemas are larger, questions span more tables and time periods, and every query carries the full schema tax. The failures are quiet and there is no obvious sign that anything went wrong. Just a number that looks right but isn't, and a token bill that climbs with every table and every user you add.

Mosaic addresses both cost and accuracy before the query runs. The business logic is already defined and the join paths are already encoded so the model never has to receive the schema or reconstruct rules on its own. For organizations moving AI analytics into production, the real question is whether the architecture can scale without token costs rising every time the schema grows or another user starts asking questions.

If your team is evaluating AI analytics for production, we'd welcome the chance to show you how Mosaic performs against your own schema. Join the world's largest deployed semantic layer and give AI the business context it needs to deliver trusted answers at scale.

For the full methodology, detailed results, and cost analysis discussed in this blog, read Strategy's AI Token Cost and Accuracy Benchmark.