Semantic Layers for Deterministic LLMs
Leah Weiss, co-founder of Preql, on delivering clean data to LLMs
The real bottleneck in enterprise AI is not answering questions, it is making sure the same question produces the same number every time. In finance and reporting, a model that changes its answer across runs breaks trust immediately, so the winning product is the layer that locks down definitions, permissions, and cleaning steps before an LLM ever sees the data. That is why semantic layers are reappearing as AI infrastructure, not old BI plumbing.
-
The concrete failure mode is simple. Revenue, margin, or headcount often exists in different dashboards, spreadsheets, and warehouse tables with slightly different logic. Preql is built to clean those inputs, reconcile conflicts, and route queries through a governed metric layer so chat interfaces do not improvise on top of messy source data.
-
This also explains why earlier semantic layer products struggled. Teams had to hand code business logic in tools like dbt or BI products, which added work for analytics engineers before the payoff was obvious. AI makes that payoff immediate because natural language interfaces are only useful when the underlying metric definitions are fixed and reusable.
-
The competitive line is between systems that organize knowledge for search, and systems that standardize business facts for decisions. Glean unifies documents, people, and app context for retrieval, while dbt and Cube center governed metrics and semantic models. Preql is positioning around the same trust layer, but with more no code cleanup and workflow capture for finance teams.
The market is heading toward agentic workflows that do not just answer questions, but trigger actions in systems like ServiceNow or UiPath. As that shift happens, the valuable layer will be the one that turns messy company data into approved, repeatable business logic, because autonomous software cannot operate on numbers that change from run to run.