Share with your CDO
Most enterprise semantic layers, the governed translation layer that turns raw data into consistent business metrics like “revenue” or “active users,” were built for BI dashboards, not AI agents. Pratik Jain of Kyvos Insights walks through the three main semantic layer architectures available today: BI-native (Power BI, Looker), data platform-native (Snowflake, Databricks), and standalone (Kyvos, AtScale, Cube). The argument is that only a warehouse-independent standalone layer can serve concurrent agentic and BI workloads without forcing them to compete for the same compute budget.
What this means for your business
The gap this piece is really describing is a sequencing problem that’s already biting organizations scaling AI past proof-of-concept. A semantic layer scoped to one BI tool works fine until an AI agent needs the same metric definition the dashboard uses, at which point someone rebuilds it by hand, inconsistently. If your current semantic layer lives inside Snowflake or Power BI, that rebuild cost isn’t theoretical. It’s sitting in your next sprint backlog, disguised as a data quality ticket.
The architectural critique of warehouse-native layers is the sharpest claim here and it holds up independently of who’s making it. When every query, whether from a Tableau report or an autonomous AI agent, hits the same warehouse compute engine, you get resource contention. Agentic workloads are bursty and unpredictable in ways that BI never was. Warehouses like Snowflake and Databricks price on compute consumption, so an AI workload that triggers thousands of small queries can inflate costs faster than any dashboard ever did. A middle-tier execution layer that absorbs that load before it reaches warehouse compute is a legitimate architectural answer to a real cost and performance problem.
Jain is a senior director at Kyvos, which sits squarely in the standalone category he recommends, so the checklist of ideal capabilities maps cleanly onto Kyvos’s own feature set. That doesn’t make the framework wrong, but it does mean the evaluation criteria deserve stress-testing against vendors he doesn’t mention. The question your renewal cycle actually forces is narrower than “which architecture wins” — it’s whether your existing semantic layer vendor has a credible agentic execution story, or whether they’re retrofitting one. If the answer is the latter, the retrofit cost compounds every quarter you wait.
Concept deep-dive: Semantic layer
A semantic layer sits between raw data in a warehouse and the tools that consume it, translating physical table structures into business concepts like “net revenue” or “churn rate.” Think of it as a shared dictionary that ensures every analyst, dashboard, and AI agent uses the same definition of a metric. Without it, the same underlying data produces different numbers depending on who’s querying it, which is the root cause of the “whose number is right” meeting that every data team knows.
Based on reporting from Comparing Semantic Layer Architectures: What options do enterprises have today?, originally published 2026-07-01 02:30:00.
