dbt Labs: AI Agents Could Succeed Where the First Semantic Layer Wave Stalled

dbt Labs staff product manager Zach Mandell used his dbt Summit 2026 session to trace what he calls the first and second waves of the semantic layer - and to argue the second, powered by AI agents, has a genuine chance where the first fell short.

The first wave, roughly five years ago, centered on defining metrics as code: one repository of metric logic served to every BI tool so revenue wasn't calculated two ways. Mandell said it largely failed to reach production for three reasons: it delivered little beyond the status quo, integrating a shared layer into existing BI tools was hard, and semantic modeling - entities, primary keys, join paths - was unfamiliar to analytics engineers whose tools had hidden that work.

What's changed, he said, is the value at stake: agents can put "a tireless analyst at the fingertips of anyone in the business," and dbt MCP server usage is doubling quarter over quarter, largely for conversational analytics. That brings a trust problem. Mandell compared it to free throws: Steph Curry's 92.3 percent career rate still yields only about a 45 percent chance of making ten straight - worse than a coin flip. Agent answers degrade the same way across multi-step chains, and better models alone won't fix it.

Mandell laid out four requirements for the second wave. Closing the integration gap through "semantic SQL," a native SQL interface letting BI tools query the dbt Semantic Layer like a database while still allowing ad hoc joins. Pushing semantics upstream into transformation, for pre-aggregation - some companies spend 30 to 40 percent of their warehouse bill on materializations, he said - and for metrics as dimensions in customer-360 models. Treating the semantic layer as the deterministic spine of a wider context layer that includes tickets, docs and calls. And configurability with access control, so agents see only the context a workflow needs, with observability into what data they touched.

In Q&A, Mandell said dbt remains active in the Open Semantic Interchange initiative and can export MetricFlow definitions as an OSI document, though he acknowledged the ecosystem hasn't settled how much fidelity survives conversion between semantic layer specs. A native SQL interface won't replace saved queries, he said, since those serve materializing aggregations. He also cited a public Anthropic post arguing semantic layers and knowledge graphs should be human-curated rather than AI-bootstrapped.

Mandell closed by inviting attendees to a same-day roundtable on how a semantic layer should be authored and served, and what holds teams back from adopting dbt's or other headless semantic layers - including those that tried and gave up a few years ago.

Source: dbt Labs Blog