Last edited on Oct 27, 2026
In this session, Zach Mandell, staff product manager for the dbt Semantic Layer at dbt Labs, walked through what he called the first and second waves of the semantic layer, and made the case for why the second one, driven by AI agents, has a real shot at succeeding where the first one stalled.
The first wave, roughly five years ago, was about defining metrics as code: centralizing metric logic in one repository and serving it to every downstream BI tool so nobody was calculating revenue two different ways. Mandell argued that wave mostly didn't reach production, for three reasons: the value it delivered was just the status quo (the same dashboards, arrived at more carefully), integrating a shared semantic layer into existing BI tools was genuinely hard, and semantic modeling itself, thinking in entities, primary keys, and join paths, was an unfamiliar discipline for analytics engineers whose tools had abstracted all of that away.
What's different now, Mandell said, is the value on the table: agents can put "a tireless analyst at the fingertips of anyone in the business," and usage of the dbt MCP server is doubling quarter over quarter, driven largely by conversational analytics use cases. But that same shift introduces a new problem: trust. Mandell illustrated it with a probability exercise: Steph Curry makes free throws at a 92.3 percent career rate, better than nine of ten, yet his odds of making ten in a row work out to only about 45 percent, worse than a coin flip.
The same math applies to an agent answering a business question through a chain of steps (finding the right table, the right column, the right aggregation); even a high per-step success rate degrades fast across a multi-step chain, and Mandell argued that better models alone won't fix that.
Mandell laid out four things he believes have to be true for the second wave to succeed. First, closing the integration gap: rather than only serving BI tools through custom connectors, dbt is working on what he called "semantic SQL," a native SQL interface that lets a BI tool query the dbt Semantic Layer as if it were querying a database directly, while still allowing arbitrary SQL joins against ungoverned, ad hoc tables when needed. Second, pushing semantics further upstream into the transformation layer itself, both for pre-aggregation (some companies he's spoken with spend 30 to 40 percent of their warehouse bill on materializations to hit latency targets, a cost he believes dbt can help reduce) and for using metrics as dimensions in customer-360-style models.
Third, treating the semantic layer as the deterministic spine of a broader "context layer" that also includes unstructured data like tickets, docs, and calls, arguing that agents reason more reliably when core entities and relationships are already defined rather than inferred from raw tables. Fourth, configurability and access control: agents should only see the exact slice of context a given workflow needs, with observability into what data an agent actually touched, since that combination of scoped access and visibility is what makes higher trust possible.
In Q&A, Mandell fielded questions on standardization and openness: dbt remains an active member of the Open Semantic Interchange (OSI) initiative and can export MetricFlow definitions as an OSI document for use elsewhere, though he acknowledged the ecosystem hasn't resolved how much fidelity survives when converting between different semantic layer specs. Asked whether a native SQL interface to the semantic layer might replace dbt's existing saved queries feature, Mandell said no, saved queries are more about materializing aggregations, a different problem. He also referenced a public Anthropic post arguing that a semantic layer or knowledge graph should be human-curated rather than AI-bootstrapped, a point he called relevant to how organizations should collaborate on semantics going forward.
Mandell closed by inviting the room to a same-day roundtable to debate two open questions with the dbt product team: how a semantic layer should be authored, and how it should be served, plus what's currently holding teams back from adopting the dbt Semantic Layer or other headless semantic layers, including teams that tried and gave up a few years ago.
Watch every keynote and breakout from dbt Summit 2026 on demand. Watch dbt Summit 2026 on demand →
Get started in dbt
Join the analytics engineers building data infrastructure that actually scales.
Install dbt Wizard CLI
Get started with an agent purpose-built for analytics engineering. It knows which tool to call, which context to pull, and checks its own work before surfacing anything to you.
Latest posts
How dbt Labs and Fivetran are moving dbt workloads onto the data lake, model by model
on Oct 27, 2026
Inside dbt v2: how a rebuilt SQL engine makes linting up to 250x faster
on Oct 27, 2026
What Open Data Infrastructure looks like with dbt Mesh
on Oct 27, 2026
Join the largest community shaping data
The dbt Community is your gateway to best practices, innovation, and direct collaboration with thousands of data leaders and AI practitioners worldwide. Ask questions, share insights, and build better with the experts.