Before an AI agent can act, it needs to understand your business

The hardest part of building a trustworthy AI agent isn't the agent itself - it's everything the agent needs to know before it can be trusted to act. That was the argument Ben Moser of dbt Labs and Kevin Kim of Fivetran made during a breakout session at dbt Summit 2026.

Without shared context, Moser said, a simple question like "what was our ARR at the end of August" splinters into a dozen sub-questions: booked or current ARR, does it include services, which table is authoritative, who's allowed to see it. "Context aims to create an accurate world for your agents to operate in," he said. "It helps them understand your business so that they can make the right decisions, answer your questions in the right way, and not cause havoc."

Moser described context as a spectrum rather than a single problem to solve at once, from simple semantic and structural context to complex situational context where meaning shifts over time. Loading everything into an agent's environment at once, he warned, is the wrong instinct. A dbt project already provides a governed map through models, tests, documentation and lineage, but the interesting problems sit at its edges - support tickets, Slack threads and institutional knowledge that never reached the warehouse. Moser framed the stakes in offensive and defensive terms: used well, agents give a business an edge; used carelessly, inaccuracies compound at machine speed, burning trust and token spend together.

That's the gap Kim's team built Fivetran Context Layer to close. It connects knowledge sources like Notion and Confluence, operational sources like Salesforce, HubSpot and Postgres, and analytics sources like dbt and Sigma, then parses it all, builds a search index for semantic search, and infers an ontology linking concepts across the three. Inside Agents Schema, a root table acts as a glossary of what's in the schema, while a concept table stores company-specific shorthand so an agent doesn't have to guess what an acronym means. Fivetran writes what it learns into the warehouse as open, ordinary tables any AI client can query over MCP, rather than locking context behind a proprietary API. In a live demo, Kim showed an agent answering "which open opportunities have unresolved support tickets" by querying Agents Schema directly. A new traces feature surfaces where an agent's confidence was low or context was missing, so analytics teams can curate and correct it over time.