Snowflake Ventures enterprise AI funds trust layer

Snowflake frames Snowflake Ventures enterprise AI as a bet on the layer between models and apps, where pilots either become products or stall.

Head of Snowflake Ventures Harsha Kapre named Dust and Gray Swan on Sep 8 as examples of that layer in production.

How governance failures turn into security failures

Kapre argues most agent programs do not fail on model quality but on missing identity-aware access, policy guardrails and a single source of truth.

Without that foundation even capable models cannot safely act across business workflows, so governance gaps become security gaps.

Think of it like giving interns master keys to file paperwork, fast work but no audit trail until something leaks.

Dust targets the workflow side, giving teams shared knowledge, search and Model Context Protocol connectors to run cross-functional agents on any model.

Gray Swan targets the runtime side, scanning for vulnerabilities and enforcing protections before agents touch production data.

Together they map to the three capabilities Snowflake calls foundational: model choice, governance at scale, and human-agent collaboration.

Why production trust is still unfinished

Snowflake already pushed this thesis days earlier with CoCo, its coding agent that keeps data inside the perimeter, so Dust and Gray Swan extend the same inside-the-governance-boundary pitch.

The overlap matters because Dust is integrating with Cortex Agents to ground agents in Snowflake data, which tightens Snowflake as the source of truth but also concentrates lock-in risk if governance lives there.

Gray Swan fills a hole Snowflake cannot credibly fill alone, runtime defense for nondeterministic agents that traditional data controls were never built to watch.

Databricks has made parallel governance bets, so Snowflake is signaling it will buy distribution for this layer rather than concede it.

Neither announcement discloses investment size, terms or product milestones, so there is no way to judge traction beyond described GTM and ops usage at startups and enterprises.

Procurement will now ask how agents inherit identity, log actions and revert bad writes, not which model they use.

Until vendors publish third-party tests for those controls and clear rollback paths, the agentic enterprise stays a governed pilot for most shops.

Snowflake is betting outsiders solve its hardest trust problem faster than it can ship natively.