Public accounts that split star schemas by topic, with one semantic view each, report the best CoWork results. Snowflake draws its internal evaluation questions from logged questions and its most-used dashboard tiles. Open CoWork one modeled domain at a time, after its semantic view has verified queries and its agent passes an evaluation set of real business questions. Publish the agent to one team first, and cap spend before access widens.
When you need this
Evaluation criteria
Setup: model one domain
Which tables answer this domain's questions, and who maintains their definitions?
Look forOne semantic view per topic, started small and grown after testing, with descriptions, synonyms, metrics, and ID columns marked as dimensions. Semantic View Autopilot can draft it. Public accounts report the best results from star schemas split by topic, each with its own semantic view.
Setup: pin the important questions
Which questions must always return the same SQL?
Look forVerified queries for each KPI. An exact match runs the validated SQL directly. A close match only guides the agent, which then writes its own SQL, so state in the orchestration instructions when the agent must not regenerate SQL.
Governance: gate on evaluations
What must the agent answer correctly before a domain opens?
Look forAn evaluation dataset of real questions with ground-truth answers, scored for answer correctness and logical consistency. Snowflake builds its own internal evaluation sets from popular dashboard tiles and logged questions.
Rollout: publish to one team
Who sees the agent first?
Look forThe agent added to the CoWork object, USAGE granted to one team's role, and default roles and warehouses set for every user so they do not switch after login.
Review practice: watch where it fails
Who reads the traces and feedback each week?
Look forA weekly pass through the agent's observability traces and thumbs-down feedback, with fixes going back into the semantic view and verified queries.
Cost control: budgets before scale
What stops one team's questions from draining the AI budget?
Look forA resource budget on the CoWork object with actions at 80% and 100%, and per-user quotas on the CoWork domain, set before access widens.
Questions to ask vendors
- Which data domain has the most repeated questions and the cleanest dimensional model to start with?
- Who owns each semantic view after launch, and how often do they review its failures?
- Should business users ask in CoWork, or from an assistant they already use through the managed MCP server?
- What share of evaluation questions must pass before a domain opens to the whole company?
Red flags
- One large semantic view covering every table, which public accounts tie to wrong joins and invented aggregations.
- No evaluation set, so the first sign of a wrong metric is a leader quoting it.
- Verified queries treated as guaranteed, without instructions that stop the agent from re-planning them.
- CORTEX_USER still granted to PUBLIC when the agent opens to a pilot team.
Products in this piece
Sources
- Official documentation
- 10
- Public source
- 5
- Snowflake docs: Getting started with Snowflake CoWork
- Snowflake docs: Overview of semantic views
- Snowflake docs: Semantic View Autopilot
- Snowflake docs: Cortex Analyst Verified Query Repository
- Snowflake docs: Snowflake CoWork reference
- Snowflake docs: Cortex Agent evaluations
- Snowflake docs: Monitor Cortex Agent requests
- Snowflake docs: User access and settings for agents
- Snowflake docs: Resource budgets for Snowflake CoWork
- Snowflake blog: Building an Internal Context Layer for AI Agents at Snowflake (Aug 17, 2026)
- Mastech: a 30-question accuracy evaluation of Cortex Analyst semantic views (Jun 2, 2026)
- InterWorks: testing Snowflake CoWork on a retail dataset (Aug 13, 2026)
- Reddit comment on topic-specific semantic views that stopped hallucinations (Jan 14, 2026)
- Reddit post on hallucinated answers despite semantic views (Jan 14, 2026)
- Reddit post on an agent ignoring a semantic view's join conditions (Oct 12, 2025)