GILD verdict
Buy Snowflake CoWork if your data team will maintain semantic views and evaluations for each domain it opens.
Snowflake reports that its own data team covered under 5% of 9,685 tables with semantic views, and that AI answered about 25% of questions correctly without that context. Budget the modeling work and AI Credits before expanding beyond one team.
Snowflake CoWork
AI Data Analysis by Snowflake
+36 PositiveFit and limits
Who it fits
- Your analytics data lives in Snowflake, and the data team can model it as semantic views with descriptions, synonyms, and metrics.
- Business users wait on the data team for ad hoc questions that a governed agent could answer under their own roles.
- Answers must respect existing row access and masking policies without a second permission model.
- You would pay per question over paying per seat, with spend capped by resource budgets and per-user quotas.
Where it falls short
- Accuracy depends on semantic views and verified queries the data team builds and maintains; Snowflake states that agent answers are not guaranteed and should be reviewed.
- On four BIRD databases, Snowflake's improved semantic models lifted accuracy from 57% to 78% in March 2025, so about one answer in five still missed.
- Spend adds up across orchestration tokens, tool calls, and warehouse compute, so the monthly total is harder to predict than a seat price.
- Snowflake's troubleshooting guide says requests often take longer than a minute, so CoWork does not replace a dashboard people open every morning.
- User skills and document generation are still in beta. Snowflake's June 30, 2026 post put Cortex Sense, which builds context for tables without semantic views, at private preview soon, starting with a single role.
Implementation reality
Model one domain first
Build a small semantic view for one domain with descriptions and synonyms, plus defined metrics. Add columns after testing. Semantic View Autopilot can draft the view. Add verified queries for the questions that matter most.
Build and evaluate the agent
Create a Cortex Agent with the Cortex Analyst tool on that semantic view, Cortex Search for documents if needed, and orchestration instructions. Run an evaluation set for answer correctness and logical consistency before anyone else sees it.
Publish to one team
Add the agent to the CoWork object so users see it. Grant USAGE to the pilot team's role. Review observability traces and user feedback to find incorrect answers.
Set budgets and access before scale
Set a resource budget on the CoWork object and per-user AI quotas. Narrow Cortex access by removing CORTEX_USER from PUBLIC. Choose a cross-region inference setting, then enable Cortex AI Guardrails if your edition supports it.
What to measure in your pilot
- Count how many of the pilot team's real questions fall outside the modeled semantic view each week, and expand to a new team only when at least 80% land inside it.
- GILD found no per-user AI Credit figures in Snowflake's docs or practitioner posts; log each pilot user's credits weekly with Deep Research and Automations on, and set quotas from week four.
- Track analyst hours spent on semantic views per domain; if Cortex Sense reaches your account, rerun the same evaluation set with it on and off before cutting modeling time.
Sources
- Official documentation
- 21
- Snowflake docs: Overview of Snowflake CoWork
- Snowflake CoWork product page
- Snowflake docs: Snowflake AI pricing
- Snowflake Service Consumption Table (effective Oct 15, 2026)
- Snowflake docs: User access and settings for agents
- Snowflake docs: Snowflake CoWork reference
- Snowflake docs: Snowflake CoWork troubleshooting
- Snowflake docs: Resource budgets for Snowflake CoWork
- Snowflake docs: Per-user quotas
- Snowflake docs: Cortex Agents
- Snowflake docs: Cortex Agent evaluations
- Snowflake docs: Monitor Cortex Agent requests
- Snowflake docs: Cortex Analyst Verified Query Repository
- Snowflake docs: Semantic View Autopilot
- Snowflake docs: Cortex AI Guardrails
- Snowflake docs: Cross-region inference
- Snowflake docs: Snowflake AI and ML
- Snowflake docs: Artifacts in Snowflake CoWork
- Snowflake engineering blog: Cortex Analyst, Evaluating Text-to-SQL Accuracy for Real-World BI (Aug 29, 2024)
- Snowflake engineering blog: Agentic Semantic Model Improvement (Mar 31, 2025)
- Snowflake blog: Cortex Sense for Enterprise AI Agents (Jun 30, 2026)