Key findings
- 23 of the 59 accounts say answer quality depends on CoWork's semantic views, the most common position GILD recorded.
- 9 accounts raise answer accuracy as a concern and 3 more discuss it. A healthcare developer's agent hallucinated when asked to list all patients.
- 6 accounts raise cost and 5 more discuss it. One team spent about US$150 and about ten hours building semantic layers, and expects the value to exceed a yearly cost of US$35,000 to US$75,000.
- 11 accounts praise how fast an agent goes live and 14 praise plain-language answers for business users. 10 say dashboards still serve recurring reports.
- 14 accounts discuss where people should ask. 11 reach Snowflake's agents from Claude, Teams, or their own app, and 3 keep questions in CoWork.
Build small, tested semantic views before business users see CoWork, and keep dashboards for recurring metrics. Public accounts trace wrong answers to the semantic views behind the agent, and access setup and cost are the next complaints.
Where it works
Accounts that praise CoWork describe clean dimensional models with a small semantic view for each topic, tested through verified queries. One small team built a separate star schema for each topic, each with its own semantic model and verified queries. The team says that stopped hallucinations and kept credit costs reasonable. Another team replaced its Streamlit chat apps with CoWork's interface. Mastech measured the gap on 30 healthcare questions. A semantic view CoCo generated from 4 tables scored 63% answer correctness, and a curated 8-table view with synonyms and SQL guidance scored 77%.
The company-wide figures come from Snowflake itself. Its product data science team reports that more than 5,600 employees and their agents ran over 320,000 queries in July 2025 on one semantic layer. The team keeps that semantic layer in dbt with peer review.
Where trust breaks
Critical accounts describe the agent going past its instructions. A healthcare developer found the agent hallucinated on simple questions such as listing all patients, despite semantic views and strict instructions. Two engineers found the agent loosened a semantic view's join conditions, which changed results and slowed queries on large tables.
Cost and speed
Cost and speed are the other recurring limits. A consultant piloting it for a healthcare client's executives found compute cost high for the output and kept Power BI. A team that measured multi-table answers at one to three minutes moved essential metrics to third-party dashboards. 5 accounts raise slow answers.
Where people ask
14 accounts discuss where business users should ask. 11 reach Snowflake's agents from Claude, ChatGPT, Teams, or their own app through the managed MCP server or the API, because people already work there. 3 keep questions in CoWork, which carries the agents and roles directly.
Teams that route questions through another assistant report two costs. The outside session runs under the user's default role, and a question passed from Claude to Cortex Analyst pays for two models.
Methodology
GILD read 59 public accounts published from Sep 5, 2025 to Oct 11, 2026. Reddit supplies 49, read through the Arctic Shift archive, and the rest are eight consultancy and company blog posts and two GitHub issues. GILD checked Snowflake's documentation on October 12, 2026. Each count is the number of those accounts that discuss a topic, so the figures mostly describe people who post on Reddit. Speed and cost figures are the posters' own measurements. The accuracy figures are Mastech's own 30-question evaluation, and the company-wide usage figures are Snowflake's, each linked in Sources.
Products in this piece
Sources
- Official documentation
- 5
- Public source
- 10
- Snowflake docs: Overview of Snowflake CoWork
- Snowflake docs: Snowflake CoWork reference
- Snowflake docs: Snowflake CoWork troubleshooting
- Snowflake docs: Snowflake-managed MCP server
- 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)
- DataTools Pro: building a first CoWork agent from the quickstart (Jun 8, 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)
- Reddit post from a consultant piloting Snowflake Intelligence against Power BI (Jan 3, 2026)
- Reddit comment on a two-day semantic layer workshop and its cost (Jul 22, 2026)
- Reddit post weighing Claude Desktop on a Snowflake MCP server against CoWork (Aug 12, 2026)
- Reddit comment measuring one to three minutes per multi-table answer (Jan 21, 2026)
- Snowflake blog: Building an Internal Context Layer for AI Agents at Snowflake (Aug 17, 2026)