Snowflake CoWork

What it is

This page covers Snowflake’s agent layer, not the warehouse: Snowflake CoWork and the Cortex Agents platform under it. CoWork launched as Snowflake Intelligence, went generally available on November 4, 2025, and was renamed CoWork at Summit on June 2, 2026, when it gained Deep Research, saved artifacts, user memory and scheduled automations (the last in public preview). A business user asks a question in plain language; a Cortex Agent plans the steps, sends structured questions to Cortex Analyst (text-to-SQL over a semantic view), sends document questions to Cortex Search, and returns an answer with a chart and the SQL that produced it. Everything runs inside your Snowflake account, so row access policies and column masking apply to every answer under the user’s role. CoWork is reachable in Snowsight, a Slack bot and an iOS app. Through MCP connectors it can act outside Snowflake: Snowflake’s docs list prebuilt connectors for Atlassian, GitHub, Glean, Linear and Salesforce, all OAuth-only. In the other direction, the Snowflake-managed MCP server, GA since November 4, 2025, exposes Cortex Analyst and Cortex Search to Claude and other MCP clients.

Do not confuse it with Anthropic’s Claude Cowork, a desktop agent for general knowledge work. Snowflake CoWork runs on Claude models but only reasons over what lives in, or is connected to, your Snowflake account.

Why it shows up in RevOps and CS stacks

• The data is already there. If Fivetran lands CRM, billing and product usage in Snowflake, CoWork answers “which renewals over $50,000 have falling weekly active users” without exporting to a BI tool or a separate AI analyst.
• Governance is inherited. Answers run under the user’s Snowflake role, and Cortex AI Guardrails, GA for CoWork and Cortex Agents since May 14, 2026, screen prompts for injection attempts. For CS teams querying telemetry with customer PII, that is usually what gets security sign-off.
• One semantic view serves several consumers. The same semantic view feeds CoWork, Claude through the managed MCP server, and BI tools like Hex that import Snowflake Semantic Views.

Pricing reality

There is no seat fee. Snowflake bills CoWork in AI Credits: $2.00 each with cross-region routing allowed, $2.20 pinned to your home region, flat across editions. Usage is metered per million tokens and adds up across every service the agent calls. Per the Service Consumption Table effective September 30, 2026, Claude Sonnet 4.5 through Cortex Agents costs 1.95 AI Credits per million input tokens and 9.76 per million output tokens — $3.90 and $19.52 at list. The SQL each answer runs then bills separately on your warehouse in regular credits ($2–$4 each by edition in AWS US East), and a Cortex Search service bills for every GB it indexes each month whether anyone queries it or not.

Our estimate, at 30,000 input and 3,000 output tokens per multi-step question, is about $0.18 in AI Credits per question. A 40-person ops team asking 1,000 questions per working day lands near $4,000 a month before warehouse compute, and the bill moves with usage. Budget from your own SNOWFLAKE_COWORK_USAGE_HISTORY after a pilot, not from this number.

Best for

RevOps and CS ops leaders at companies whose CRM, billing and product data is already modeled in Snowflake, who want 20–200 business users answering their own account and pipeline questions under existing warehouse governance — without buying a second analytics seat for each of them.

Skip it if your data is not in Snowflake: CoWork has no native Salesforce or HubSpot data connector and needs the data ingested first. Skip it if users need to edit numbers — forecast submissions and quota overrides belong in a writeback tool.

Versus the alternatives

Microsoft Power BI Copilot is the incumbent by installed base and runs on any paid Fabric capacity from F2 upward. Pick it when the company already lives in Microsoft 365 and dashboards are the main output. Pick CoWork when the data sits in Snowflake and you want no second semantic model to maintain.

Databricks Genie is the direct warehouse-native rival. Since July 6, 2026 it bills pay-as-you-go in DBUs after 150 free DBUs per user per month. Pick Genie if your lakehouse is Databricks; the two are not interchangeable.

Hex Threads is the fastest-moving entrant, with seat pricing — Team is $75 per editor per month — and answers that escalate into a notebook an analyst can audit. Pick Hex when a data team wants to own and review what the AI answers. Pick CoWork when you want zero extra seats and Snowflake-native governance.

Sigma wins when the job is writeback apps on top of the same Snowflake data. To push CoWork-identified segments into the CRM on a schedule, pair it with Hightouch.

If none fit because nobody agrees what an “active account” is, fix the model first — see data warehouse vs CDP.

Watch-outs

• There is no hard spending cap by default. Token billing scales with question complexity, and Deep Research runs parallel sub-investigations for up to ten minutes. Guard: set per-user AI credit limits, attach a resource monitor to the warehouse CoWork uses, and review SNOWFLAKE_COWORK_USAGE_HISTORY weekly for the first 60 days.
• Answer quality is only as good as the semantic view. Vague column names and missing metric definitions produce confident wrong SQL. Guard: define your 20–30 core metrics in a semantic view, keep it under roughly 100 columns, and run a fixed set of known questions monthly to catch drift.
• A role misconfiguration silently changes the data people see. CoWork starts under each user’s default role and falls back to PUBLIC if that role is not granted, and shared artifacts re-run under each viewer’s role, so two people can see different numbers. Guard: grant CORTEX_AGENT_USER explicitly to a dedicated role, set it as the default role for CoWork users, and test artifacts with a low-privilege account before sharing.
• MCP connectors are OAuth-only, tool-only, and unverified by Snowflake. Snowflake does not maintain the external servers, and access to a server does not grant its tools. Guard: grant write tools such as Salesforce record updates per tool to a small role, and keep external actions read-only until you have an audit trail you trust.