Snowflake CoWork Adds Profiles and Dashboards, Builds on Automations

Profiles route questions to the right agent

According to Snowflake, companies that build more agents create a new problem for users. A sales rep, a finance analyst and a store manager each need different data and tools, and if they are shown a list of agents, many will pick the wrong one, get a weak answer and give up.

A Profile removes that step. An admin configures a Profile for a group of people, defining which agents it can reach, how it routes questions and which instructions, skills and connectors come with it. Users simply ask, and the Profile sends each question to the agent best suited to answer it. Those who already know which agent they want can mention it by name.

Snowflake gives an example of how the same words can mean different things. A product team member asking what is happening with a customer account could be routed to a product usage agent, while a sales team member asking the same question could be routed to the sales agent.

Profiles do not widen access. CoWork runs with each user's own privileges, and a Profile will not route a question to an agent the user is not allowed to use. Admins can also control the rollout by granting a new Profile to a small test role, confirming that routing works, and then extending it to more people.

Dashboards you can talk to

Snowflake Dashboards are in public preview. An analyst builds a Dashboard in the Snowsight interface by describing it to Snowflake CoCo, then shares it with the right roles. Business users can open it in CoWork or, optionally, in Snowsight.

In CoWork, the Dashboard becomes something users can ask questions of. They can request a summary, ask which region drove a metric drop, or ask for a forecast of whether adoption growth will continue next quarter. According to Snowflake, the agent starts from the Dashboard as the user sees it, then queries the data behind it to answer.

Snowflake argues this changes the purpose of a Dashboard. Instead of being the point where analysis stops and a ticket to the data team begins, it becomes a shared starting point where anyone can ask the next question against the same governed data, without exporting it to another tool. Agents can query the underlying data, but admins have options to limit this. Row access and masking policies still restrict what each viewer sees. Analysts who want a more narrative format can use HTML reports.

Automations and the rest of the work agent

Automations let CoWork run recurring work on a schedule. A user describes the task once in plain language, for example a Monday 7 a.m. review of last week's sales pipeline that flags slipped deals and summarizes three accounts to call first. CoWork runs it on schedule and sends the result, and opening it returns the user to the full conversation for follow-up questions.

A scheduled run uses the same agent, scope and permissions as a question the user asks personally, so scheduling widens nothing, according to Snowflake. Automations run on Snowflake tasks, which data teams already operate. Some also take actions, such as creating a Jira ticket or sending a Slack message, and CoWork lists the operations each Automation is allowed to perform.

These features sit on top of the personal work agent, which also includes:

• User memory: keeps facts and preferences such as role, team or report formatting, which users can review and delete in settings.
• User skills: multistep tasks saved from chat and rerun with a slash command.
• Connectors: MCP connectors approved by the organization bring in context from tools such as Gmail and Google Drive.
• Code execution: code written and run in a sandbox to analyze data and prepare files.
• Document generation: output in PDF, PPTX, DOCX and HTML.
• Analytical search: answers questions spanning thousands of documents using Cortex Search, AI functions and SQL, and shows its plan before it runs.

Snowflake says all of these are generally available except analytical search, which is in preview.

Governance and what comes next

Snowflake stresses that each person sees only what they are allowed to see, in tables and now documents, because Cortex Search helps enforce the row access and masking policies on source tables. Admins decide what each agent can reach using monitoring and evaluation tools. Cortex AI Gateway, now in public preview, gives them one place to govern model access and spend. According to Snowflake, this applies whether a question comes from a person in chat, a Dashboard or a scheduled Automation.

The company describes Automations as the first step toward agents that work longer and more independently. Next, it says it is building agents that notice what changed, follow up without being asked and reach people wherever they work. Snowflake notes that statements about future offerings are not a commitment to deliver them.

For teams, Snowflake suggests starting by creating a first Automation, publishing a Dashboard and asking CoWork questions about it, and rolling out a Profile to a small test group first.