What Is a Dashboard Agent?

A dashboard agent is a conversational AI assistant that answers follow-up questions about a dashboard and helps a viewer explore its data. It can query governed metrics behind the charts and, when supported controls are present, change the viewer's filters or time granularity. The useful distinction is that a dashboard agent works with the data model behind the dashboard, not only the pixels already on screen.

What a dashboard agent does

A dashboard is good at making known questions easy to revisit: monitor a metric, compare a few segments, or scan the latest operating numbers. The next question is often more specific than the view the author prepared. A dashboard agent gives the viewer a way to ask that question next to the chart instead of leaving the dashboard and starting over.

For example, someone looking at monthly orders could ask why the latest month changed, break the number down by region, or compare a longer period. The agent interprets the request, chooses metrics and dimensions from the analytics model, runs a query, and explains the result. This is a focused form of agentic analytics: the agent takes analytical steps, while the dashboard supplies useful context for the conversation.

The agent should not simply rewrite a chart title or summarize the numbers already visible. A dashboard is a useful starting point, but it is not always enough to answer a follow-up. A well-grounded agent can query the underlying model for a new grouping or time range, then return an answer tied to business definitions the team maintains.

How it works with governed metrics

Dashboard agents are only as reliable as the definitions behind their queries. Pointed at raw tables, an AI model has to guess which column means revenue, how tables join, and which rows the viewer may access. Those guesses can change from prompt to prompt, even when the question stays the same.

A semantic layer gives the agent a governed map of metrics, dimensions, entities, joins, and access rules. The viewer can ask for “revenue by region,” while the model supplies the approved meaning of revenue, the valid grouping, and the access policy that applies to that person. The warehouse remains the storage and compute layer; the semantic layer provides the business meaning and compiles the request.

That architecture also makes permissions part of the query path instead of an instruction in a prompt. For customer-facing dashboards, the host application can pass the viewer's tenant and role context so the same model returns only the rows that person is allowed to see. The production requirements for AI agents for data analysis apply here too: the answer should be consistent, permission-aware, and traceable to the metric and filters that produced it.

In Cube, the Dashboard Agent answers questions using live queries against the semantic model behind a published or embedded dashboard. It can drill into a metric, break it down by a dimension, examine a longer time range, summarize trends, explain a metric or dimension, and look up available dimension values. That means a viewer can ask about the business question behind a chart, rather than being limited to its current visualization.

Change the current view without changing the shared dashboard

A viewer may want to filter to the last 30 days, group a chart by month, or switch a field from revenue to profit. The agent can apply these requests to the controls already configured on the dashboard. It can choose among allowed values, but it cannot invent a new filter option or control that the author did not provide.

These changes apply to the viewer's session. They do not save new defaults to the published dashboard or change what another person sees. The resulting view is reflected in the URL, so the viewer can bookmark or share that state. If the dashboard has no matching control, the agent explains that it cannot make the requested change.

This boundary is useful for both viewers and authors. People can explore without making a shared report unpredictable, and authors retain control over which dimensions and values are available. A dashboard agent should not be described as a dashboard builder: the published agent does not edit the saved dashboard, add or remove widgets, create workbooks, or change the data model.

Dashboard Agent or Workbook Agent?

The right agent depends on whether the person is viewing a published dashboard or creating analytics content. A dashboard agent is for a viewer who wants to understand a chart, ask a follow-up, or change their own dashboard controls. A workbook agent supports authoring work: building reports, creating and editing dashboards, and doing analysis inside a workbook.

Keeping those jobs distinct makes permissions and expectations clearer. A viewer can explore without accidentally changing a shared artifact; an author can work in a space designed for creating and reviewing analytics. Teams that manage report definitions as code can also use dashboard as code with Cube to version and deploy saved dashboards through a reviewable workflow.

Internal and embedded dashboard agents

Dashboard agents are useful for internal BI, where employees want to investigate a metric without waiting for an analyst to prepare another view. They also fit embedded analytics, where a customer's next question comes up inside the product that already contains the dashboard. The experience changes with the audience, but both need the same foundations: a governed model, fast queries, and permissions that follow the viewer.

The embedded agentic analytics guide covers the additional work required for customer-facing AI, including tenant isolation and production performance. Cube's agentic analytics platform brings dashboards, workbooks, Analytics Chat, and embedded surfaces together around the semantic layer. The same definitions can support employees inside the company and customers using analytics within a product.

The tradeoff is modeling investment. Teams need to define metrics and access rules before an agent can answer with consistency. That work is not unique to conversational interfaces; it is what lets the dashboard, workbook, and agent agree on what a number means.

For current capabilities and deployment-specific availability, see the Dashboard Agent documentation. Chart-level follow-ups and attached-chart context may roll out separately, so check the documentation for what is available in a given deployment.