Analysis Agent
Start chats with Sundial's main agent to ask lookup and deep-analysis questions with clear trust cues.
The Analysis Agent is the main agent in Sundial and the primary surface where users interact. You ask questions in natural language; it answers using your semantic layer, playbooks, and other context so results stay consistent with how your company defines metrics and analyses.
Where to find it
Start a conversation by clicking New chat in the sidebar and selecting Analysis Agent. Analysis Agent is selected by default.
Outputs the agent generates (reports, dashboards, and more) are collected in the Library.
Types of questions
Lookup questions
Lookup questions ask for a number or a simple chart over time, for example workspaces by region, or a metric trend. These usually resolve through the semantic layer: the agent uses governed metric definitions instead of writing one-off warehouse SQL.
Deeper analyses
More complex questions (why a metric changed, how an onboarding experiment performed, where a funnel drops off, which early behaviors predict retention) are typically handled by playbooks. The agent matches your question to a relevant playbook and runs that analysis flowchart against your data.
For example, if you notice ARR spiked on a particular day and ask "what drove the increase in ARR?", the agent recognizes a common metric-change question, loads the matching playbook, and works through it, decomposing the change into price vs. quantity and finding the segments that explain it, rather than improvising a one-off approach.
Trust cues in the answer
When the agent responds, look for:
- Chain of reasoning: the steps it took, including whether it used the semantic layer.
- Confidence badge: a trust signal (high, medium, low, or unverified) reflecting how the answer was produced. High confidence means the answer is grounded in governed context (for example a verified metric) rather than ad-hoc SQL; lower signals are a cue to double-check the result.
- Supporting detail: for semantic metrics, the agent shows what definition supported the statement; when it must run a custom query, it can surface that query so you can inspect it.
See Observability for how these signals are assessed and how the data team acts on them.