docs: add Autopilot AI agent guide

docs.json

@@ -145,7 +145,8 @@
                   "guides/ai-agents/evaluations",
                   "guides/ai-agents/data-access",
                   "guides/ai-agents/best-practices",
-                  "guides/ai-agents/self-improvement"
+                  "guides/ai-agents/self-improvement",
+                  "guides/ai-agents/autopilot"
                 ]
               },
               "references/integrations/lightdash-mcp",

guides/ai-agents/autopilot.mdx

@@ -0,0 +1,84 @@
+---
+title: "Autopilot"
+sidebarTitle: "Autopilot"
+description: "Scheduled AI agent that keeps your project clean — fixes broken charts, flags stale content, and suggests new ones."
+---
+
+<Info>
+  **Availability:** Autopilot is a [Beta](/references/workspace/feature-maturity-levels) feature available on **Enterprise plans** only.
+</Info>
+
+Autopilot is a scheduled agent that runs against a project on its own and keeps it healthy. Each run, it reviews recent activity, finds problems, fixes what it safely can, and flags the rest for review.
+
+<Frame>
+  <img
+    src="/images/guides/ai-agents/autopilot-home-card.png"
+    alt="Autopilot card on the project home page showing run summary"
+  />
+</Frame>
+
+Unlike conversational AI agents that respond to user questions, Autopilot is autonomous — it works in the background on a schedule you set.
+
+## What Autopilot does
+
+Each run, Autopilot works through a fixed checklist:
+
+| Step | What it does |
+|------|--------------|
+| **Stale content** | Soft-deletes charts/dashboards with 0 views older than 30 days. Flags content not viewed in 3+ months. |
+| **Broken content** | Finds charts referencing fields that no longer exist. Tries to auto-fix by removing invalid references. Flags any it can't fix. |
+| **Slow queries** | Flags charts and dashboards with consistently slow warehouse queries (>2s). |
+| **Preview cleanup** | Flags preview projects older than 3 months. |
+| **Chart suggestions** | Reviews recent user questions to the AI assistant. If users repeatedly ask about a topic with no saved chart, Autopilot drafts one in an "Agent Suggestions" space for admin review. |
+| **Insights** | Surfaces popular content that isn't pinned, or popular content stuck in a private space. |
+
+Every action is recorded in the activity log with a description of *why* Autopilot did it.
+
+## Action types
+
+Each row in the activity log is tagged with one of:
+
+- **Auto-fixed** — broken chart Autopilot fixed itself
+- **Cleaned up** — content soft-deleted (admins can restore)
+- **Suggested** — new chart drafted for review
+- **Flagged stale / broken / slow** — recorded for review; future runs may escalate (e.g. soft-delete a stale flag that's been ignored for 24h+)
+- **Insight** — observation about popular content
+
+## Safety rules
+
+Autopilot is conservative by default:
+
+- Never touches content less than 30 days old
+- Never soft-deletes a chart that's the only chart on a dashboard
+- Never modifies content it created itself (slugs prefixed `agent-`)
+- Prefers flagging over deleting when in doubt
+- Soft-delete only — admins can restore from the trash
+- Reverses its own actions if it detects a mistake on the next run
+
+## Setup
+
+Autopilot is configured per project from the project Home page.
+
+**Schedule** — choose how often it runs (every 6 hours, daily, etc.).
+
+**Permissions** — three capability toggles:
+
+- **Read content** *(always on)* — reads project health signals, recent actions, and chart definitions
+- **Create content** — drafts new chart suggestions in the "Agent Suggestions" space
+- **Modify existing content** — soft-deletes stale content, fixes broken charts, reverses its own past actions
+
+Disable both modify capabilities to run Autopilot in read-only mode (flags and insights only).
+
+**Slack updates** *(optional)* — post a summary of each run to a Slack channel. Requires the Slack integration.
+
+## Reviewing activity
+
+The Autopilot activity page shows every action across runs, with a sidebar that surfaces:
+
+- Affected chart or dashboard context (space, last modified, views, verification status)
+- Validation error details for broken content
+- Execution time for slow queries
+- Chart-as-code definition for suggested charts
+- Autopilot's reasoning for the action
+
+Use the activity log to spot-check decisions, restore anything Autopilot got wrong, and accept or dismiss suggestions.