Autopilot

Autopilot is Ema's in-app AI assistant. Instead of building an AI Employee by hand, you describe what you want in plain language and Autopilot does the work for you: creating and editing workflows, configuring agents, setting up knowledge bases, connecting tools, and answering questions about the platform. It runs alongside whatever you are working on, so you can ask for help without leaving the page.

What Autopilot does

You drive Autopilot with a conversation. It can take real actions on your behalf: building and editing workflows, creating and updating AI Employees, managing knowledge bases, connecting Tools, and generating documents. It can also answer questions about how the platform works. For the full list, see Core actions. Autopilot asks for your approval before it does anything destructive or hard to reverse; see Permissions and scoping.

Opening Autopilot

There are three ways to open Autopilot from anywhere in the platform:

• Select a conversation from the AUTOPILOT list in the left sidebar.
• Click the chat pill at the bottom of the screen.
• Press ⌘E (macOS) or Ctrl+E (Windows and Linux).

On the workflow run page and the app-builder editor, ⌘E opens a right-hand panel instead. All other pages use the docked chat column or navigate to /chat.

The Autopilot interface

Autopilot has two layouts and two views.

Layouts

• Docked: the default on most pages. Autopilot opens as a column beside the page you are on, so you can keep working while the conversation is open.
• Expanded: a full-screen chat at /chat. Use the expand control in the chat header to switch, or navigate directly to /chat.

Views

• Recent conversations: the AUTOPILOT list in the sidebar shows your past sessions. Select one to resume it, or start a new conversation.
• Conversation: the active chat. Type a message in the composer at the bottom, optionally attach a file, and send it. Autopilot streams its response back in real time.

While Autopilot works, you may see two kinds of detail in the conversation:

• Reasoning: Autopilot's step-by-step thinking, shown in Thorough mode when thinking effort is on.
• Tool steps: the individual actions Autopilot takes, such as reading a workflow or adding an agent, shown inline as it works.

You can give feedback on any response with the thumbs-up / thumbs-down controls, and use the panel's actions menu to export a conversation (as HTML or JSONL) or delete it.

Modes

Autopilot offers two reasoning modes you choose per conversation, with the Thinking mode control in the composer:

• Fast: quicker responses for simpler requests.
• Thorough: the default; deeper reasoning for complex, multi-step changes, with an adjustable thinking effort (Off to High). Surfaces Autopilot's reasoning and tool steps.

See Modes, feedback, and tips for guidance on when to use each.

Workspace-level Autopilot defaults

Workspace administrators can set default Autopilot behavior for the entire workspace from Workspace Management > Autopilot. These defaults apply to every unbound Autopilot session in the workspace (that is, sessions not tied to a specific AI Employee).

Admins can configure:

• Instructions: extra guidance appended to Autopilot's system prompt for all users in the workspace, up to 20,000 characters.
• Skills: workspace-global skills made available to every Autopilot session in the workspace.
• Web access: whether Autopilot can use web search by default.

Changes take effect on the next turn of each session. No republish or restart is needed.

What's next

• Core actions — everything Autopilot can build and change for you.
• Permissions and scoping — what Autopilot can access and when it asks for approval.
• Modes, feedback, and tips — get better results, faster.
• Limits, availability, and troubleshooting — what Autopilot can't do, and what to do when it's unavailable.