AI Copilot

AI Copilot is a chat assistant built into both the workflow builder and the journey builder. Instead of manually adding nodes and edges, you describe what you want in plain language, and Copilot proposes the change as a set of structured edits you can review, apply, and undo.

Copilot works from the graph you currently have open — it sees your nodes, edges, prompts, and settings automatically, so you don't need to paste anything in.

Opening Copilot#

Inside a workflow or a journey, click the Sparkles button floating in the bottom-right corner of the canvas to open the AI Copilot panel. The panel opens as a side drawer and is open by default when you land on the page.

The first time you open Copilot on a workflow or journey that doesn't have a chat history yet, it automatically reviews the current graph and posts its findings as suggestions — you don't have to ask for a review to get one.

Modes#

A mode selector at the bottom of the chat box controls how Copilot approaches your next message:

Copilot can also enter two modes on its own that aren't in the selector: Review, which it uses for the automatic first-open suggestions (and for the explicit review action described below), and — for workflows only — Setup, used during guided agent creation (see below).

Reviewing and applying changes#

When Copilot's reply includes proposed edits, it shows:

• Show Diff / Hide Diff — preview exactly what would change before committing to it
• Apply Changes — apply the edits to your canvas and save
• Undo — revert the last applied Copilot change
• Confirm risky change — required before Copilot can delete nodes or replace the whole graph; a confirmation dialog explains what's about to happen

Deleting nodes or replacing an entire workflow/journey always requires this explicit confirmation step — Copilot never removes nodes or overwrites your graph silently.

Copilot's replies also include a short row of suggested follow-up chips (e.g. "Add extraction for the caller's name and phone") tailored to what it just found or changed — click one instead of typing to continue the conversation.

Copilot for voice agent workflows#

Inside a voice agent workflow, Copilot can act as an AI workflow engineer for the whole graph or, if you have a node or edge selected on the canvas, for just that node or edge. Requests it can act on include, for example:

• Rewriting a node's prompt or greeting (tone, length, language)
• Adding, renaming, or deleting nodes (start, agent, end call, global, webhook, QA, or trigger nodes)
• Adding, editing, or removing edges and their transition conditions
• Turning on or updating variable extraction on a node
• Updating the Global Node prompt (agent name, company name, primary goal, tone, language)
• Updating template variables
• Renaming the workflow
• Generating a full replacement flow for the current agent, or generating a brand-new agent as a separate workflow — both ask for confirmation first

Some example prompts you can send it:

• "Review this workflow and suggest improvements"
• "Make the greeting warmer and more concise"
• "Add extraction for the caller's name and phone"
• "Simplify the flow — remove redundant nodes"

In Ask mode, try things like "What does this workflow do end-to-end?" or "Where are the riskiest failure points?" — Copilot will discuss and critique without changing anything.

Guided setup for new agents#

When you create a new voice agent, Copilot opens automatically in a guided setup flow instead of the general chat above. It asks you a short, fixed sequence of questions — your use case and agent name, whether the agent is inbound or outbound, what should happen on the call, and whether to capture any extraction variables — and fills in the canvas as you answer, without requiring you to click Apply for each step.

You can answer by typing, or start a live voice call with Copilot using the phone icon in the composer and answer the same questions out loud; your spoken answers are turned into the same setup steps. See Voice Agent Builder for more on creating a new agent.

Copilot for journeys#

Inside the journey builder, Copilot understands journey-specific node types — trigger, call (place a voice call through a Lemu voice agent), message (WhatsApp), condition (branch on a field/operator/value, with true/false paths), wait (delay), http (HTTP request), emit event, load history, schedule, slack_message, jira_issue, calendar_event (see Integrations for those three), and end. Requests it can act on include:

• Updating a node's configuration (e.g. a condition's field/operator, an HTTP request's URL, a WhatsApp template) or its label
• Adding, renaming, or deleting steps
• Adding or removing branches and connections, including the true/false paths out of a condition node
• Simplifying a flow by removing redundant steps or improving field mappings between steps
• Updating journey-level settings
• Replacing the entire journey with a new one (asks for confirmation first)

Some example prompts:

• "Review this journey and suggest improvements"
• "Add clearer error handling after failed steps"
• "Rename nodes so the flow is easier to follow"
• "Explain what this journey does end-to-end"

As with workflows, Ask mode ("Explain this journey step by step", "Where could this flow get stuck?") is discussion-only and never proposes edits.

Starting over#

Use the eraser icon at the top of the Copilot panel to clear the current conversation. This clears the chat history for that workflow or journey; it does not undo edits you've already applied to the canvas.