AI Flow Builder

The AI Flow Builder is an optional panel that sits on top of the Flow Designer. It lets you describe what the flow should do in natural language and applies the result to the canvas — useful when you want to draft a new flow from scratch, restructure several nodes at once, or convert a written brief into a concrete graph.

⚑Feature required: aiFlowBuilder

When this flag is off the AI panel is hidden completely. The Flow Designer still works fully — manual canvas edits and the source view stay available.

What it does

• Drafts a flow from a written brief ("a German appointment-booking line that transfers to Office for high-value callers and hangs up on abuse").
• Edits the current flow — add a fallback, rename an agent, attach a tool, add a webhook.
• Migrates older flows — when launching from /apps/new with an existing legacy scenario, the panel can read the v1 definition and propose a current Flow Designer graph.
• Iterates — targeted edits update the requested part of the flow while preserving unrelated provider, tool and layout settings. Review the preview before saving; checkpoints and version history remain available.
• Checkpoints work — v0.9.13 saves AI checkpoints so you can return to a previous assistant state while editing.
• Retries failed turns — transient provider or network failures can be retried without forcing a full page refresh.
• Tracks long-running drafts — background runs keep polling, show elapsed time, and render structured JSON replies in a readable format.

The panel writes the same JSON that the manual UI and source view produce, so a flow built by the AI is indistinguishable afterwards from one built by hand.

Screenshotguide/users/flow-designer/ai-panelneeds_review

AI Flow Builder panel above the canvas

Highlight: briefing chat with a few turns of conversation, the resulting graph rendered on the canvas below

Reason: Patch2 targeted assistant edits and publish-time tool validation require visual review.

Workflow

1. Open the panel from the toolbar (or it auto-opens for a freshly created app).
2. Brief it — a paragraph or two describing the goal, channels, languages, key constraints. The model uses the team's variables, tools, and inline provider configuration as context.
3. Review — the panel writes a draft definition; the canvas updates immediately.
4. Iterate — ask follow-ups ("add an end node when the caller says they're done"), edit individual nodes manually, or both.
5. Use checkpoints — if an iteration goes in the wrong direction, restore a previous AI checkpoint or use Version History to return the draft to an earlier definition.
6. Test — run a draft test call from the toolbar (see Draft and publish).
7. Publish when satisfied.

When to drop into manual editing

The AI panel is good at the structure of a flow but it is still a layer on top. For precise tweaks the manual canvas and inspectors are faster:

• Renaming a single transition.
• Pinning a model or voice on a specific agent.
• Wiring a complex condition with multiple decision rows.
• Editing tool JSON-schemas.

You can switch back and forth — the AI panel re-reads the current definition every time it runs.

Privacy and providers

The AI Flow Builder runs against the same LLM provider configured by the platform admin. It does not have access to call recordings or production data — it only sees the current flow definition, your brief, and the team's tool / variable catalogue.

See also

• Flow Designer overview
• Source view (flow definition JSON)
• Draft and publish
• Feature flags

Tool validation before publishing

Generated and pasted tool code is validated against the sandbox API used by the editor. Correct reported errors before publishing; drafts remain editable. For keypad collection, read result.digits, result.valid and result.reason, not result.collected. Tool code must be executable JavaScript, without TypeScript-only syntax. Test the corrected draft before publishing it.