What the workflow editor is
The workflow editor is a canvas where you draw the call flow as nodes and edges. The Finn follows that graph instead of deciding the order of the conversation itself.
Most Finns do not need one. Free-form mode (identity, prompting, knowledge base) handles the majority of use cases and is easier to maintain. Build a workflow when you have a concrete reason to enforce order: a compliance script that must be read verbatim, a step that must happen before another step, or clear decision points where the caller's answer changes the path.
The signal that you need one: you keep writing "you must do X before Y" into the prompt and the Finn keeps getting it wrong.
Opening the editor
The editor is a dashboard surface. There is no API for building or editing workflows.
From the sidebar: View All Finn → click the Finn → Edit Workflow. The editor opens full-screen with the canvas in the middle, the node palette on the left, a mini-map bottom-right, and Save / Test buttons top-right.
A new workflow starts with a single Start node. Everything else you drag in.
Canvas controls
| Action | How |
|---|---|
| Add a node | Click and drag from the left toolbar onto the canvas |
| Move a node | Click and drag the node |
| Connect two nodes | Drag from one node's edge dot to another node |
| Delete a node or edge | Right-click → Delete, or select and press Delete |
| Undo | ⌘/Ctrl + Z |
| Save | ⌘/Ctrl + S, or the Save button |
| Reset zoom | Click "Fit to screen" |
Autosave is off by default. Nothing you draw is committed until you save. Close the tab mid-edit and the changes are gone. This is deliberate, so a half-finished edit cannot overwrite a working flow, but it means an unsaved session is a real way to lose work.
Building a flow
Six node types exist: Start, Say, Ask, Decision, Action, End. Full field-by-field detail is on workflow nodes.
The shape most flows take:
- Start — one per workflow. The call enters here. Set an opening message on the node, or leave it blank and let the welcome message handle it.
- Ask nodes to collect what you need. Each Ask saves its answer to a named variable.
- Decision nodes at the points where the answer changes the path.
- Action nodes for side effects: book a calendar slot, transfer the call, write to CRM, send SMS, call a webhook.
- End nodes, one per outcome. Multiple End nodes are worth the effort because each becomes its own category in call outcomes.
A worked example: an urgent-care intake Finn greets, asks name and date of birth, asks symptoms, asks severity 1-10, then branches. Severity 8 or higher transfers to the on-call doctor. Severity 4-7 books same-day. Below 4 books a regular appointment. All three paths converge on a Say node that reads the disclosure, then End. The Finn cannot skip the disclosure or book before checking severity, because the graph does not allow it.
Data moving through the flow
Ask nodes and Action nodes produce variables that persist for the whole call. Reference them with double curly braces anywhere text is accepted: Say text, Ask question text, Action parameters, Decision conditions.
{{contact.name}}comes from the audience row the call is for{{state.severity}}comes from the Ask node that saved toseverity{{state.appointment_id}}comes from a Book Calendar action
If a variable name does not match, the reference is not substituted and the literal {{state.x}} string ends up in what the Finn says. See workflow variables.
Test as you build
Do not design the whole graph before testing any of it. Build a small section, test it, extend it.
Test on Chat opens a text simulator in the browser. Each node shows its inputs and outputs as the run proceeds, so you can see which branch fired and why.
Test on Call dials your phone for an end-to-end run. Use this before pointing real traffic at the workflow. More in workflow testing.
What goes wrong
Callers say things your Ask node cannot parse. This is the most common failure and it strands the call. Add a fallback edge to every Ask node so there is somewhere to go when parsing fails. Switching the node to AI-evaluated also lets the model interpret paraphrased answers.
The Finn sounds robotic. Long chains of Say nodes leave no room for conversation. Use the End node's hand-off-to-free-form option so the Finn exits the workflow and continues conversationally on its normal identity and knowledge base.
The editor gets sluggish. Flows past roughly 30 nodes start to lag. Group sub-flows into Group nodes (Pro plan), or split the flow across multiple Finns joined by transfer.
Saving, versions, and export
Every save creates a new version. New deployments pick up the current version. Deployments already running stay on the version they launched against. Roll back through Edit Workflow → History → Restore, which makes the restored version current. See workflow versioning.
Workflows export to JSON through the ⋯ menu → Export as JSON, and import the same way. Use it to copy a flow between Finns or accounts, or to keep the JSON under version control outside the platform.