Three ways to test
| Method | Where | What it runs | What it can't tell you |
|---|---|---|---|
| Test run | Canvas → More actions → Test run | Nothing. It traces a path through your steps on screen. No call is placed. | How the AI routes, what it says, whether tools work |
| Simulation chat | Finn page → Test → Simulation chat | A chat model given your flow as text instructions | Tools, voice, and how the call engine moves between steps |
| Phone call | Finn page → Test → Phone call | A real call with the saved canvas, tools included | Contact data from an audience: a test call has none |
Only a Phone call runs your flow the way a deployment does. Use the other two to catch mistakes cheaply first. Test calls need wallet credits. See testing for the Test menu in general.
Simulation chat and Phone call use the saved canvas, so click Save canvas before testing an edit. Test run uses whatever is on the canvas, saved or not.
Before you test: the issues list
The canvas toolbar shows OK or a count of issues. Clear the list first. A dead end, an unreachable step, a step with no prompt, or two unlabelled paths out of one step will all misbehave on a call, and the canvas doesn't stop you saving them. The messages are listed on workflow editor.
Test run
Test run starts at the step connected to Start and follows the first transition out of each step until it reaches End Call, a step with no transitions, or a step it has already visited. It shows the first part of each step's instruction prompt.
Use it to check that Start is connected and the main path reaches an end. It ignores transition conditions, so it only ever shows one path.
Simulation chat
Simulation chat gives a chat model your Finn's identity, guidelines and the canvas written out step by step, and you play the caller by typing.
Use it to check wording and whether your transition conditions make sense. Keep its limits in mind:
- Tools don't run. Calendar checks, bookings, transfers and API calls don't happen.
- It doesn't use the call engine that moves between steps on a real call, so a path that works in chat can still route differently on the phone.
- On a large canvas, some steps may be left out to fit the model's context.
Phone call
A Phone call test runs the saved canvas on a real call: step routing, tools, the filler lines, End Call. This is the only test that tells you a flow works.
Go through every path, not only the cooperative one:
| Check | How |
|---|---|
| Every transition | Give the answer that should trigger each condition, once per test call if you need to |
| The fallback | Give answers that match no condition. After a few turns the Finn should move on to the transition you left without a condition. |
| Off-topic questions | Ask something unrelated and check the step's Steer back brings the call back |
| Each tool | Check Calendar returns real slots, a booking appears in Cal.com, a transfer rings the destination, your endpoint receives the request |
| A slow or failing endpoint | Point a tool at an endpoint that times out, and listen for how long the caller waits and what the Finn says |
| Every End Call | Reach each one and check the goodbye message |
A test call has no audience, so {{column}} placeholders in the welcome message aren't filled from a CSV. To check personalisation, run a small deployment to your own number with an audience file that contains your row. See workflow variables.
Reading the result
Test calls appear in Analytics with real calls. Open the call for the transcript and audio. See call logs.
When a call goes wrong, find the first turn where it went off course and check the step the Finn was on:
| Symptom | Likely cause |
|---|---|
| Finn moved to the wrong step | Two transition conditions overlap, or a condition is too vague |
| Finn kept asking the same thing, then moved on | No condition matched the caller's reply. Rewrite the conditions in the caller's terms. |
| Finn moved on before getting an answer | The step's only transition has no condition, so it moves on as soon as the step's goal looks done |
| A tool didn't fire | It's attached to a different step, or the step prompt doesn't say when to use it |
| Finn hung up too soon | A step's only exit leads to End Call, so the Finn ends the call on the caller's next reply |
| Finn paraphrased a line that must be exact | Mark it verbatim. See workflow nodes. |
After testing
Saving overwrites the canvas and there's no history. Once a flow tests well, use Export JSON to keep a copy before you edit it again.
Related
- Workflow editor
- Workflow nodes
- Testing: the Test menu, phone number format, wallet credits
- Evaluations