Product documentation

Persona simulation

Stress-test a draft assistant by auto-running a batch of conversations in the voice of a tricky customer persona, scored on completion, fallback rate, and off-topic recovery.

Persona simulation answers a simple question before you publish: if 50 difficult customers talked to this draft right now, how many would get stuck? Instead of hand-typing a few test messages, it auto-runs a batch of conversations in the voice of a tricky customer persona and scores how your flow holds up.

When to use it

Run it on a draft flow before publishing, and again after every fix. The accumulated run history lets you confirm a change actually moved the score in the right direction.

The personas

Real customers don't talk like the happy-path script you built the flow against. You pick one persona per run:

PersonaBehaves like
ImpatientTerse, demanding, wants the answer immediately with no pleasantries.
Typo-proneTypes fast and makes frequent spelling and spacing mistakes.
Non-native speakerPolite, simple grammar, occasional misphrasing, sometimes asks you to repeat.
ElderlyPolite and a little verbose, less familiar with technology, asks for reassurance.

Run a stress-test

  1. Open the assistant's Flow builder.
  2. Make sure your latest changes are saved as a draft.
  3. Click Stress-test in the top bar.
  4. Pick a persona and how many conversations to simulate (10, 25, or 50).
  5. Click Run stress-test — the batch runs inline and scores appear in a few seconds.

Each run is saved to the history table at the bottom of the panel, so you can compare a persona's scores across edits.

Reading the scores

ScoreMeaningDirection
CompletionShare of conversations that reached the end of the flow. A low number means customers get stuck partway.Higher is better
Fallback rateShare of conversations where the assistant didn't understand and had to fall back — usually a sign intents or prompts need more coverage.Lower is better
Off-topic recoveryWhen a customer wanders off-topic mid-conversation, how often the assistant still gets them back on track and finishes.Higher is better

Scores are colour-coded — green is healthy, amber is worth a look, red needs attention — so you can scan a run at a glance.

Tips

  • Run every persona before you publish — a flow can score well for a patient customer and badly for an impatient or typo-prone one.
  • Re-run after each fix and use the history table to confirm the score moved in the right direction.
  • Start small: a 10-conversation run is quick for a first pass; bump it to 50 once the flow is close to ready.
  • Pair it with the live chat simulator — investigate a specific failure there, then re-run the persona batch to confirm the fix holds across many conversations.

What it does not do

Persona simulation tests your draft flow, not the published one — the whole point is to catch problems before they go live. It exercises the conversation logic with synthetic customers; it does not place real orders, call live tools against production systems, or contact real customers.