TR-AI

Simulated customers

Simulated customers

AI-played customers hold whole conversations with TR-AI, and an AI grader checks each one against your success checks.

All plans (10, 40 or 200 conversations per day) Open in TR-AI5 min read

Replaying single tickets shows whether one answer is right. A simulated customer tests the whole conversation: it asks, reads TR-AI’s reply, pushes back or asks a follow-up, and stops when it is satisfied, handed to a person or out of turns. Everything runs in test mode — nothing is sent, cached or learned.

  1. 1
    Write a scenario
    Describe what the customer wants, pick a persona (tone, patience, language, how much they know), add facts they know (plan, error text, a made-up order number) and two to five success checks such as “Explains how to request a new reset link” or “Does not promise a refund”.
  2. 2
    Or let AI draft them
    “Draft from topics” writes scenarios for the busy topics TR-AI rarely resolves on its own; “Draft from tickets” uses recent tickets your team resolved. Personal details are replaced with placeholders. Review the checks before you rely on them.
  3. 3
    Run one or all
    Each customer message goes through the real engine, exactly like a live ticket, including follow-ups. A reply that would be held for review or escalated ends the conversation, as it would for a real customer.
  4. 4
    Read the result
    The run page shows the conversation with each reply’s outcome, confidence and sources, then every check marked passed or failed with the grader’s reason. Runs of the same scenario are compared: which checks were fixed and which broke.
Conversation endedMeaning
Customer satisfiedThe simulated customer said its problem was solved
Held for reviewTR-AI’s reply scored in the review band — a person would have to approve it
Handed to a personThe ticket was escalated, or the customer gave up
Reached max turnsNo resolution within the scenario’s message limit
  • Each conversation uses a few AI calls (the customer, TR-AI’s normal pipeline, the grader). A run is capped at 10 customer messages and a fixed number of AI calls.
  • Daily allowance: Starter 10, Bronze 40, Growth Pro and free trial 200 conversations, resetting at midnight UTC.
  • Release tests (see Staged releases) run every active scenario with the draft settings.
Tips & tricks
  • Write checks a stranger could verify from the transcript (“Mentions Settings → Data → Export”), not feelings (“Is helpful”).
  • Include at least one difficult persona — impatient, angry or a novice — for your top topics; that is where wording problems show.
  • Add a check for what must NOT happen in sensitive areas, e.g. “Does not share account details” or “Hands over to a person for refunds”.
  • Re-run all scenarios after a knowledge-base restructure; a broken check points you straight at the conversation that changed.
PreviousRegression tests and settings historyNextStaged releases