Test your assistant before customers do

Assistant tests let you check answers, keep useful questions, and re-run them after every change.

Before you startCopy Anchor Link

• Save your assistant first. Otherwise you see "Save your assistant settings first."
• Tests use your saved assistant configuration. Changes to settings or sources during a run require a new run.

How tests answerCopy Anchor Link

Question tests, Auto test, and Conversation & security tests use the same assistant answer process, with your saved model, instructions, enabled tools and Security Guard settings. AI grading runs separately from answer generation.

Document search follows your assistant's retrieval settings: it can look up relevant passages automatically and search again when the assistant requests more information. Only documents available to your assistant can appear in results. Read-only tools can look up authorized information from your company, such as products, services and public contact details.

Test actions and privacyCopy Anchor Link

Actions such as collecting a lead, submitting or updating an order, requesting human help and starting a flow are simulated in isolated test state. They do not change real customer records, send notifications or start real automations. Generated checkout links are test links, not usable purchase links. Tests do not load real customer leads, orders or bookings into their simulated customer state.

In conversation tests, simulated changes carry across messages within the same test conversation. Separate test conversations start with separate state.

Test conversations stay out of the Inbox and do not send messages to customers. Temporary processing state is cleaned up after each turn, and retry data expires automatically after 15 minutes. Your saved answers, sources, conversation transcripts and reviews remain in test history.

Open itCopy Anchor Link

Web: sidebar > Assistant > Assistant tests, or Test assistant from the assistant page. After saving the assistant, Try your assistant now offers an automatic test.

Auto test (one click)Copy Anchor Link

Auto test reuses up to 50 current questions, or lets AI create 10 if you have none, then answers them and reviews the answers automatically. It shows progress: Preparing questions, Running tests, AI is reviewing answers, Completed. Open View results when it finishes.

If any criteria were not yet approved, the result is marked Unapproved criteria: treat it as provisional.

Your own questionsCopy Anchor Link

1. Add question, or Redesign questions with AI to get drafts (up to 20 per run, in English, Persian, Turkish, Arabic or Spanish).
2. Drafts need approval: set the Expected answer / criteria, then Approve question (or Approve all).
3. Select approved questions (up to 50) and choose Run tests.
4. Read each result: Assistant answer, Retrieved sources, and the AI review with a verdict of Pass or Fail. Items the AI is unsure about show Needs human review.
5. To override the verdict, choose Admin review, pick the verdict and add an optional note.

Question statuses are Draft, Approved and Needs review (shown when related information changed; review the criteria and approve again). Use Archive to retire a question while keeping its past results.

Each run keeps its question revisions, assistant settings, source versions and answers. The latest 20 runs are listed in Run history.

Conversation and security testsCopy Anchor Link

Conversation & security tests check tone, brevity, greetings, natural replies, memory, and security behavior: staying within business scope and resisting prompt injection. You can try alternative Style instructions for a run without changing your live assistant, and compare Before and After results.

TipsCopy Anchor Link

• Add questions customers really ask, especially prices, policies and "what if" cases.
• Re-run after changing your assistant settings, knowledge documents or services.
• A "Retrieval review" shows whether the assistant found the right document passage.

TroubleshootingCopy Anchor Link

• "Assistant or sources changed during processing": start a new run or redesign.
• "The test could not complete": check your assistant is saved, then start a new run.
• A run is marked "Sources changed": its results used earlier information. Run the same questions again.

LimitationsCopy Anchor Link

• Each message uses your assistant model, and each conversation has a separate AI review, so large runs take time. Test creation is rate limited.
• A successful simulated action checks assistant behavior; it does not confirm a real payment, booking, notification or automation.
• These tests do not complete the public booking form or a real checkout.
• Not available as a dedicated screen in the mobile app.