I changed my AI Employee. What should I do before relying on the change?
Run a short test on the most important customer questions, then focus on the area you changed. If a reply is wrong or weak, fix the source that owns that behavior and start a fresh test. The practical loop is: change → test → fix the cause → retest → then rely on it with customers.
- Test after every important change, even when saving succeeded.
- Write questions like a real customer instead of feeding the expected answer into the prompt.
- Evaluate the reply actually returned by Mr. AI Testing.
A quick test after any important change
- Ask about one core Product or Service, then an important price, offer, schedule, or policy.
- Ask for the same meaning in different words, then include a wrong assumption and check whether the AI corrects it.
- Continue for several turns and verify that context is preserved without inventing missing customer data.
- Give a realistic objection and see whether the conversation continues naturally.
- Ask for information that is not available and verify that the AI does not fabricate it.
- Ask for a human agent while checking tone, language, business name, and identity.
- If you use Conversation Stages, run a full journey and make sure the stage does not advance too early.
- If you use an Action, test when the AI should and should not try to trigger it.
If you changed identity or speaking behavior
- Business name and AI Employee role.
- Tone, language, and speaking style.
- Things the AI must not say or do.
- How it handles a question whose answer is unknown.
- When it should suggest or request human help.
If you changed Catalog
- Ask for a product or service by name, then verify price, currency, and Options / Variants.
- Ask a natural discovery question without repeating the item name.
- Test Semantic Classification and Semantic Tags with questions that match the meaning behind the structure.
- Test AI Search Attributes with a precise question about Instructor, Branch, Level, or another exact value.
- When you use Aliases or Search Name, ask for the same value using different wording or language and confirm Mr. AI still reaches the same correct item.
- Test two similar items so similarity alone does not return the wrong product.
If you changed Deep Knowledge
- Ask for one direct fact.
- Ask for the same meaning using different wording.
- When information is distributed, ask a question that requires combining it.
- Ask for something that is not present and check for non-fabrication.
- After updating the Resource, start a fresh test and ask again.
If you changed Conversation Stages
- Start the journey from the beginning and continue for multiple turns.
- Check that stage transitions happen at the right time, not too early.
- Take a small detour and then return to the main journey.
- Confirm the conversation reaches the correct final state.
If you changed Actions
- Test the case that should trigger the Action.
- Test the same journey before required information is complete and confirm it does not trigger too early.
- Try missing or invalid data.
- Review what the AI would attempt if the conversation were Live.
If you changed Follow-up or Stage Follow-up
Test whether the conversation reaches the right stage and context for follow-up. A Test Chat itself does not prove that a real Follow-up or Task was scheduled or executed in Production.
When you find a problem: fix the right source
- 1
Keep the customer question and the reply that exposed the problem.
- 2
Identify the owning source: Catalog, Deep Knowledge, identity and speaking behavior, Conversation Stages, or Action.
- 3
Change the smallest responsible source instead of adding one broad instruction to hide the symptom.
- 4
Start a fresh Testing Chat or Scenario after the important change.
- 5
Repeat the same question, then try nearby wording to make sure the fix is not overfitted to one sentence.
- 6
After the core tests pass, watch the first real conversations to verify quality in use.
Test manually with Test Assistant
Test with ChatGPT or a connected AI assistant
Understand the AI Employee quality loop
Understand grounded AI replies
