Pre-launch stress-test script for an AI customer-service bot

You are a quality lead for a customer-service system. My AI customer service is about to launch — help me write a pre-launch stress-test script.

Service scope: (e.g., order lookup, returns/exchange, product consultation)
Knowledge-base coverage: (e.g., 120 FAQ articles)
What it shouldn't handle: (e.g., can't promise compensation amounts, can't process personal-data changes)
Audience traits: (e.g., wide age range, messages often have typos or use bopomofo)

Output:
1. 30 normal-scenario test questions covering the highest-frequency question types you think of
2. 20 "tricky but reasonable" questions: asking three things at once, self-contradictory, incomplete info, emotionally charged
3. 15 out-of-scope questions to test whether it makes wild promises or answers
4. 10 Taiwan-user-specific expressions (typos, bopomofo text, direct Taiwanese translation, abbreviations)
5. Scoring criteria per category: what answer counts as pass, as fail, and as needing a human
6. Suggested human-handoff trigger conditions (specific enough to write as rules)

Output as a table with columns: number / category / test question / expected behavior / failure mode.