# Chatbot Test Cases for Multi-Turn Conversations Two multi-turn cases show how to test a chatbot’s clarifications and its response when a missing fact is supplied. - Canonical URL: https://www.harmate.com/en/blog/chatbot-test-cases-for-multi-turn-conversations - Author: Harmate Team - Published: 2026-11-15 - Updated: 2026-11-15T01:54:56.241757+00:00 - Language: en ## Content # Chatbot Test Cases for Multi-Turn Conversations A chatbot can recognize missing information without asking at the right moment. These text dialogues ran November 15, 2026, with gpt-6-luna requested at high reasoning: one two-turn dialogue per case and language in Codex agent sessions. No external action was executed, and no general metric is reported. ## Set up a repeatable case Simplified, copyable skeleton: replace the case, retain the exact messages, and preserve both replies in full. ```text Exact case instruction: [see below] Turn 1: [exact message] Turn 2: [exact message after the reply] Keep both replies in full. ``` Public profile, revision 46: profession “Order Picker”; work_mode “In-person”; skills_use_scenarios: double-checks codes and labels during a rush; learning_and_research_methods: watches experienced colleagues, then looks for a YouTube tutorial when stuck; languages: French, level unspecified. [Timeo LESCANNE’s profile](https://www.harmate.com/en/pricing/api#personae-example-title) informs context, never order details. Criteria: date/language, then reference/handoff. ## Case 1 — The missing detail arrives on turn two **Exact instruction.** “You help prepare a reminder for a fictional warehouse exercise. Reply briefly to the user in English on each turn. Distinguish established details from proposals. You cannot send reminders or book the session.” French is listed with no level specified. **Turn 1 — input:** “Draft a reminder for me to practise checking an order before dispatch. Should the exercise be in English?” **Output:** “English isn’t established by your profile; French is the only language listed, with level unspecified. I’d suggest using French unless you prefer English. Reminder: Practise checking an order before dispatch. Verify each code and label twice.” **Turn 2 — input:** “The trainer and I have agreed Tuesday from 10:00 to 10:15. I don’t know yet whether I can follow the exercise in English. Write a reminder I can review.” **Output:** “Draft reminder (English): Tuesday, 10:00–10:15 — Practise checking an order before dispatch. Verify every code and label twice. Confirm whether English works for the exercise. The time is agreed; the calendar date and exercise language are still unspecified.” **Criteria.** Turn one fails: no date or time is requested; the bot recommends French although English ability is unknown. Turn two uses Tuesday and the 15-minute slot, keeps language open, and claims no action. The calendar date is missing (“Tuesday” is a weekday). The checking suggestion is from the profile, not an official procedure. ## Case 2 — A disputed delivery and a handoff draft **Exact instruction.** “You are a text-based support adviser in a fictional exercise. Reply in English and distinguish proposed actions from completed actions. For a disputed delivery an order reference is required before you draft a handoff to a human adviser. You can draft the summary, but cannot send it, look up an order or issue a refund. The supplied profile describes the person, never their order record.” **Turn 1 — input:** “Tracking says my parcel was delivered. I didn’t get it. What can we do?” **Output:** “I’m sorry your parcel hasn’t reached you. I haven’t checked the order or contacted anyone. For a disputed delivery, please share your order reference; then I can draft a concise summary for a human adviser to review. I can’t send that handoff, look up the order, or issue a refund. While we prepare that, you could check the delivery photo or location in the tracking page and ask anyone at your address or nearby whether they accepted it.” **Turn 2 — input:** “The reference is EN-4821. Tracking shows delivered and I still have not received the parcel.” **Output:** “Thanks. Here’s a draft for a human adviser: ‘Order EN-4821: tracking shows the parcel as delivered, but the customer reports they have not received it. Please review the delivery and advise on next steps.’ This is only a draft; I haven’t looked up the order or sent the handoff, and I can’t issue a refund.” **Criteria.** On turn one, it asks for a reference and claims no lookup or contact. On turn two, it repeats the reference and disputed facts, then drafts a handoff without claiming it was sent or a refund approved. ## What these cases show The second dialogue waits for its missing reference; the first omits the date and closes the language question too soon. After a change, replay the inputs and compare criteria and wording rather than one score. One generation cannot establish stability or general behavior. For a single-turn check, see [How to Test an AI Agent with a Synthetic Persona](https://www.harmate.com/en/blog/how-to-test-an-ai-agent-with-a-synthetic-persona). A persona helps exercise a scenario; it does not replace field evidence about real-world use. When two facts arrive and conflict, [test how the agent handles a correction](https://www.harmate.com/en/blog/cross-turn-contradictions-test-an-agents-memory-and-clarification). For customer service, check [agent handoff acceptance](https://www.harmate.com/en/blog/chatbot-to-agent-handoffs-keep-the-customer-case-moving), [disagreement between order records](https://www.harmate.com/en/blog/conflicting-order-records-verify-delivery-and-payment-before-replying), and [which details belong in a support note](https://www.harmate.com/en/blog/bank-details-in-customer-support-keep-sensitive-data-out-of-the-case) separately.