Short answer
Ask about the answer you just got, not about the topic in general: why it is true, what evidence supports it, what else could explain it, and what would change your mind. A structured method such as the five whys turns a chatbot from an answer machine into a thinking partner that helps you reach the cause of a problem.
Follow-up question How do you know when to stop asking?
The first answer a chatbot gives is usually the most common answer to the question as asked. That is often fine. But when you are trying to understand a problem, the common answer is a starting point, and the useful insight sits two or three questions further down. Follow-up questions are how you get there.
A Beginner's Guide to ChatGPT recommends working in rounds: start broad, read the answer, then narrow, clarify or add context and ask again. Iterative prompting usually improves the answer. Digging deeper improves your understanding: each new question targets one specific statement in the last reply and asks what lies beneath it.
Four kinds of follow-up that go deeper
Most weak follow-ups are variations of "tell me more," which produces a longer version of the same answer. These four types each push in a different direction:
| Type | What it does | Example |
|---|---|---|
| Why | Moves from a symptom toward a cause | "Why would that happen in a small shop specifically?" |
| Evidence | Tests whether a claim holds up | "What would I see in my sales data if that were true?" |
| Alternative | Opens other explanations | "What else could produce the same pattern?" |
| Consequence | Checks what follows | "If that's the cause, what should change when I fix it?" |
The "why" question is the heart of a well-known method from manufacturing, the five whys. It was developed by Sakichi Toyoda and became part of problem-solving training in the Toyota Production System. You state a problem, ask why it happens, then ask why about that answer, and keep going. Five rounds is a rule of thumb, not a law. The aim is to reach a cause you can actually act on, which is usually a process that is missing or not working.
A worked example: a café with falling lunch sales
Here an AI plays facilitator. Its job is to question you, because you hold the facts about your business and it doesn't. The replies are summarized and illustrative.
Act as a facilitator for a five-whys analysis. My problem: weekday lunch sales at my café have dropped about 20% over three months. Ask me one "why" question at a time, based on my last answer. Don't suggest solutions until we reach a cause I can act on. If my answer is a guess, ask me how I could check it.Why 1. The AI asks why fewer people are buying lunch. The owner answers that regulars from the office building next door come in less often.
Why 2. Why are they coming less? The owner thinks it might be the new sandwich shop on the corner. The AI notes that this is a guess and asks how she could check. She realizes she hasn't asked anyone.
She asks three regulars the next day and comes back with a different answer: they still like the café, but the wait at noon has become too long for a 30-minute break.
Why 3. Why has the wait grown? Since the summer, all hot sandwiches are made to order, and the menu has expanded.
Why 4. Why does that slow things at noon? One person runs the press and the till at the same time during the rush.
Why 5. Why is one person doing both? The schedule was set before the menu changed and nobody revisited it.
The root cause is not the competitor. It is a staffing schedule that no longer matches the menu, something she can test next week by adding a second person from 11:45 to 1:15, or by offering a few pre-made options at noon.
Two moments in that conversation did the real work: the AI flagging a guess, and the owner leaving the chat to collect a fact. A chatbot can structure the questioning, but it cannot know what your customers think.
Try it: Pick a small, real problem at work or at home, paste the facilitator prompt above with your own problem, and answer honestly. Each time you answer with "probably" or "I think," stop and find out before the next why.
Watch for the method's blind spots
The five whys has well-known critics, and their points apply when you use it with AI. People tend to stop at a symptom instead of a cause. They also tend to look for one root cause when several are contributing. You can counter both with follow-ups: "Is this still a symptom of something else?" and "What other causes could be contributing at the same time?" For tangled problems with many causes, a branching diagram such as a fishbone chart, which maps several causes side by side, works better than a single chain.
How do you know when to stop asking?
Stop when the last answer points to something you can change and test. "Customers are busy" is not actionable; "the noon shift has one person for two jobs" is. If asking one more why takes you to causes outside your control, such as the economy or the weather, you have gone one step too far: back up to the last answer you can act on.
Stop, too, when the answers stop being facts. If you are repeatedly guessing, the conversation has run ahead of what you know. That is the moment to leave the chat and gather information, then return with it.
Finally, stop when the replies start repeating themselves in new words. A model will happily keep generating plausible "deeper" layers forever, because producing plausible text is what it does. Depth comes from your facts, not from the number of questions. Three good whys backed by evidence beat seven that are built on guesses.
Sources
- A Beginner's Guide to ChatGPT — Holistic Hybrid, April 18, 2023
- Five whys (revision of 23 May 2024) — Wikipedia, May 23, 2024