Role-play case study chatbots for active discussion

Case-based teaching has a scaling problem. The pedagogy assumes a small seminar where every student is on the spot, has to defend a position, and gets challenged by faculty and peers in real time. In a 50-person section — or a 30-student MBA class with a 75-minute window — most students never get pushed. They observe a discussion that happens to a few of their classmates.

A pattern that’s working at MSU: build a chatbot that plays one side of the case, and require every student to engage with it before class. In the bot, every student gets a one-on-one, every student has to defend their reasoning, and class time can move on to the higher-order discussion.

The pattern

For each case study:

• The chatbot is loaded with a detailed summary of the case (or the full case if licensing permits).
• The bot is prompted to play a specific role — the company’s CFO, a board member, an investor, an opposing expert witness, a skeptical client.
• The student plays the other side — the analyst making a recommendation, the advisor, the lawyer making the argument.
• The interaction runs for a fixed number of rounds (typically three), with each turn requiring the student to cite specific evidence from the case.

Students are graded on whether they complete the rounds and on the quality of their citations and reasoning. The transcript becomes a low-stakes pre-class artifact.

Why it works

A few things happen when you do this:

• Every student has to take a position. No more drifting through a class period without engaging.
• The bot does not let students bluff. A well-prompted bot will press on weak reasoning or unsupported claims. Students learn quickly that vague answers don’t move the conversation forward.
• Class discussion gets better. When every student arrives having already worked through the case, in-class time stops being for first-pass analysis and starts being for synthesis, comparison, and the harder questions.
• Faculty get a window into student thinking. Skim a few transcripts before class. You’ll know exactly which concepts are landing and which are not.

Setup steps

1. Write a detailed case brief for the bot. If you can’t load the case PDF for copyright reasons (Harvard Business Publishing cases are a common example), build a 1,500–2,500 word internal summary the bot can use. Include the facts, the tension, key actors, and the data the student is expected to engage with.

2. Define the bot’s role precisely. Generic “play the CEO” instructions get generic responses. Specify:

• The role (CEO, board chair, lead investor, opposing counsel, regulator)
• The role’s perspective and priorities (e.g., “You are most concerned about quarterly earnings; you are skeptical of long-horizon investment without clear ROI”)
• The role’s behavior in the conversation (e.g., “Press the student on numbers. Ask for sources. Push back if the recommendation feels under-justified.”)

3. Set the engagement rules.

• Number of rounds (three works well)
• Required citation of evidence from the case (this is what makes the exchange substantive)
• Word limit per turn (keeps the conversation moving)

4. Make the transcript the deliverable. Students submit the transcript before class. Grade it on completion and citation quality, not on whether the bot was “convinced.” The bot should not be easy to convince — that’s the point.

A starter system prompt

Adapt the role and topic to your case:

You are [role] in the case study titled [case title]. You have read the full case. Your perspective and priorities are: [list 2–4 priorities specific to the role].

A student in this course will play the part of [opposing role]. They will present a recommendation. Your job is to press them — ask for evidence from the case, challenge weak reasoning, and require them to defend their position with specific facts from the materials.

Engage for exactly three rounds. After each student response, ask one or two follow-up questions or counter-arguments. Do not be persuaded easily. Do not concede a point unless the student has cited specific evidence from the case to support it.

Keep your responses under 150 words. Stay in role at all times. If the student asks an off-topic question, redirect them back to the case.

At the end of the third round, summarize the student’s strongest point, their weakest point, and one specific piece of evidence they did not use but should have.

Platform options

Same set as the Socratic tutor: Copilot Studio, Gemini Gems, Claude Projects, ChatGPT Custom GPTs. For role-play specifically, Claude tends to stay in character most reliably; Gemini and Copilot are easier for students to access through their NetID.

Copyright and access caveats

A real practical issue: many published cases (Harvard Business Publishing, Ivey, etc.) cannot be uploaded directly to a chatbot platform under their licensing terms. The workarounds:

• Write your own case summary for the bot. Time-consuming but reusable across semesters.
• Use cases written by MSU faculty or in the public domain. Not always available for the topic you want.
• Have students paste relevant excerpts in their own conversation. Pushes the copyright issue onto the student’s individual session, which is generally permitted under fair-use educational provisions, but ask your library or general counsel before relying on this.

This is the most common reason this pattern stalls in adoption. Decide upfront whether you have the time to write summaries or whether you can work with cases you’re allowed to load.

Practical adoption notes

• Pilot on one case before redesigning a whole course around this. Pick the case where students most consistently come unprepared and see if a pre-class bot exchange changes the in-class conversation.
• Read the transcripts. Even a quick scan tells you what students are missing. This is the highest-value part of the data the pattern produces.
• Be ready for the bot to be too aggressive or too passive. Tune the system prompt across the first few uses. You’re calibrating for “challenging but not punitive.”
• Don’t grade students on whether they “won.” Grade on completion, citation quality, and reasoning. The bot is not a judge of correctness; it’s a sparring partner.

What this pattern is good for

Any course built on case studies — business, law, public policy, ethics, medicine, engineering ethics. It works less well in courses where the curriculum is built on problem sets or labs rather than narrative cases. For those, see the Socratic AI tutor pattern instead.