Nine Entrepreneurship Assignments AI Cannot Do For Your Students

Every entrepreneurship instructor I talk to has arrived at the same place by a slightly different route: the assignments that used to work are now trivially generated, and the obvious replacements are either unenforceable or busywork.

The way out is not better detection. It is choosing assignments whose output depends on a human being responding to your student. AI can produce any document. It cannot produce a stranger’s willingness to take a call, a refusal, or a deposit.

Here are nine that hold up. All of them fit inside a course you already run.

1. The twenty-minute stranger

Find one person outside your network who has the problem. Talk to them for twenty minutes. Submit the date, their first name and role, how you found them, and your raw notes.

The deliverable is not insight. It is contact. For most undergraduates this is the hardest single thing in the course, and the fear is the point. A student who has done it once will do it ten more times. A student who has not will avoid it all semester behind generated personas.

2. The belief memo, collected before anything else

Due on a fixed date before any interviewing: what you believe about the customer, the problem, and the willingness to pay, in a form that could be proven wrong.

This is the only artifact in a course that a student cannot reconstruct afterward, which makes it the anchor for everything downstream. Every later assignment is graded against it. No revisions.

3. The disconfirmation report

Report the three strongest pieces of evidence against your idea, with sources.

Students find this genuinely difficult, and not because of AI. It runs against the instinct that the assignment is to defend the idea. Weight it heavily and say so in advance. The inversion is most of the learning in a first course.

4. The second conversation

Go back to someone you interviewed earlier. Show them how your thinking changed. Record what they say about the change.

Almost no student does this unprompted, and it is where the real information lives. It is also structurally impossible to fabricate convincingly, because it depends on a specific earlier conversation that you have already seen.

5. The price conversation

Not “would you pay for this.” Ask what they use today, what it costs them in money or hours, and what would have to be true for them to switch. Submit what they actually said.

The honest answers are uncomfortable and specific, which is exactly why generated versions are obvious. A synthetic customer says they would pay twenty dollars a month. A real one says they already pay a part-time bookkeeper and switching would mean retraining her.

6. The kill criteria, written in advance

Before testing: what would you have to learn to abandon this idea?

Then, at the end of the term: did that happen, and what did you do about it? Full marks for killing it. You have to say that early, loudly, and more than once, because no student believes it the first time.

7. The five-minute evidence review, delivered live

Five minutes, no slides. What did you believe, what did you do, what changed, what do you recommend?

No slides is doing real work here. It removes the artifact that AI is best at producing and leaves the student with only what they actually know. Questions from the room do the rest.

8. The teardown of a real failure

Take a company that died. Find the assumption that killed it. Trace it to a decision someone made and a test they could have run.

AI will cheerfully produce a plausible narrative here, so the grading requirement is primary sources: the postmortem, the founder’s own account, press from the time. Grade the sourcing, not the story.

9. The handoff

Give your evidence to a classmate. They present your idea, and your recommendation, to the room.

Nobody can present evidence they do not understand, and nobody can present fabricated evidence at all, because the questions come from a peer who had to read it closely enough to argue it. This is the single most effective integrity mechanism I have used, and it was never designed as one.

The one rule that makes the rest work

Require provenance, not purity.

Interview dates. Raw unedited notes alongside any summary. A contactable name for a sample. Then announce that you will spot-check, and actually check two per class where students can see you doing it.

You do not need to catch anyone. The declared possibility of a phone call changes behaviour far more reliably than any detector, and unlike a detector it never falsely accuses a student who did the work.

Fabricated interviews, when you do read them, are easy to spot once you know the tell. They are smooth, agreeable, on topic, and they confirm the thesis. Real interviews contain a tangent about somebody’s dog, a complaint about something unrelated, and one sentence that clearly stopped the student cold.

---

Try the ToolSuite free with your class Customer discovery and business model canvas tools built for coursework. Free to start, with academic pricing for programs. Check out our software free →

Several of these assignments need somewhere for students to log interviews with dates and sources, and somewhere for you to see the work between class sessions rather than at the end. That is what the customer discovery tools in Startup Ignition ToolSuite are for, and academic pricing is available for accredited programs.

If you want the grading side in more detail, there is a customer discovery rubric and a guide to grading the business model canvas.