Your Students Can Ship an MVP on Day One. That's the Problem.
A student walks into class in week one with a working product.
Not a wireframe, not a mockup, a functioning application built over a weekend with AI. Ten years ago that was the output of the entire course. Today it is the thing they did before the syllabus was handed out.
Most entrepreneurship programs were designed around a constraint that no longer exists. We built our courses on the assumption that building was the slow, expensive, learning-rich part of starting a company. It was a reasonable assumption for about forty years. It stopped being true in roughly eighteen months.
I taught entrepreneurship at BYU for nearly two decades. Since 2015 I have run the Startup Ignition Bootcamp, now past 50 cohorts and more than 1,000 ventures, and I sit on the investor side of the table as well. I am writing this for the people doing the same job I did: professors, program directors, and accelerator managers who are looking at a cohort full of finished products and wondering what exactly they are supposed to teach now.
The short answer is that almost everything in the method survives. The assignment that used to deliver it does not.
What AI actually broke in entrepreneurship education
Steve Blank wrote about this from the Stanford classroom in a four part series for Poets&Quants in September 2026. In his Spring 2026 Lean LaunchPad, every presenting team arrived with a finished, functional product built before the course began. He calls the artifact an Initial Untested Product rather than a Minimum Viable Product, and he uses the phrase “evidence theater” for polished work that reflects very little real customer contact. Both terms are his, and they are the most useful vocabulary anyone has given this problem. I have reviewed the whole series part by part for anyone who wants the detail.
He is careful about the size of the claim and so am I. Customer Development holds. The Business Model Canvas holds. Hypotheses, interviews, evidence, pivots: all of it holds. What broke is narrow and specific. Building a product stopped being proof that any learning happened.
I can add one thing to his account, which is that this is not a student behavior. We see the identical pattern in self funded adults in our bootcamp who have no grade to optimize for, several of whom have quit a job to be in the room. When the same effect shows up in graded students at Stanford and in forty year old founders spending their own money in Provo, the grade was not causing it. The artifact was.
That matters for curriculum design, because it means you cannot fix this by changing the incentive. You have to change the assignment.
Auditing an entrepreneurship curriculum: the weeks that stopped working
Here is the uncomfortable audit I would run on any entrepreneurship course right now.
Take your semester and mark every week whose primary function was to make students build something. Scoping. Technical planning. Sprint check ins. The mid semester prototype review. For most programs that is somewhere between a third and half of the calendar.
Now ask what each of those weeks was teaching that was not the building itself.
For my whole career, the answer was: quite a lot, accidentally. The cost of building forced a set of questions nobody had to assign. Before you committed six months of engineering or a semester of a team’s time, somebody asked who this was for, whether the problem was real, and whether anyone would pay. Those questions did not get asked because students were disciplined. They got asked because the alternative was expensive.
That forcing function is gone. It is genuinely a gift, and I do not want it back. Students who could never have afforded to test an idea can now try five. But we removed a constraint that was quietly doing pedagogical work, and nothing replaced it. The discipline it produced was never voluntary in the first place.
So the weeks are still on the calendar. They are just no longer teaching the thing they used to teach by accident. That time is now available, and what it should be spent on is the part that did not get cheaper.
Building got easy. Finding product market fit is exactly as hard as it has always been.
Students can fake customer discovery with AI, and no one has a policy for it
This is the one I would want a department to think about before next semester, because it is newer than the MVP problem and considerably more awkward.
Students can now fabricate the evidence too.
A student can generate twenty plausible customer interview transcripts in the time it takes to walk across campus. Personas, quotes, objections, a tidy synthesis with themes. They will read like good work, because the models are good at producing the shape of good work. If you are grading a stack of interview write ups on a Sunday evening, you cannot reliably tell which ones happened.
Every professor I know has thought hard about AI and the essay. Very few have thought about AI and the interview log, and the interview log is the one that matters more, because in this discipline it is not a writing exercise. It is the evidence the entire grade is supposed to rest on.
I do not think detection is the answer. It will not work, and it puts you in an adversarial posture with the students who are doing it honestly. Provenance works better, and it is cheap:
Require the raw artifact, not the summary. A recording, or an unedited transcript with the date and the interviewee’s first name and role. Summaries are generated. Raw material is harder to fake convincingly and much less pleasant to fake.
Require a contactable name for a sample. Tell students up front that you will verify two interviews per team, chosen at random, by sending a one line email. You will almost never need to send it. The announcement does the work.
Grade the surprises, not the themes. A generated transcript converges on what the student already believed, because that is what the student prompted. Real conversations produce the thing nobody expected, usually in an aside. Ask for the single most surprising sentence a customer said, and ask what the student believed before hearing it. That question is very hard to answer from synthetic material, and it happens to be the most valuable thing in the whole exercise.
What is still scarce, and therefore what to teach
If the build is no longer scarce, and the write up is no longer scarce, it is worth being precise about what is.
A student choosing which assumption would kill the business if it were wrong, and testing that one first instead of the one that is fun to test.
A student getting a stranger to tell them something uncomfortable. Not a survey response. A real person being honest about what they actually do today, which is almost always duller and more entrenched than the founder’s story requires.
A student recognizing a result that kills the idea, and saying so out loud, on the record, to a room.
A student changing their mind in public and being able to name the exact conversation that changed it.
None of those can be prompted. All of them can be assigned, practiced, and graded directly. That is the course now.
Six ways to teach startup validation in the classroom
These are not theoretical. They come out of the bootcamp, where we have three days rather than fifteen weeks and therefore no room for anything that does not work. Take any of them.
1. Make them write the hypothesis before they are allowed to build. One page, before any code: who the customer is, what the problem is, what they do about it today, and what would have to be true for this to be a business. Date it and lock it. It becomes the baseline that every later claim is measured against, and it is the only honest way to tell later whether a student learned something or just rearranged the pitch.
2. Hold the demo until the back half. Not a ban on AI, which teaches a constraint that will not exist in their careers. A rule about sequence. Nothing gets shown to a customer until a first round of real conversations exists. A founder holding a finished product finds it genuinely difficult to ask an open question. They stop asking “what do you do about this today” and start asking “would you use this,” and the answer to the second question is worthless, because it is an opinion about a demo rather than evidence about a life.
3. Teach the costly signal. The single most useful exercise we run is asking a customer who has just said yes to do something small that costs them something. A five dollar deposit that will not be cashed unless the product ships. A calendar slot. An introduction to their boss. The gap between “yes, I would buy that” and an actual commitment is where the entire lesson lives, and most students have never once pushed a conversation that far. They are shocked by how fast the yes evaporates. That shock is the curriculum.
4. Grade the pivot, not the polish. Put a line item in the rubric for a hypothesis the team killed, with the evidence that killed it. Teams that kill nothing in a semester did not learn anything, and under the old system you could not see that clearly because they were busy building. Now you can.
5. Require a “what surprised you” memo every two weeks. Short. Three sentences. What did you believe, what did you hear, what changed. It is the cheapest high signal instrument I know, it is nearly impossible to generate convincingly, and reading a semester of them back to back tells you more about a team than any final presentation.
6. Make the final deliverable the evidence file, not the pitch. Let them pitch, it is a real skill. But the graded artifact should be the accumulated record: the dated hypotheses, the raw transcripts, the commitments obtained or refused, the pivots and their causes. A pitch deck shows what a student can assemble. The evidence file shows whether they understand a market.
Are AI-simulated customer interviews useful for teaching?
I should address this directly, because we build software in this space and it would be convenient for me to be vague about it.
AI simulated customer interviews are useful for rehearsal. They are a good way for a student to practice question design, to hear what a leading question sounds like coming back at them, and to prepare before spending a real stranger’s forty five minutes. Used that way they raise the quality of the real conversations considerably, which is the whole point.
They are not evidence, and a program that lets them become evidence has quietly removed the one irreplaceable part of the course. A synthetic customer has no budget, no switching costs, no boss, and no reason to disappoint you. It will agree with you. Agreement is the precise thing customer discovery was invented to protect founders from.
The distinction I would teach explicitly: simulation is for preparing the question, a real person is for getting the answer. Students who understand that line use the tools well. Students who do not will produce a very confident, very tidy, completely unfounded business.
Why established university entrepreneurship programs struggle with this
Some of the most established entrepreneurship programs in the country are having the most trouble with this, and it is not because the people running them are not good at their jobs.
It is because the programs were designed well, for the world that existed when they were designed. A program built around students forming teams and building something over two semesters made excellent sense when building was the hard part, when a team had to recruit a technical cofounder, when shipping anything at all demonstrated real capability. Those programs produced good outcomes for years and the design was correct.
The design is not correct anymore, and the longer the build cycle the harder it is to see. Over two semesters a team holding a finished product has plenty of time to look productive. They iterate, they add features, they present at the showcase. The absence of learning hides inside the motion, sometimes for the entire program.
Our bootcamp is three days. That compression is not a selling point, it is a diagnostic. There is no interval in which building can substitute for knowing. On day one we work the problem, the customer, and what would have to be true. By day two they are talking to actual customers. The pattern Steve describes shows up inside seventy two hours, which makes the causation very hard to miss.
Students from well regarded university programs have come through our cohorts and said afterwards that they learned more in three days than in months of their program. I do not repeat that to be unkind to anyone. I repeat it because it is diagnostic of the same thing: they had spent months building, which they were good at, and almost no time being told something they did not want to hear.
What I would change in an entrepreneurship course next semester
If I were handed a course tomorrow, the change is smaller than it sounds.
Keep the method. It is not broken and the people who wrote it have said so.
Take back the weeks that were doing nothing and give them to customer discovery. Move the build to the back half, and require that it reflect something learned rather than something assumed.
Change what you grade. Hypotheses dated before contact, documented changes of mind, costly signals from real customers, hypotheses killed.
Ask for provenance on interviews and tell students you will spot check it.
And accept the part that is genuinely harder now. The budget used to teach discipline for free. You have to teach it on purpose.
ToolSuite for universities and accelerators The validation sequence we run in the bootcamp, as software your cohort works through. Built for programs with a new intake every semester. See how programs use it →If you run an entrepreneurship program and want to talk through any of this, including the parts where you think I am wrong, I am happy to. I spent twenty years on your side of this problem and I would rather compare notes than pitch anything. Reach me through the schools page and ask for a conversation with John.
Frequently Asked Questions
How should universities teach entrepreneurship now that AI can build the product?
Shift the unit of work from the artifact to the evidence. Building a product used to take weeks and taught students something along the way. It now takes a weekend and teaches them almost nothing. The scarce skills are the ones AI cannot produce: choosing what to test, getting a stranger to tell you an uncomfortable truth, recognizing when the answer kills your idea, and changing your mind in public. Those have to be assigned, practiced, and graded directly, because the build no longer forces them.
Can students fake customer discovery with AI?
Yes, and this is the newest problem in entrepreneurship education. A student can generate plausible interview transcripts, personas, and survey results in minutes, and a professor reading them cannot reliably tell them from real ones. The fix is provenance rather than detection: require recordings, raw unedited transcripts, interview dates, and a contactable name for a sample of interviews, and grade the surprises rather than the summaries.
Is the Lean Startup method still relevant for university courses?
More relevant, not less. Steve Blank, who wrote the Lean LaunchPad, has been explicit that the foundations hold: Customer Development, the Business Model Canvas, hypotheses, interviews, evidence, and pivots. What broke is narrower. Building a product stopped being proof that any of that happened. The method is intact. The assignment that used to deliver it is not.
Should students be allowed to use AI to build their product in class?
Yes. Banning it teaches a constraint that will not exist in their careers, and the build is no longer where the learning was. The useful restriction is on sequence, not on tools. Hold the build until a documented hypothesis and a first round of real customer conversations exist, so the product reflects something the student learned instead of deciding it in advance.
What should professors grade instead of the demo?
Four things AI cannot generate on a student’s behalf: a written hypothesis recorded before any customer contact, a documented change of mind with the specific conversation that caused it, a costly signal from a real customer such as a deposit or a scheduled pilot, and a hypothesis the student killed. A demo shows what a student can prompt. These show whether the student understands a market.
Are AI-simulated customer interviews useful for teaching?
They are useful for rehearsal, not for evidence. Simulated interviews are a good way for a student to practice question design, hear how a leading question sounds, and prepare before spending a real stranger’s time. They cannot supply the finding. A synthetic customer has no budget, no switching cost, and no reason to disappoint you, and agreement is the exact thing customer discovery exists to guard against.
How long should a university entrepreneurship course spend on validation?
Longer than most do, and earlier. When building took most of a semester, validation got whatever time was left. Now that the build is a weekend, the released time belongs to customer discovery. In our own three-day bootcamp format, founders are talking to real customers by day two, and the compression is deliberate: a long build cycle gives a team room to look productive while learning nothing.
Back to All News
