The business plan was never the point

A student can now produce a market analysis, competitive landscape, five year financial model, and brand identity in an afternoon. The document will look professional. The numbers will be internally consistent.

And we may have almost no idea whether that student understands how to start a company.

That is not an AI problem.

It is a business plan competition problem that AI has finally exposed.

The uncomfortable part

If a team can generate an entire submission in an afternoon, what exactly are we scoring?

Here is the thing I think we should say out loud. The business plan was never a measure of whether someone could build a company. It was a measure of whether someone could produce documents about a company, on the assumption that the two travel together. That assumption was always shaky. Every one of us has watched a beautifully argued plan win a competition and then dissolve on contact with a first customer. Polished documents have been outrunning founder capability for as long as this format has existed.

AI did not create that gap. AI collapsed the cost of the artifact to nearly zero and made the gap impossible to ignore.

That is a favor to us, although it does not feel like one. We were grading a proxy. The proxy has stopped working. The honest response is not to defend the proxy. It is to ask what we were trying to see through it.

The credit card terminal

The first company I started needed a credit card terminal. I set it up myself.

I did not know what I was doing. I had it configured backwards, so every time I believed I was charging a customer, I was actually giving them credit.

The bank called and asked me what exactly I was running.

I was fortunate. It had only happened with four or five customers, and I was young enough that the person on the phone smiled, was patient with me, and walked me through fixing it. I called each customer back and charged them properly.

No business plan contains that. No financial model contains it. There is no line item for the founder who does not know how his own payment system works. And yet that afternoon taught me more about operating a company than any document I wrote that year.

Operating a company turns out to be a thousand small mechanical facts that you cannot learn by describing a company.

Pressure, contact, and change

If the document is no longer evidence, the evidence has to come from somewhere else. I would suggest that the next generation of venture competitions score three things we have historically undervalued.

Pressure. Can the founder defend her own assumptions live, without a script? Not a rehearsed pitch followed by polite inquiry, but real weight on a single number. If a team cannot defend its customer acquisition cost under questioning, the model is not theirs, no matter who typed the prompt.

Contact. Has the founder actually met the market? How many customer conversations, what she expected to hear, what she actually heard. A model can fabricate the transcript. It cannot fabricate the learning, and questioning exposes the difference in about ninety seconds.

Change. What is different between the first submission and the final one, and why? A team that only polished did not learn anything. A team that changed its mind about the customer learned everything.

None of these are new ideas. All of them have historically been worth a few points at the end of a rubric dominated by the document. That weighting is now backwards.

One real encounter with the market

Contact deserves a hard requirement rather than a scoring line, because it is the one thing a student cannot produce from a desk.

Before the final round, each team should have to make one real encounter with the market. Ideally that means taking one real payment from one real customer, using a payment system they set up themselves. Not a mockup, not a prototype checkout. A real account, real fees, real settlement.

Where that is not legally or practically possible, and it often will not be, given university rules, minors, regulated sectors, ventures that are not yet incorporated, and deeptech that is years from a first sale, the requirement should be satisfied by another verifiable market action. A signed letter of intent. A real supplier quote. A live pilot with a named organization. A preorder. Anything that creates consequences outside the classroom.

Then they report what broke. Something always breaks. The account gets held for review. The fee structure is not what the landing page promised. The money takes four days to arrive and the model assumed it was instant. The supplier quote comes back at three times the assumption. Every one of those is a lesson that no amount of research produces, and the report on what broke is worth more than the market analysis it sits behind.

It is also the kind of requirement a language model cannot complete on a student's behalf.

Where was AI wrong?

Most institutions are asking students whether they used AI. That is the less interesting question, and policing the answer is unenforceable anyway.

Do not ask students merely to disclose that they used AI. Ask them where it was wrong.

What did you verify? What did you reject? What did the model tell you confidently that turned out not to be true?

That is not a compliance exercise. It is evidence of judgment. A student who understands what she is doing will write something specific and slightly embarrassing about a market size figure she caught before it embarrassed her in front of judges. A student who does not will write something vague about using AI for research support. The gap between those two answers tells you more than the financial model does.

A category for AI ventures

One practical implication follows from all of this.

We should create a distinct category for AI new venture ideas and businesses. The obvious objection is that AI is a tool, and we do not create categories for spreadsheets. That was true until recently. AI can now sit at the center of the value proposition, the business model, the product itself, and the way a venture is created and scaled. A venture whose defensibility rests entirely on how it deploys a model is not the same object as a venture that uses a model to write its emails.

I would also let AI ventures compete in both their traditional category and the new one. This recognizes AI as a dimension of entrepreneurship rather than a separate country. A logistics venture built on AI is still a logistics venture.

One warning about that category

The category only works if the judging panel can tell the difference between a venture where AI is genuinely the value proposition and a thin wrapper around somebody else's model. Most panels cannot make that call today. If we create the category without a plan for recruiting judges who can, we will have created a prize for the best demo.

The dual entry rule also needs a line on whether a team can win twice. My instinct is that they should choose. Otherwise the new category becomes a second lottery ticket for the same work, and the traditional winners will notice.

What we are actually certifying

Every time we hand a student a check and a photograph, we make a claim about that student to the world. We say that this person is worth backing.

That claim was always a little generous. It is now considerably more generous than most of us are comfortable admitting, because the thing we evaluated can be assembled by a machine while the student sleeps.

I am not arguing that we retreat from AI. I am arguing that the arrival of a tool which produces flawless plans should push us toward the only thing that was ever worth measuring, which is the founder. Not the document. The person who has to answer the hard question in the room, with nothing in front of her, about a number she chose.

The bank that called me was not evaluating my business plan. It wanted to know whether I understood the machine I had switched on.

Perhaps that is the question we should be asking our students now.