Steve Blank Says the MVP Is Dead. I Think the Problem Goes Deeper.

I have a slightly uncomfortable relationship with Lean Startup. I believe in it. I built companies that way, and Build, Measure, Learn is still one of the most useful ideas in entrepreneurship. But I also think one of its central mechanisms has become dangerously misleading: the Minimum Viable Product.

Not because MVPs are suddenly bad. Quite the opposite. AI has made them ridiculously easy to build, and that is precisely the problem. For most of the startup era, building something cost enough time, money and engineering effort that the artifact itself carried information. A team had to make choices. It had to narrow the problem, decide which assumptions mattered and commit scarce resources to testing them. Today, much of that friction has disappeared. Give Claude or Cursor a prompt on Friday and you can have something surprisingly polished by Monday. It looks like progress. It may tell you almost nothing.

The artifact stopped being evidence

Last week, Steve Blank wrote that AI had killed the MVP. That got my attention. Blank helped create the methodology that taught a generation of founders to get out of the building, test hypotheses and search for a business model before scaling one. Now he is changing the course he has taught for fifteen years. His students can build so much, so quickly, that the amount of product they produce no longer reflects the amount they have learned. He calls these AI-generated day-one products Initial Untested Products, or IUPs.

I think he's right. But I think the problem goes one step further. AI didn't just kill the MVP. It broke the relationship between building and learning. And that relationship was doing more work than we realised.

When building was expensive, being wrong created friction. You could not test every idea, build every feature or pursue every customer segment simultaneously. Constraints forced choices. AI removes much of that friction, which is wonderful for execution and potentially terrible for learning. You can now build five versions instead of deciding which assumption matters. You can respond to every objection with another feature. You can keep moving without ever confronting whether the market is moving with you. We have become dramatically better at producing things and not necessarily any better at deciding which things deserve to exist.

The scarcity moved

The obvious response would be to replace the MVP with a new kind of artifact. I think that misses the deeper shift. There is no artifact clever enough to solve this problem, because when artifacts become abundant, their information value collapses. The scarcity has moved somewhere else: to the customer. More precisely, to a real person doing something that costs them enough that their behaviour becomes difficult to fake.

That is the idea at the centre of Human Signal. The artifact used to be evidence. Now it's the starting line, and everyone reaches it. What matters is what happens after you ship it. Does someone come back without being reminded? Do they book the call? Introduce you to their boss? Sign the letter of intent? Put down a deposit? Pay? Buy again? Recommend you without being asked?

Those actions are not equally valuable, and that gives us a more useful way to think about validation in the AI era. Instead of asking, What is the minimum product we can build?, I increasingly think founders should ask a different question: What is the minimum action we can get a real customer to take that costs them enough to teach us something?

Call it the Minimum Costly Signal.

A compliment costs nothing. An email address costs almost nothing. A 30-minute call costs something. Introducing you to a colleague costs more. Signing a pilot creates internal risk. A deposit costs money. A repeat purchase costs money twice. The more costly the behaviour, the harder it becomes to explain away.

That does not mean payment is always the only signal that matters. Cost can be money, time, effort, reputation, switching behaviour or organisational commitment. But there should be a cost. Otherwise founders are remarkably good at turning politeness into product-market fit.

AI makes this problem worse in a wonderfully ironic way. It is perhaps the most patient validation partner ever invented. It will analyse your market, improve your pitch, generate personas, challenge your positioning if you ask it to, praise the idea if you ask it differently, and then help you build the product. You can now spend an entire weekend receiving sophisticated feedback on your startup without encountering a single person who actually wants it.

The least reliable instrument is still the founder

There is another part of this that no model fixes for us. The instrument reading all this evidence is the founder, and the founder is usually the least reliable instrument in the building.

Finding product-market fit requires a strange psychological switch. Founders need conviction to recruit people, raise money and survive long enough to have a chance. Customer discovery requires almost the opposite: doubt. You need to search actively for evidence that your beautiful theory is wrong. We tell a story airtight enough to convince employees, investors and ourselves, then somehow have to switch out of sell mode the moment we speak to a customer.

AI can amplify this problem because it gives confirmation bias industrial machinery. A better prototype, another market analysis, a prettier deck, ten synthetic personas, a new feature addressing the objection from yesterday's meeting. More output can now hide a lack of learning remarkably well.

So the discipline has to move upstream. Before running an experiment, write down what result would make you change your mind. Before launching the landing page, define the number that means stop. Before showing the prototype, decide what behaviour will count as evidence. Write the kill-line before you see the result. Otherwise you are not really testing a hypothesis. You are negotiating with yourself, and founders tend to be excellent negotiators when the continued existence of their company is on the table.

Lean Startup isn't dead. Its unit of progress changed.

This is why I disagree slightly with the idea that AI killed Lean Startup. It may have killed the MVP as a meaningful unit of evidence, but the most important insight behind Lean survives perfectly: a startup is a search, not an execution. Steve Blank was right about that before almost anyone else, and AI has made the distinction more important, not less.

What changed is the economics of the search. Building used to be expensive, so Lean taught us to waste less of it. Now building is cheap. The scarce resources are customer attention, committed behaviour and founder judgment. That requires an update to how we measure progress.

Shipping isn't learning. A prototype isn't validation. A signup isn't demand. A successful demo isn't product-market fit. And an AI-generated product that looks remarkably like a startup may contain approximately zero evidence that a startup should exist.

So build the product. Build it in a weekend. Vibe-code the hell out of it. There is no prize for taking three months to create something AI can help you create in three days. Just don't congratulate yourself on Monday.

Go find a human being and ask them to do something that costs them.

Then watch what happens.

The product is no longer the proof. The human signal is.

I explore this shift in Human Signal: How to Find Real Product-Market Fit in an AI-Built World. Steve Blank's series on how AI is changing Lean LaunchPad is worth reading in full.

Sources:

• Steve Blank: AI Killed the MVP – Long Live the IUP
• Steve Blank: The Year AI Came For Us