A Product Is No Longer Proof: Customer Discovery in the AI Era

A Product Is No Longer Proof: Customer Discovery in the AI Era

Founders now walk into our bootcamp holding a finished product.

Not a sketch, not a wireframe, not a landing page — a working application, built with AI before they arrived. It looks like a running start. In our experience it is usually the opposite, and after more than 50 cohorts I have stopped treating it as a good sign.

We are not the only ones seeing this. Over the past week Steve Blank published a four-part series on what AI did to the Lean LaunchPad, the Stanford course he built in 2011. Every presenting team this year arrived on day one with a finished, AI-built product — and learned less than the classes that had to build their own. He names the artifact an Initial Untested Product rather than a Minimum Viable Product, because it no longer carries the two signals an MVP used to: that the team could build, and that what they built reflected something they had learned.

He is careful about the size of the claim, and worth quoting on it: “we had confused the symptoms (the MVP is dead) with the much larger real problems facing the class.” Customer Development, the Business Model Canvas, hypotheses, interviews, evidence, pivots — all of it stays. What broke is narrower. AI did not eliminate the MVP. It eliminated the assumption that having one is evidence you have learned anything. That is his argument, not mine; I have summarised the whole series here for anyone who wants it part by part.

He’s right. But I’ve watched this from a different set of seats than he has, and the view from those seats adds something to what he’s describing.


The same pattern, in founders who aren’t students

I taught entrepreneurship at BYU for nearly two decades, so I recognize the classroom version of this immediately, and I have written separately about what this means for teaching entrepreneurship if you run a program. But Startup Ignition Bootcamp isn’t a classroom. We started it in 2015 and have since worked with more than 1,000 ventures.

The people in our room are not students. There is no grade. Most of them are spending their own money, and several have quit something to be there.

That distinction matters more than it might sound, because the easiest way to dismiss Steve’s observation is to say students optimize for the assignment. Of course an ambitious Stanford team shows up with a product — it looks like effort, and effort gets graded.

That explanation doesn’t work for our room. Nobody is being evaluated. And we see the same pattern anyway: founders arriving with something polished, and asking noticeably worse questions because of it.

When an effect shows up in graded students at Stanford and in self-funded adults in Provo with nothing to score, the grade wasn’t the cause. That’s the contribution I can make to Steve’s argument — not a new theory, but an independent replication in a population that rules out the most convenient explanation.

A founder in our Provo cohort last week is a good example. He arrived with an AI tool for small property managers: it reads a tenant’s maintenance message, diagnoses the issue, replies to the tenant, and dispatches a contractor. He had a working version in four days, built with Claude.

What he had was an IUP in Blank’s sense, and it is worth watching what that costs him. The product had already decided something important. It was designed to send replies without the manager in the loop, which meant it assumed managers want maintenance off their desks as fast as possible. Reasonable. Untested.

He demoed a tenant reporting a leak and asked a property manager whether it would save her team time. She said yes. He started writing that down as validation — and then, almost as an afterthought, asked whether she would let it send the reply on its own.

She said no, she needs to see the message before it goes out. Last week a tenant reported a leak and the photos showed they had left the bathtub running. Who pays for that depends on what the manager says next.

That is not a feature request. That is the business. The value was never speed — it was judgment about liability, exercised in the exact moment his product was designed to skip.

Here is the part that matters for Steve’s argument. The founder’s first instinct after her objection was to show her another feature. He had already built the automatic reply flow, so he heard a missing setting rather than a wrong premise. The build had made one answer much cheaper to defend than to abandon.

He got there. By the end of the week he had stopped pitching automatic dispatch and was interviewing managers about which decisions they could delegate, which ones they had to review, and why — which points at a tool that assembles the evidence and drafts a response for approval. A better product, arrived at by giving up the one he had already built.


Three days compresses what ten weeks reveals slowly

Our bootcamp is three days, not ten weeks. That difference turns out to be diagnostic.

Over ten weeks, a team holding a finished product has time to look productive. They iterate, they add features, they present weekly. The absence of learning can hide inside the motion for quite a while.

Three days has no room for motion. On day one we work through the problem statement, the customer, and what would have to be true for the business to work. By day two founders are talking to real customers. There’s no interval in which building substitutes for knowing.

What we watch happen in that compressed window is the thing Steve describes, but sped up and visible: a founder holding a working product finds it genuinely difficult to ask an open question. They don’t ask “what do you do about this today?” They ask “would you use this?” The artifact in their hands changes the grammar of the conversation, and the answer they get back is worthless because it’s an opinion about a demo instead of evidence about a life.

We have taught the same exchange in our curriculum for years. It is the clearest illustration I know of what a polite “yes” is actually worth:

Founder: Let’s assume the product you just learned about existed today. Is it safe to assume you would be interested in purchasing?

Customer: Yes.

Founder: So if it existed right here, right now, you would part with money and buy it.

Customer: I think so, yes.

Founder: It’s actually very close to being ready. Would you be willing to write a check right now for $5 — the first month — and it won’t be cashed until the product launches?

Customer: Now that’s different. I’m not saying I’d do that.

Founder: Why? The check won’t be cashed ever if the product does not launch.

Customer: Actually, I was just trying to be friendly. I like you and your idea, but I think I can provide that service for my family on my own with free things already available on the web.

Founder: So really, this is not that great of an idea, because you have other ways to get this done.

Customer: Right.

Four exchanges to get from “yes” to “right.” The entire gap between those two words is the difference between an opinion and evidence, and $5 is what it costs to find out.

That exchange was hard enough to run before AI. It is harder now, because a founder holding a finished product almost never gets to the check. The demo is so much more pleasant to show than a commitment is to ask for, and the polite yes it earns feels like the answer.

Watching it happen over three days rather than ten weeks makes the causation hard to miss. The product doesn’t just fail to help discovery. It actively degrades it.


The forcing function nobody designed

Here is the part I would add to Steve’s diagnosis, and it is the thing I have thought about most since reading him.

For my whole career, the cost of building protected founders from themselves. Not on purpose — nobody designed it that way. But writing a cheque for six months of engineering made you stop and ask who this was for, whether the problem was real, and whether anyone would pay. Those questions did not get asked because founders were disciplined. They got asked because the alternative was expensive.

That constraint is gone, and it is genuinely a gift. Founders who could never have afforded to test an idea can now try five. But we removed a forcing function without replacing it, and the discipline it used to produce was never voluntary in the first place.

Which is the whole problem in one line: the cost of building has collapsed. The cost of building the wrong thing has not.

Six months of your life, the cofounder who left a job for this, the customers you burned on version one, the eighteen months of runway — none of those got cheaper. Only the code did. And because the code is the visible part, the collapse of that one cost reads like everything got cheaper.


What evidence theater looks like from the investor side

This is the part I don’t think Steve’s vantage point covers, and it’s where I’d add the most.

Steve sees the classroom. I’ve also spent two decades on the other side of the table — over 200 angel investments, running UtahAngels, co-founding BoomStartup, and now Startup Ignition Ventures. So I see what happens to these companies six months after the course ends, when they’re asking someone for money.

Here is what changed in the last eighteen months. A pre-seed pitch used to open with a deck and a story, and a working product was a meaningful signal partway through diligence. Now the product arrives first, and it is beautiful, and it tells me nothing.

I’ve started asking a different question early in those meetings: what did you believe three months ago that you no longer believe?

The answers separate founders faster than any demo. A founder who has done real customer discovery answers immediately and specifically — they thought the buyer was the ops manager, it’s actually the CFO; they thought pricing was the objection, it’s the migration. A founder who has been building rather than learning gives you a version of “we’ve refined the positioning.”

That’s the practical consequence of Steve’s argument for anyone writing checks. If the product is no longer evidence, diligence has to move upstream to the founder’s own record of being wrong. I now weight a specific, dated account of a changed mind above any demo I’m shown.

Which leads to the question the whole series turns on: if the MVP is no longer proof, what is? We had to answer that for ourselves before he published his answer, so let me give ours first.


What we ask for instead of a product

We are adopting Blank’s vocabulary for this, because naming the thing changes how founders hold it: what arrives on day one is an IUP, and it earns the name MVP only once something has been tested against a customer. After more than 50 cohorts, here is what we ask for in the meantime. None of it requires the product to exist.

  1. What you believed before, and what you believe now. Dated, specific, and different. If nothing changed, you weren’t doing discovery — you were doing demos.
  2. What a customer did that cost them something. Not what they said. A deposit, a signed LOI, a workflow they changed, two hours of their time they won’t get back. Enthusiasm is free; that’s why it’s worthless.
  3. The assumption that would kill this. Every founder can name what has to go right. Few can name what would end it. The ones who can are the ones doing the work.
  4. Why this is still yours in two years. Code is not the answer anymore. Distribution, proprietary data, workflow integration, switching costs, brand, network effects, genuine expertise — something that compounds rather than something that can be regenerated over a weekend.

A founder who can answer those four has done the thing the MVP used to stand in for. A founder who can’t hasn’t, no matter what’s running on their laptop.


Design partners, and the version of that a first-time founder can actually get

His third part answers the question this one raises. If a built product is no longer evidence, what is? His answer is a design partner — a customer who co-develops with you and commits something scarce to do it: staff time, data, integration effort, permissions, test environments, changes to their workflow.

I want to point at the overlap honestly, because it is the most encouraging thing in this series. Five days before that piece ran, the list above said the second thing we ask for is what a customer did that cost them something — a deposit, a signed LOI, a workflow they changed, two hours of their time they won’t get back. He arrived at the same test from Stanford and deep tech; we arrived at it from a three-day room full of first-time founders. Neither of us copied the other. That is usually a sign the underlying thing is real. The same happened with moats — his conclusion that defensibility can no longer be deferred is the fourth thing on that list, phrased as why this is still yours in two years. Two people looking at the same collapsed cost of building end up asking the same two questions.

Where I would add something is what happens when you are not selling to an enterprise.

A design partner is an enterprise-shaped instrument. It assumes a customer with staff to assign, systems to integrate, and procurement to navigate — which describes most of Blank’s teams and almost none of ours. A founder building for dentists, contractors, restaurant owners or consumers has no counterpart who can commit an integration effort. Told the standard is a design partner, they will conclude the standard does not apply to them, and go back to showing the demo.

It does apply. It just has to be translated down. The general principle is that evidence is whatever the customer cannot give you for free, and at small scale that currency is different: a deposit, a pre-order, a cancelled subscription to something else, an introduction to their boss, a spreadsheet they actually send you, a Tuesday they gave up. Five dollars from a plumber is the same instrument as a Fortune 500 committing a test environment. Both cost the person something. Both are impossible to fake with enthusiasm.

Blank’s other point in that piece — that product/market fit is fragmenting into category-specific versions — lands on my side of the table too. “Do you have product-market fit?” is quietly ceasing to be one question in a pitch meeting, and an investor who keeps asking the generic version will keep getting the generic answer.


The uncomfortable part: we also sell the accelerant

I’d be dodging if I didn’t say this plainly. Startup Ignition builds AI tools for founders. Our ToolSuite exists to make idea analysis, market research and business modeling faster. We are, in a direct sense, part of what Steve is describing.

I don’t think that’s a contradiction, but it did force us to be specific about a design principle: AI should compress the time it takes to learn something, not the time it takes to build something.

Those sound similar and they are opposites in practice. A tool that writes your interview questions, stress-tests your assumptions, or tells you which assumption to test first accelerates learning. A tool that turns your unexamined idea into a deployable app accelerates commitment to a belief you haven’t checked. The first makes discovery cheaper. The second makes discovery feel unnecessary.

We build for the first, and we say so in the room. Our free AI Idea Validator deliberately hands back an assumption to go test rather than a product to go show. That’s not a small distinction — it’s the whole difference between the two futures this series points at.


Where this leaves founders

AI has given entrepreneurs one of the most powerful tools we’ve ever had, and founders should use it aggressively — to research faster, prototype faster, analyze faster, and experiment faster.

But Steve’s central point deserves repeating, because it is genuinely new: the existence of a product is no longer evidence of anything. That was true in his classroom, it’s true in our bootcamp, and it’s now true in the meetings where companies get funded.

When everyone can build, knowing what to build is the entire advantage. Pre-AI entrepreneurship was constrained by execution. AI-era entrepreneurship is constrained by judgment. Which means customer discovery — the least glamorous, most avoidable part of this work — is the only part that still separates founders from each other.

If you’re at the start of this, our step-by-step guide to validating a startup idea is the place to begin. If you’re already holding something you built in a weekend, the more useful exercise is the one nobody enjoys: write down what result would make you stop, then go find out whether it’s true.

Blank’s fourth part — the redesigned course — has not run yet. I am more interested in it than in anything else being written about AI and startups right now, because he is doing the thing most commentary avoids: rebuilding a working method after admitting the old one broke. We will be reading it closely, and changing our own three days where he is right.

Startup Ignition Bootcamp 3-day intensive. Zero equity. Validate the idea before you build it, with exited founders and active investors. See Upcoming Cohorts →

Frequently Asked Questions

What did Steve Blank say about AI and entrepreneurship education?

In ‘The Year AI Came For Us’ (Poets&Quants, September 2026), Steve Blank reported that in his Spring 2026 Lean LaunchPad course at Stanford, every presenting team arrived with a finished, functional product built with AI before class started. He argues the Minimal Viable Product has stopped working as an artifact for learning, coins the terms ‘Initial Untested Products’ and ‘evidence theater’ for polished work that reflects little real customer discovery, and concludes the competitive bottleneck has shifted from building speed to judgment about what to build.

Is the Lean Startup method still relevant with AI?

More relevant, not less. Steve Blank, who wrote the Lean LaunchPad, is explicit that the foundations hold — Customer Development, the Business Model Canvas, hypotheses, interviews, evidence and pivots. What broke is narrower: building the product stopped counting as proof that any of that happened. AI didn’t eliminate the MVP, it eliminated the assumption that having one is evidence you’ve learned anything.

What is customer discovery?

Customer discovery is the practice of testing your assumptions about a problem, a customer, and a willingness to pay before you build a solution. It means turning beliefs into hypotheses, talking directly to the people who experience the problem, and gathering evidence that either supports or kills the idea. It is the first phase of Steve Blank’s Customer Development framework and the foundation of Lean Startup.

Does a working AI-built product prove product-market fit?

No. Ten years ago a functioning product signaled that a team had invested real time, money and technical effort. Today it signals that someone spent an afternoon. The existence of a product is no longer evidence that customers care, that anyone will pay, or that the founder understands the market. Those still have to be proven separately.

Can AI replace customer interviews?

No. AI can help you prepare interview questions, summarize transcripts, and spot patterns across conversations you have already had. It cannot generate the evidence itself. A synthetic customer will agree with you, and agreement is the one thing customer discovery is designed to guard against.

What should founders show investors instead of a product?

Evidence that the founder has been surprised. Specifically: what you believed before customer conversations, what you believe now, what changed your mind, and what a customer did that cost them something — paid a deposit, gave up time, switched a workflow. A demo shows what you can build. Those four things show whether you understand the market.

John Richards is co-founder of Startup Ignition. He taught entrepreneurship at BYU for nearly two decades, has made more than 200 angel investments, and has worked with more than 1,000 ventures since launching Startup Ignition Bootcamp in 2015.

Back to All News