AI Made Prototypes Easy. Real Customers Are Still the Hard Part.

Does AI replace talking to real customers?

No. AI has made prototypes faster, cheaper and better-looking than ever, and that’s good news. But a prototype is a question, not an answer. Only a real customer, doing something real, can answer it, and no model can stand in for that person. Finding real customers and understanding what they actually do is still the job, just like the old days.

The lure was never the hard part

Indulge me while my impractical master’s degree in aquaculture finally earns its keep.

Ask any Wisconsin angler what makes fishing hard and nobody will say the lures. Every bait shop in the state has a wall of them. The hard part is finding the fish: where they’re holding this week, what they’re feeding on, and whether they’ll bite at all.

AI just handed every product team a machine that ties lures. Clickable prototypes, landing pages, working apps and investor decks, a dozen before lunch, each one shinier than the last. Don’t get me wrong: the machine is wonderful. But it didn’t come with a lake full of hungry fish.

The fish still get the final vote.

What did AI actually change about the lean startup?

Jeff Gothelf, co-author of Lean UX, made the case in his October 2026 newsletter, AI made the lean startup even more important. It’s a point I intend to keep repeating until teams are sick of hearing it.

Jeff was answering a September 2026 Harvard Business Review article by Victor P. Seidel, Bret Greenstein and Thomas H. Davenport. Their argument: one person with AI tools can now generate ideas, build prototypes, put up test websites and run several experiments at once, for very little money. The authors call it ‘abundance entrepreneurship.’ About the abundance, they’re right. There has never been a cheaper time to build something.

Jeff’s answer, and mine, is that the lean startup was never mainly about saving money. Companies with plenty of cash adopted it too. It was about uncertainty. Nobody knows whether customers want the thing, so you build the smallest test that will tell you and watch what people do with it. AI made that test cheaper and faster to build. It did nothing about the uncertainty.

Only one thing has ever reduced it: evidence of what real customers actually do. AI can’t manufacture that evidence. It can only help you go get it faster.

Jeff saves his sharpest point for big companies, which he says adopted the language of the lean startup far more than the practice. You’ve probably sat in those meetings. Every slide says ‘MVP,’ and nobody in the room has talked to a customer since the last reorg. Now AI lets that same team produce a working prototype and a gorgeous business case in an afternoon. When everything looks finished, looking finished stops meaning anything.

Your newest teammate is a great Hacker and a lousy Hustler

In Enquire to Inspire, I describe the dispositions a strong discovery team needs. The Designer hears the customer and turns wants into needs. The Hacker whips up a quick prototype out of leftover parts. The Visionary keeps the end in sight. And the Hustler finds customers and gets the prototype in front of them, which I called, back in 2023, often the biggest challenge for a discovery team.

AI is now the best Hacker most teams have ever had. It will even sketch whatever the Designer imagines. But the Designer’s real job is hearing the customer, and the Hustler’s job is finding one to hear. An AI agent will happily draft the recruiting email and send ten thousand copies. It can’t make a stranger care enough to answer. And it can’t sit at a kitchen table and notice the sticky note on the monitor that explains the whole problem.

So the bottleneck moved. When building took weeks, the build hid how few customers a team actually talked to. Now the build takes an afternoon, and an empty calendar of customer conversations is the only thing left between your team and a decision. AI didn’t remove the bottleneck. It moved it to the part most teams were already worst at.

When did your team last sit with a customer? Not survey one. Sit with one. Observe one. Understand one better than they understand themself.

Why can’t a synthetic customer stand in for a real one?

The newest shortcut is the ‘synthetic customer’: ask an AI to play your persona, then interview it. It’s fast, it’s free, and it answers every question with total confidence. What could go wrong?

In the book, I tell the story of sticking my finger in an electrical outlet as a kid, after being told to stay away from it. One hundred twenty volts taught me what no warning could. Every team gets that jolt once. They build something because a bunch of customers said they’d buy it, and then the customers don’t. People rarely do what they say they’ll do. Not because they’re bad people, but because they’re people.

A synthetic customer can only ever say. It can’t do. It never forgets its password, never abandons the cart, never sticks with the clunky spreadsheet it already knows, and never, ever pulls out their credit card and gives you their info. And these models lean agreeable. They’re programmed to be helpful and pleasing to you, and that always drifts toward telling you what you hoped to hear. Ask one whether it would use your product and it will usually find a reason to say yes — even if it’s a two-story outhouse. Jeff warns that this is how you manufacture false traction. I call it interviewing your own reflection.

Use a synthetic customer to rehearse your interview and sharpen your questions. Never use one to answer them.

How do you find real customers when AI makes everything else fast?

The same way we always have, just sooner. Some free advice: start before the prototype is done, not after.

• Name them before you build. Describe your customer well enough to recognize one at the hardware store: the job they’re trying to get done, the problem that ruins their Tuesday, where they go for help. If you can’t say where they gather, you can’t recruit them, and no prototype will fix that.
• Book the conversations the day you start building. With AI, the prototype will be ready long before your calendar is. Put five customer conversations on it before anyone writes a prompt. Run at least three of those interviews until you get a repeating story.
• Go where they already are. The trade association meeting, the online group, the job site, the waiting room, your own customer list. Ask for twenty minutes, not an hour, and offer to share what you learn.
• Watch, don’t just ask. Have them show you how they handle the problem today. The workaround they’ve rigged up tells you more than any opinion they’ll offer. That’s where observations turn into insights.
• Ask for something that costs them. An email address, a signature, a deposit, an hour of their team’s time. Compliments are free. Behavior isn’t. That’s the whole idea behind a good lean startup experiment.

Jeff closes his piece with four changes for anyone who approves new work. Two of them I’d tape to every team-room wall: agree on the customer behavior you expect to see before the work starts, and keep count of the ideas you kill. A team that kills nothing isn’t learning. It’s reporting.

Five questions to ask when someone demos an AI-built prototype

You knew I’d get to questions eventually. If you sponsor or coach a team, these five cut through the polish fast:

1. “Who has used this, by name?” Not ‘users.’ Names. If the answer is ‘nobody yet,’ you just found next week’s work.
2. “What did they do, not say?” Listen for verbs: signed up, came back, paid, quit. ‘They loved it’ is a compliment, not evidence.
3. “Which parts came from a customer, and which came from a model?” Both can be useful. Confusing the two is how polished guesses get funded.
4. “What would make you kill it?” If nobody can answer, nobody is testing anything.
5. “What has surprised you?” It’s one of the twenty questions in my book for a reason. No surprises usually means nobody left the building.

Then stay quiet long enough to hear the answer. That’s the hard part.

The short version

• Let AI tie the lures. Get in the boat and find the fish yourself.
• Book customer conversations the day you start building, or earlier.
• Count what customers do, not what they say.
• Treat synthetic research as a must-test hypothesis, never as evidence.
• Kill bad ideas fast, and keep score.
• And when the prototype is done before lunch, take a customer to lunch.

Frequently asked questions

Can AI replace customer interviews?

No. AI can help you plan interviews, sharpen your questions and sort your notes afterward, but it can’t stand in for the customer. A model can only tell you what someone might say, and the gap between what people say and what they do is where a lot of product ideas die. Use AI to prepare for real conversations, not to skip them.

Should my team use synthetic users or AI personas?

Use them to rehearse, never as evidence. A synthetic user is an AI model prompted to play your customer persona, and it can help you sharpen questions and find holes in your interview guide. It has never felt your customer’s problem, though, and these models tend to agree with whoever is asking.

Is the lean startup still relevant now that AI can build prototypes?

More than ever. The lean startup exists to reduce uncertainty about customers, and AI only reduced the cost of building. Teams can now run more experiments, but every one of them still needs real customers to produce evidence. When building is cheap, testing with real people is what separates a team that learns from a team that just ships more guesses.

How many customers should we talk to before we build?

Start with five to ten people who actually have the problem. You’re looking for patterns in behavior, not statistical proof, and a handful of honest conversations will usually tell you whether the problem is real. If they surprise you, talk to five more before you build anything bigger.

Where do you find real customers to test a prototype with?

Go where they already gather: trade associations, online groups, job sites, waiting rooms and your own customer list. Ask for twenty minutes instead of an hour, and offer to share what you learn. Start recruiting the day you start building, because with AI the prototype will be ready long before your calendar is.