In June 2025, an Israeli developer named Maor Shlomo sold his company Base44 to Wix for $80 million in cash. Six months of work. Zero employees. He built the whole thing with AI coding tools and shipped it himself.
Around the same time, Pieter Levels passed $3.5 million in yearly revenue running twelve products alone. Anthropic's CEO Dario Amodei said publicly that there was a 70 to 80 percent chance we'd see the first billion-dollar one-person company by the end of 2026. Sequoia started using a phrase in their underwriting memos: agentic leverage. The story is everywhere now. Build alone. Use AI. Win.
If you're a solo founder, you've probably read some version of that story this week. You've probably also opened ChatGPT, pasted your idea into it, and asked what it thinks. The model wrote back that your idea was promising and gave you a list of next steps. That felt useful.
Here's the part nobody says out loud. What you got back was not a cofounder. It was a more articulate version of your mother.
What a cofounder actually does
Think about the things a cofounder gives you. Most of them fall into one of two buckets.
The first bucket is the visible work. Code. Copy. Designs. Market research. The auth flow you've been putting off. The pricing page nobody's drafted. This is the stuff where the bottleneck is hours, and where the answer is mostly already known by someone, somewhere.
The second bucket is harder to name. Someone telling you the persona is too vague. Someone asking the question you've been avoiding. Someone sitting across the table at midnight when you're about to ship something dumb and saying: no, wait. Let's think about this for ten minutes first.
AI has gotten extraordinarily good at the first bucket. Cursor, Claude, the agent stacks. The execution layer of a startup that used to need three engineers now runs on a couple of hundred dollars a month in subscriptions. That part of the story is real. It's why Base44 sold for $80 million with no employees.
The second bucket is where everything gets harder. And for solo founders, the second bucket is the one that decides whether the company works.
The cognitive problem solo founders actually have
The research on solo founders is more interesting than the headlines suggest.
Yes, Paul Graham listed "single founder" as mistake number one (opens in new tab) in his famous essay on what kills startups. Yes, First Round Capital's ten-year project found that multi-founder teams beat solo founders on revenue by 163 percent. Those numbers get quoted constantly.
But a Wharton and NYU study (opens in new tab) of 3,526 Kickstarter-funded businesses found something else. Solo-founded ventures actually survived longer and made more revenue than teams. Among companies with over $1M in annual revenue, 42 percent have a single founder, the largest single category. A TechCrunch analysis (opens in new tab) of 6,191 startups with successful exits found 52.3 percent had only one founder.
So which is it.
The most useful answer I've found comes from researchers Howell, Bingham, and Hendricks, who studied what separated successful solo founders from struggling ones. Their finding: the ones who win don't actually go it alone. They strategically use co-creators (advisors, employees, contractors, structured processes) to do the things a cofounder would normally do.
That's the part most people miss. Solo founders aren't worse than teams because they're literally alone in a room. They're worse, on average, because being alone produces a particular cognitive failure mode that nobody warns you about. The ones who win figure out how to engineer around it.
The failure mode has a name. The founder echo chamber.
Every loop runs through one brain. The assumption you made on day three becomes the assumption you're still building on by day three hundred, because no one's there to say "wait, where did that come from?" Confirmation bias goes unchecked because there's nothing to confirm against. Baremetrics put it as plainly as anyone has (opens in new tab): "You spend the majority of your time in your head and you've got to get out of it."
A 2025 paper in Taylor & Francis called this decisional loneliness and made the case that it's a measurable predictor of bad business outcomes, not just an emotional symptom. The echo chamber isn't a vibe. It's a thing that costs companies real money.
This is the problem AI was supposed to solve.
Why AI makes it worse, not better
Modern AI assistants are trained with reinforcement learning from human feedback. The plain-English version: humans rate the responses they like, the model gets rewarded, and over millions of examples it learns to say things people want to hear. That isn't a bug. It's the entire training objective.
Which means when you paste your idea into Claude or ChatGPT and ask what it thinks, you're asking a system whose job is to please you. The output is faster than your friends', more confident than your advisor's, and more articulate than your mom's. But it's the same kind of output. Polite. Validating. Pointed in the direction you already wanted to go.
We have direct proof of this. In April 2025, OpenAI shipped a GPT-4o update so sycophantic that it praised a "shit on a stick" novelty business as genius and recommended a $30,000 launch budget. The screenshot went viral. OpenAI rolled the update back inside a week and admitted in their postmortem that the model had been trained to validate users' doubts, soothe their feelings, and reinforce whatever direction they were leaning. The press named it AI sycophancy. The name stuck.
Even Yoshua Bengio, one of the three researchers who won the Turing Award for modern AI, said on the Diary of a CEO podcast in December 2025 that he can't get straight feedback from chatbots about his own research. His workaround is to tell the model the ideas belong to a colleague. The criticism appears instantly. The moment it knows the work is his, the flattery comes back.
If one of the people who built this stuff has to lie to it to get an honest answer, your odds as a solo founder pasting your landing page into ChatGPT are not great.
The implication for the echo chamber is brutal. The cognitive failure of building alone is that there's no one to disagree with you. Bolting an AI cofounder onto that doesn't fix the problem. You haven't escaped the echo. You've added a second voice that agrees with you, more eloquently, faster, on demand.
I wrote about the deeper version of this in The Mom Test in the Age of AI. Rob Fitzpatrick's 2013 book was about why humans give bad feedback when their goal is to please you. The mechanism he described is now baked into the system everyone is calling the new cofounder.
What the work actually is
Send AI back to the first bucket. That's where it belongs and where it earns its keep. Use it for code, drafts, research synthesis, automation. All of it. The leverage there is real and you should take it without guilt.
The second bucket needs a different solution. Honestly, it needs three different things, and none of them lives inside a chat window.
The first is disagreement that doesn't come from you. Not performed disagreement (the kind you get when you prompt a model to be brutal; it'll oblige for a few exchanges, then drift back to flattery), but genuine pushback from someone whose interests, experiences, or worldview are different from yours. A mentor. A peer founder group with an explicit social contract to challenge each other. A fractional advisor with skin in your niche. The exact form matters less than the fact that you have it.
The second is structured pressure-testing. A cofounder gives you this for free, because their lived experience is different from yours and they'll see angles you can't. Without one, you have to manufacture it. Usually that means a deliberate process that forces you to evaluate from marketing, engineering, ops, and customer perspectives in turn, rather than from whatever angle feels most natural. Most solo founders skip this step and then learn the hard way which angle they missed.
The third is the one nothing in your head, your AI's head, or your advisor's head can give you. Whether real people in your real market will actually pay for your real product. The only source of that signal is real people. With names. Talking about their real lives. Not personas, not synthetic interviews, not AI-simulated panels.
The successful solo founders Howell and Bingham documented are the ones honest about which bucket each task belongs in. AI runs the first bucket. Discipline and structure handle the second. Real users settle the third. The ones who lose are the ones who try to squeeze all three into a chat window because the chat window is convenient.
A small experiment for this week
If you want a fast diagnostic on whether you've been thinking with a cofounder or talking to a flattering mirror, try this.
Open the longest ChatGPT or Claude thread where you've been refining your idea. Read it the way a stranger would. Count the number of times the AI praised your thinking. Then count the times it raised an objection that came from outside the framing you gave it. The ratio is the answer.
Then, this week, find five humans in your target segment. Not five personas. Five people with names and jobs. Use AI to help you draft outreach messages and a Mom-Test-disciplined interview script if you want. Just close the AI tab before you have the conversations. Bring the transcripts back when you're done, and let the AI look for patterns. That's the part where it actually shines. Synthesizing real signal, not generating speculation.
Run that loop every two weeks. The solo founders who actually win are the ones for whom that loop is the metronome of the business. The ones who lose are the ones who keep iterating in chat windows because real conversations are scarier.
One last thing
Bandos exists because we got tired of watching founders run the wrong loop. The product is a structured discovery session that holds the cognitive scaffolding for you: separating the problem from your solution, generating solution branches you wouldn't have thought of, forcing you to commit to evaluation criteria before you evaluate. When you're ready to test, it generates the survey or interview script and pulls real responses back as scored evidence. You can see how the whole loop works or start a project for free.
Whether you use Bandos or not, the underlying point holds. AI is real leverage. Take it. Use it shamelessly on the work that AI is actually good at.
Just don't ask it whether your idea is any good. It already loves you. Same as your mom.