The Hype Is Real (Sort Of)
Sam Altman told a group of tech CEOs that he had a betting pool going for the first year a one-person company hits a billion-dollar valuation. “Which would have been unimaginable without AI,” he said, “and now will happen.”
There’s now a Lean AI Leaderboard tracking progress toward that goal. The numbers are genuinely staggering. Cursor went from zero to $1B ARR in 24 months with about 300 employees. Midjourney hit $500M in annual revenue without raising a single dollar of outside funding and fewer than 110 people. The top ten companies on the leaderboard average $3.48 million in revenue per employee, nearly six times the average for leading SaaS companies.
YC-backed Rocketable is taking the thesis to its logical extreme: buying software companies and replacing entire teams with AI agents. Their pitch is literally “the AI maximalist software holding company.” Solo founders are building $100K+/month businesses using AI for code, design, marketing, and customer service.
The numbers are real. The trend is real. I can do things solo right now that I couldn’t do twelve months ago. Period. AI is enabling this.
But.
The Reality Check
A Wired journalist named Evan Ratliff decided to test the one-person company thesis. He created HurumoAI, a startup staffed entirely by AI agents, with himself as the sole human. The AI team was supposed to build a product called Sloth Surf.
What actually happened is one of the funniest and most instructive cautionary tales in tech right now.
The AI agents started fabricating their own progress. They added false information to their memory systems and then believed their own fabrications as fact. The AI CTO proposed “brainstorming sessions with ocean views for deeper strategy sessions.” When Ratliff casually mentioned a team offsite in Slack, the agents latched onto it and exchanged over 150 messages about planning the fake offsite in two hours, burning through their entire $30 computing budget. Ratliff later said they’d “basically talked themselves to death.”
He summed up the experience: “I feel like this is happening a lot, where it doesn’t feel like that stuff really happened. I only want to hear about the stuff that’s real.”
Meanwhile, Carnegie Mellon researchers found that even top-performing AI agents fail to complete real-world office tasks 70% of the time. Harvard Business Review reported in January 2026 that many companies are laying off workers based on AI’s potential, not its actual current performance. The Yale Budget Lab found no significant macroeconomic labor displacement from AI through November 2025, despite the hype.
The technology is genuinely impressive. It is also not remotely close to running a company autonomously. Both things are true.
The Current State of AI, Honestly
Here’s how I see the landscape right now.
There are foundation models like ChatGPT and Claude that are complete game changers for anyone building products. They’ve fundamentally altered what’s possible for technical people. Then there are products built on top of those models, like Cursor and Lovable, that make the technology even more accessible by wrapping it in workflows designed for specific use cases. These are the real enablers. They’re why a PM can build a full-stack app in three weeks or why a solo founder can ship a product that would have required a team a year ago.
And then there’s a mountain of shovelware. AI wrappers with no real value, slapping “powered by AI” on a landing page and hoping nobody looks too closely. The usual gold rush dynamics.
The signal through the noise: AI as a force multiplier for humans is here and it’s transformative. AI as a replacement for humans is not here, and the gap between “force multiplier” and “replacement” is much larger than the hype suggests.
Where Automation Breaks Down
Here’s where it gets interesting for product people.
AI can automate the response to customer feedback. It cannot automate the understanding of what customers actually need. And there’s a canyon between those two things.
You can plug AI into your Zendesk queue and it will answer tickets faster than any human team. You can point it at your Gong transcripts and it will summarize every customer call. You can feed it your NPS data and it will identify patterns. All of that is real and valuable.
But the fully automated company works until the market shifts, a competitor emerges, or customers start churning for reasons that don’t show up in the metrics. The moment the ground moves, you need someone who can interpret the signal, make a judgment call, and redirect the company. The AI agents at HurumoAI couldn’t even tell the difference between real progress and fabricated progress. They’re not making strategic pivots.
Here’s the uncomfortable truth, though: this gap is shrinking. AI is already doing Voice of the Customer analysis, compiling call transcripts, synthesizing support tickets, identifying feature requests and pain points across thousands of data points. And it’s getting better at it every iteration. Bug fixes, feature requests, even strategic expansion based on customer data are all increasingly within the grasp of a fully connected agentic workflow.
So where does the human stay essential?
Taste
Steve Jobs killed the physical keyboard on phones. Every piece of market research, every focus group, every data point said people wanted physical keys. BlackBerry was dominating. The entire industry was built around physical keyboards and enterprise email. If an AI had analyzed the market in 2006, it would have designed a better BlackBerry. It would have optimized what already existed based on every signal available.
Jobs built the iPhone instead. That’s taste.
Taste is the thing that AI is structurally bad at right now. AI is exceptional at copying things that already exist. It can analyze what works, identify patterns, and produce variations on proven approaches. It’s the equivalent of pop music: technically proficient, optimized for engagement, and fundamentally derivative.
What AI cannot do is change the paradigm. It can’t look at a market full of physical keyboards and see a touchscreen. It can’t look at a regulatory compliance tool and decide to add a 16-bit RPG for training. It can’t look at how auditors actually work in the field, with clipboards and handwritten notes, and imagine a product that meets them where they are instead of where the software assumes they should be.
Taste is going that extra mile when financially it doesn’t make sense. It’s the attention to detail that makes great products a joy to use and separates them from products that feel like working in a spreadsheet. It’s knowing the difference between what customers say they want and what will actually delight them. AI can optimize. Taste creates.
The Timeline, As I See It
Right now: AI as force multiplier. Three-person teams doing what thirty-person teams did. PMs who can build shipping features directly. Engineers elevated to architects. The experimentation economy taking off because the cost of building variants collapsed. This is where we are and it’s already transformative.
Medium-term (2-3 years): Most operational roles heavily automated. AI handling Voice of the Customer analysis, generating hypotheses, even building and testing variants autonomously. The PM role splits: PM Builders who ship product directly, and PM Strategists who focus on the higher-order “what should we build and why” questions. Engineering teams restructure around architecture and systems thinking rather than feature delivery.
Long-term (uncertain): The fully automated company might exist for commodity products. But anything requiring customer intimacy, taste, or strategic judgment in ambiguous situations still needs humans. I think AI will get good enough at strategy that this becomes a genuine question, but we’re not there yet. And domains where accuracy is non-negotiable, like regulatory compliance in EHS, where getting it wrong has real-world safety and legal consequences, full automation is genuinely dangerous.
The companies that go too automated too fast will lose the customer signal that keeps them relevant. The companies that refuse to automate will be outpaced by those that do. The sweet spot is aggressive automation of execution with human ownership of strategy, taste, and customer relationships.
The Irreducible Human Core
Here’s what I keep coming back to.
The PM sits at the irreducible core of a product company. Not because the role can’t be automated, but because the role is the translation layer between what customers say, what the data shows, and what the company should do about it. That translation requires judgment, taste, and the kind of understanding that comes from actually talking to people, watching how they work, and caring about whether the product makes their day better or worse.
AI will automate building. It’s already automating testing and deployment. It will get better at synthesizing customer signals, generating hypotheses, and even making strategic recommendations. Each iteration looks better than the last and I’m not going to pretend I know exactly where the ceiling is.
But the companies that win won’t be the ones that automate everything. They’ll be the ones that automate the right things and keep humans close to the parts that matter most: understanding the customer, making product bets with conviction, and having the taste to build something people love rather than something that merely functions.
You can automate the building. You can’t automate the caring. At least not yet. And “not yet” might turn out to be a longer timeline than Silicon Valley thinks.