Founder Mode in the Age of AI

These days you hear all the time that with AI you can do everything on your own. So on a Tuesday at 9 a.m., I hosted a Mixer called Founder Mode. (Mixers are Bloom's small paid gatherings, 5 to 7 people around a table.) It was a chance for people who had started a company, or were thinking about it, to meet at a cafe and share their raw thinking. Of all the sessions we ran that day, this early morning one drew the most sign-ups.

Mixers follow what we call the two-pizza rule: conversation works best with as many people as two pizzas can feed, which means 5 to 7. That morning we had a big turnout, so we did introductions together and then split into two tables, mixing people who had founded several companies with first-timers.

The fun of a Mixer is that it runs without a host. But running Bloom, I barely get to network myself, so every now and then I host one. One person's experience goes much deeper than anything you'll find online, and good ideas usually come out of chance encounters.

Founder Mode draws as many "potential founders" as actual founders. These are people who work hard at their day jobs while making a point of meeting founders, the ones at their companies who are unusually tuned in to trends.

Once introductions had gone around the room, I noticed something. In an era when AI supposedly lets you do it all alone, a good share of the people there were looking for a cofounder, a CTO, a developer or a marketer.

The day agents pay each other

The conversation that ran longest at my table was one founder's plan to build a bank whose customers are AI agents. A bank for a world where agents hand work to other agents and pay them for it, rather than a bank for people.

The gap he sees is between verification and payment. On the verification side, checking whether an agent's output is right, companies that verify vibe-coded software have already raised a lot of money. On the payment side, there are machine-to-machine payments like Stripe, Tempo and x402. But nobody has tied the two together yet.

So the settlement he wants to build is closer to holding money in escrow than to processing a payment. The money is released the moment the work checks out and the client approves it. Since nothing is perfect, he says it also needs something to carry the risk, the way an insurer or a credit company does.

I asked whether the card company doesn't already guarantee things when a person pays by card. He said agents fire off tiny transactions every second, and today's banking system can't keep up with that speed.

Payment looks like the last step, but in his view it's actually the beginning. He applied to a big blockchain company's developer program with the same idea and didn't get in. He takes that to mean the market hasn't caught on yet, and he's looking at the US.

We had just run an event with Ripple, so I asked something I'd been curious about. The blockchain industry wants to attach itself to AI too, so it's moving toward agent payments and agentic commerce, where agents buy and sell on your behalf. On the surface it makes sense, but I still can't tell whether blockchain is really the answer or whether the industry is just pulling things its own way.

He cared less about whether it's blockchain and more about the fact that AI spreads without anyone having to be persuaded. It used to be that you had to convince people with logic before the world changed. Now something becomes the norm before people are even convinced, so the power to choose has already slipped away.

He said that when he first really encountered AI, it hit him ten times harder than the Industrial Revolution must have hit someone in the 18th century.

The rules of starting a company are changing

I asked him one more thing: "Do you really need VC money?" People used to raise money because they needed a lot of headcount. AI costs something, but nothing like that.

He said the balance of power has flipped as AI has cut development budgets. Startups no longer go chasing VCs. VCs come looking for startups.

That said, a business like his, with layers of regulatory sandboxes and security, ends up needing big money, and it's a game where the incumbents have to move. He also thinks it won't be easy to raise money for AI in Korea, since the center of the AI world is in the US.

Since anyone can build the same thing quickly, he said, with patents you lock in the date first. Few patent attorneys understand AI, so he drew the flowcharts himself and filed a provisional application. He plans to file the full applications in Korea and the US within a year.

His first exit could come once someone sees the value of this IP and wants it, and the job of growing the business could go to someone better at running customer operations with AI. I asked if he finds his reward in creating an idea and having it validated through an exit. He said yes.

For an AI startup, he thinks two people are enough: one who faces customers and one for everything else. When VCs say they won't back solo founders, it's because they haven't yet felt how much one person can do alone.

A cofounder brings a lot, but you also have to carry a lot. Back when he ran a startup, he deliberately didn't hire experienced people. It was faster to give advice to someone starting from a blank page than to undo the fixed ideas of someone who already had a picture in their head.

Because this is an era when you can do everything yourself, everyone is looking for people, but nobody is looking for just anyone.

Agents still can't book a restaurant

On the same day we talked about a future where agents pay each other, I moved to the other table and confessed the opposite reality. Today's agents can't even book a restaurant for you.

Naver Booking and CatchTable, Korea's big reservation platforms, block outside agents from making reservations. Some similar apps in the US have opened their doors to agents.

Because of restaurant reservations, we close Mixer sign-ups a day or two ahead. If an agent could make the booking, running Mixers would be much easier. As it is, agents can find a place I don't know but can't actually book it, which always bugs me.

The only options are to ask the reservation platforms to open up just for us, or to build a separate booking service made for the AI era. They block agents to prevent abuse, I'm sure, but a person who's determined can also grab every table. Deposits or no-show rules would solve that.

It feels strange now, but someday handing your reservations to an agent will be completely normal.

Naver isn't sitting still either. With the Place Agent that Naver Map launched in September, if a restaurant uses Naver Booking, you can just say the date, time and party size in a chat and it will submit the reservation request. So the door is shut to outside agents but open to Naver's own.

The Claude I use is especially cautious about this kind of task for security reasons. Codex tends to do a bit more.

At the same table was someone who wants to start by changing things from the restaurant side. Restaurants involve too many competing interests, so they plan to widen what they can actually reach: get close to one place first, steer gatherings there, become friends with the owner, and then add reservations.

Now people are the bottleneck

Someone who runs an agency was already handing AI the busywork: emails, inquiries, consultations, quotes. That work took a lot of effort because it needed long explanations. His goal is to focus only on the core of his business and pass the rest of the communication to AI.

But once AI took over the busywork, people became the bottleneck. AI reviews and brings back far more material than he could on his own, so reading it all and pulling it together is left to the human. And because he runs several AIs at once and merges their results, the amount he has to review has actually grown.

When there's no precedent, build the case

The same person brought up another problem. When he proposes something new, like a pop-up, the client first asks for projected sales and precedents, and without data it's hard to put numbers on it. My take was that without past data it's ultimately a hypothesis, so you either make a qualitative case or point to reference cases.

He said the right way to see it is that it will work precisely because nobody has tried it. But the client is the one paying and executing, so you need a separate skill for putting it into language they can accept. The model you show and the model you actually run can't be the same, and that's even more true in the AI era.

If you still need numbers, you can use Nemotron-Personas-Korea, a synthetic dataset of Korean personas that NVIDIA released in April. It has about 7 million virtual people across 1 million records, so you can run a probability model of how many people in each age group and lifestyle would show up for the idea, and use that as your evidence.

It's similar to what consultants have always done. I knew about this dataset but had never thought of using it that way.

We also talked about how the numbers a CEO wants to see differ from the numbers a CEO needs. Heads of professional services firms usually want to see how much marketing traffic turns into inquiries, and whether their expensive staff are producing results worth what they cost.

But the first is closer to a number for checking whether you can cancel the service, and the second is more of a CEO's heartfelt wish. What really changes decisions is how fast the business is growing. You need to see that to decide how far to scale it.

A lot of people just look at this month's revenue and move on. He said he hates that. He likes designs that give clients what they want while laying what they actually need underneath.

He added that people don't differ as much in how deeply they think as you'd expect, and that if you keep looking at things from a certain seat, you become the right person for that seat.

Developers rushing into K-beauty

Someone mentioned that with K-beauty taking off, developers who experienced Korean culture firsthand abroad are starting AI-based cosmetics companies. Ask whether they use AI for marketing or for product development, and the answer is both. Everyone wondered whether people coming from outside cosmetics will do well in this market.

There's also a company that raised a big round by bundling Korean cosmetics into a subscription box aimed at middle-aged and older customers in parts of the US with no Olive Young store (Olive Young is Korea's dominant beauty chain). But it's a red ocean with huge marketing costs: seeding products to influencers, pop-ups, under-the-radar marketing, VIP events for seeding. That also means the market is big.

Beauty leads into plastic surgery and dermatology. He said people around him often point out that when you look at what foreigners spend in Korea, medical spending is far bigger than shopping.

Pharmacies are changing too, as drug companies launch cosmetics as quasi-drugs. Drug companies are best placed to meet the ingredient concentration standards, pharmacies are going back to being drugstores, and there are more pharmacist influencers.

AI has made building things easy, but in this market the weight still sits with marketing and regulation. In other words, the domain.

Your craft over a moat

At lunch I hosted another Mixer, this one on career building in the age of AI. When someone asked how I see careers in the AI era, I gave this answer, which might be a bit surprising from someone who has written a string of posts about moats.

Honestly, I don't really get all the talk about moats. If you're good at your craft, that becomes your moat. The same way a superstar developer gets paid a lot.

AI era or not, being bad at your job is still a problem. What has changed is whether you're good at using AI, and whether you're in an environment where you can use it freely.

What I've been wrestling with lately is that AI has made individuals faster while organizations have stayed the same. If the team structure stays the same, so many POs and so many backend engineers, products ship at the same speed as before.

So I've been thinking about splitting roles so they don't overlap. Overlap requires coordination, but when areas are fully separate, coordination turns into a division of labor. On speed alone, one person is the fastest, and in the early days that speed decides who wins.

A B2B marketer at my table said it depends on the product and the situation. If there's still a lot left to try and polish matters, you need to go deeper. I thought that was right.

Hiring hasn't gotten easier either. I heard people complain that thanks to AI, resumes look great, but when people actually show up they're different, and the gap is worse than it used to be.

Rather than the buzzword "N-jobber," Korean slang for someone juggling several jobs, I feel the era of the individual freelancer is coming. Side jobs are common in the US, but in Korea the rules are tight and it's hard. Rhee Chang-yong, the former governor of the Bank of Korea, has also said the labor market needs to become more flexible. That kind of flexibility might actually be a better setup for everyone.

It comes down to who shows up

At the other table, someone said that even when friends get together to talk about AI, it always drifts back to the usual topics. In the end, what matters most is who's in the room.

In an era when AI lets you do everything alone, what the people at Founder Mode were looking for, in the end, was people.

Originally written in Korean on Brunch, October 21, 2026.