Andrew Wilkinson has made hundreds of millions of dollars buying and building companies.
And right now he's waking up at 3 in the morning to sit in a terminal.
Not because he has to. Because he can't stop.
I watched his latest sit-down with Greg Isenberg, and it's the most honest look I've seen at how a guy at the very top is actually using AI agents day to day. Not the hype version. The real version, with the dollar amounts attached.
So let me break down what he's doing, the numbers behind it, and what you can steal from it as a founder who is not sitting on a billion-dollar holding company.
The December moment
Andrew describes December 2025 like a switch flipped.
He'd played with Claude Code and Replit earlier in the year. It was interesting. It wasn't there yet.
Then in December something changed. His words: "I feel like I started shooting heroin."
He started rolling out of bed at 3 or 4 in the morning with a big smile, sitting in the terminal with 10 tabs of Claude Code open. He's been doing it ever since.
I bring this up because the energy matters. The founders who are going to win the next few years are not the ones who took a course on AI. They're the ones who got obsessed, opened the terminal, and started building ugly stuff at 3am.
You don't need permission. You need a problem you can't put down.
He replaced his payroll with a $40,000 Claude bill
This is the line that stopped me.
Andrew runs a family office (the rich-person word for a personal holding company). His CFO and his president both got obsessed with Claude Code.
So instead of scaling employees, they scaled the API cost.
His exact words: "Instead of a payroll, we just have a $40,000 a month Claude bill. And it's doing everything."
Now, $40K a month sounds like a lot. But think about what a CFO, a president, and a team of analysts cost. He's running the back office of a conglomerate on the price of one mid-level hire.
The example he showed: a product called Addepar that lets rich people track their portfolios. It charges $50,000 to $100,000 a year. His team just built their own version internally with Claude. Software that used to be a six-figure vendor line is now a thing his team vibe-coded in a weekend.
That's the pattern to watch. A lot of "enterprise software" is about to get rebuilt internally by small teams who got good at agents.
He built a vector database on his entire investment history
Here's where it gets wild for anyone running multiple things.
Andrew built a vector database on his family office, Foley Partners. Then he just asked it questions in plain English.
He asked: break down how many minority venture investments I've made, and how many are in the money, bankrupt, or declined.
It went off, ran the query, and came back: 132 direct investments. $16 million invested, now worth $36 million. Plus a full breakdown of write-offs and everything else.
He did the same thing for Tiny, where he owns something like 24 businesses with years of historical data.
The problem when you run a conglomerate is you can't actually hold it all in your head. Are any companies spending too much? Are any CEOs full of it? Are any businesses quietly trending the wrong way? You used to need a team of analysts to answer that. Now he just asks.
You probably don't have 24 businesses. But you have a Stripe account, a Postgres database, a pile of customer emails, and a bunch of questions you never get around to answering. Same move. Dump it into a place an agent can read, then ask.
The app he's actually building: Deep Personality
The best part is he's not just optimizing the empire. He's building a new product like a regular indie hacker.
It started personal. He and his wife took a deep personality test, exported the JSON, dropped it into ChatGPT, and asked it to describe their relationship without knowing anything about them.
It nailed every fight they have. Every issue at home. Their jaws dropped.
So he saw the product opportunity and started building Deep Personality.
The early numbers: about $20,000 in revenue so far. Small, and he says so. But here's the interesting part. He hooked the app into PostHog, and the agent manages a Meta and Reddit ads account on its own. It runs multivariate tests, creates the ad creative, and sets the budgets.
His next move is to hand it a $100,000-a-month ad budget and see what happens when the agent is the one finding the winning creatives.
If you've ever wanted to test the "AI runs my growth" idea, this is the cleanest version of it I've seen described. Small product, real revenue, agent on the ad account, and a founder willing to pour fuel on it.
For more on where these AI-native products are headed, this pairs well with our roundup of the best startup ideas for 2026.
The 24-hour app
One more story that shows how fast the floor has dropped.
Andrew was annoyed that the Limitless pendant (the wearable that records your day) existed when the iPhone already has a microphone. He said out loud, "Why can't the iPhone just record all the time?"
Someone he knows went, "Well, I can build that."
24 hours later there was an app called Hearsay. You set what times of day you want it to trigger, it records your day, builds transcripts, and ships them to iCloud. Andrew now feeds those transcripts into his agents so they have full context on his life.
An idea spoken out loud on Monday was a working app on Tuesday. That's the speed we're all competing at now.
His honest advice for a 20-year-old who can vibe code
Greg asked him the obvious question. If you were young and obsessed with vibe coding right now, what would you do?
Andrew's answer was refreshingly un-hyped.
Make one to two million dollars by building a product like Deep Personality. Don't try to build the next billion-dollar company on day one. Build a real thing that real people pay for, and get yourself to a couple million.
Then (his words, and he admitted it was a sad answer) take that money and invest it in the picks-and-shovels: data center stocks, TSMC. Because that's what he's actually doing.
I don't fully agree with the "go buy chip stocks" ending. But the first half is gold. The opportunity right now is not waiting for a unicorn. It's building a tight product, getting it to real revenue, and compounding from there. A one-to-two-million-dollar business is a fantastic life.
What's still broken
He was honest about the pain too, which I appreciated.
He reckons he spends 50% of his time debugging the agents, 30% making them better, and only 20% actually being productive. Open Claw is "hard but amazing when it works."
And the real ceiling is memory. With a 1 million token context window, an agent can maybe remember a day. He compared it to the movie Memento, where the guy wakes up with no memory every morning. Until context windows hit 5 or 10 million, agents can't truly run a whole company end to end.
So if you've been feeling dumb because your agents keep forgetting things and breaking, relax. The richest, most obsessed user in the room is hitting the exact same wall.
What I'd actually take from this
Strip away the family office and the chip stocks, and here's the founder playbook hiding in this conversation:
Get obsessed, not educated. Andrew didn't take a course. He opened the terminal at 3am and didn't stop.
Put your data where an agent can read it. Your Stripe, your DB, your customer emails. Then ask it the questions you've been avoiding.
Let an agent run one real thing. For him it's a $20K product with an agent on the ad account. For you it might be support, or onboarding, or one growth channel. If you want to turn that skill into income directly, here's how founders are building an AI services business off the back of it.
Aim for a couple million, not a unicorn. A tight, profitable product beats a lottery ticket.
This is the exact stuff we trade notes on inside Profitable Founder Club: what's actually working with agents, what's a waste of time, and the real numbers behind it. No theory, just founders comparing receipts.
FAQ
How is Andrew Wilkinson using AI agents in his business?
He's using tools like Open Claw and Claude Code to run his family office back office (replacing payroll with a ~$40,000/month Claude bill), to query his entire investment history through a vector database, to build and grow a new app called Deep Personality, and to manage his personal life through daily transcripts.
What is Deep Personality?
It's an app Andrew started after a deep personality test described his marriage scarily well. It's done about $20,000 in revenue so far, and an AI agent runs its Meta and Reddit ad accounts (creative, budgets, and multivariate testing) through a PostHog integration.
Did he really replace employees with AI?
In his family office, largely yes. His CFO and president got obsessed with Claude Code and scaled API spend instead of headcount, landing on roughly a $40,000/month Claude bill that handles a lot of what a back-office team used to do.
What's the biggest limitation he hit?
Memory. He says today's ~1 million token context window lets an agent remember about a day, like the movie Memento. He thinks agents won't run entire companies until context windows reach 5 to 10 million tokens.
What's his advice for young builders?
Build a real product (like Deep Personality), get it to one or two million dollars in revenue, and don't obsess over chasing a billion-dollar outcome on day one.
Want more founder stories like this?
I break down a real founder's numbers and tactics every week on the Profitable Founder Podcast. Bootstrapped SaaS, indie apps, the whole thing, with the receipts.
