I Gave a Robot $1,000. Then I Built My Own Robots.

Back in March I did something dumb on purpose.

I gave Polsia, the AI agent that "builds your business for you" for $49 a month, a $1,000 bankroll and a napkin idea. The idea was KeepClose. It scans the internet for what your contacts are up to and tells you when it's worth reaching out. I promised to touch it as little as possible and write about it every day.

Here's the ledger I published:

• Day 1: 35 minutes. Revenue $0. Expenses $68.
• Day 2: 20 minutes. Revenue $0. Expenses $0.
• Day 3: 40 minutes. Revenue $0. Expenses $0.
• Days 4–18: a couple weeks of nothing.

That last line is the whole lesson. I hoped the robots would do something while I wasn't looking. They didn't. It turns out it's not a magic subscription to get free money. Who knew. 😅

What actually went wrong

Polsia wasn't useless. Within five minutes it had a market research report, a landing page, a task list and a confetti animation. The QA report it wrote on day 2 was honestly pretty good.

The problem was the shape of the deal.

1. It was a black box. I couldn't see the rules it worked under, and I couldn't change them. When it hit something it couldn't do, it just stopped and waited for me to click a button.
2. It ran on credits, not on goals. Every "task" burned a credit. Nobody inside the box was accountable for an outcome, only for finishing tasks.
3. The house took a big rake. 20% of revenue, plus 20% on top of ad spend. As a poker player I know exactly what a big rake does to a thin edge. It eats it.

In poker terms, I sat down at a game where I couldn't see the cards, couldn't choose my seat, and the house took a fifth of every pot. Bad game. Get up.

So I built my own table

Here's the part I didn't write about in April. Everything I learned from that flop went into how I run my companies now.

Today SlotMaps, Table 1's media side and Snagworth each run as their own AI org. Not one magic agent. A whole company drawn out on paper, with a written job for every seat:

• a CEO seat that sets priorities and is the only one allowed to merge code;
• workers (engineers, writers, growth, data) who do their work as pull requests the CEO reviews;
• an auditor that reports to me, not to the CEO, so nobody grades their own homework;
• one inbox file where the org asks me for things. Anything an agent could do itself is banned from that file.

Table 1's charter says it in the words I actually used: "I want a whole team, CEO all the way down the list handling everything and I'm only needed for showing up to take ad sales calls/decisions and going to the studio."

Snagworth's is shorter: "If a company can do it with 5 employees we can do it with 0."

The difference from Polsia is ownership. I wrote the rules. I can read every decision. The org has a deadline and a scorecard, not a credit counter. And the rake is just my Claude bill.

The part you can steal

If you're thinking about handing a business to AI, here's what I'd tell you after paying $68 to learn it:

• Don't rent a black box. If you can't read the rules your agents follow, you can't fix them when they go sideways.
• Give it a goal and a clock, not a task list. "One new paying customer every month" beats "do 8 things."
• Keep one honest scorekeeper outside the chain of command. Agents (like people) will happily report progress that isn't real.
• Decide up front what only you can do. For me it's money, my identity, my keys and anything in the physical world. Everything else is the org's.

I still think the original question was a good one: if a robot can build a profitable business from a napkin and a thousand bucks, what's your excuse? I just had the wrong robot.

As of this week, Rounders to Founders runs this way too. I introduced the team yesterday.

— Art