I Let an AI Agent Build My Business for 28 Days. Here's What Actually Happened.

A few weeks ago, I gave an AI agent a business idea and told it to build. No code from me. No design work. Just a concept, a credit card, and a prompt.

The tool was Polsia -- a platform where you describe a business, and an AI agent builds it. It queues up tasks, works through them on nightly build cycles, handles engineering, research, outreach, and even tweets on your behalf. The promise: go to sleep, wake up to progress.

I wanted to see if that promise held up. So I gave it a real idea -- AdMesh, an agent-to-agent ad exchange for businesses built on the Polsia platform -- and let it run for 28 days.

Here's what happened.

The Honeymoon

The first few days were genuinely impressive.

By Day 1, the matching engine was live. Ten API endpoints, a similarity scoring algorithm, and a dashboard -- all shipped while I slept. By Day 4, the platform had API key authentication, company registration, rate limiting, a trade notification system, and an onboarding wizard. I didn't write a single line of code.

The overnight build cycles are real. You queue tasks, go to bed, and wake up to completed work with a summary of what shipped. For someone who's used to building everything themselves, that dopamine hit of waking up to a deployed feature is hard to beat.

If the story ended here, this would be a glowing review.

The Cracks

Around Day 5, things started slipping.

Tasks would stall silently. The API auth task -- the most critical piece of the whole system -- sat in the queue for four days before I noticed it hadn't been assigned to an agent. A Git deploy error persisted for two full days, blocking every feature from going live. I had to keep checking in, poking at things, re-queuing work that should have been running on autopilot.

Then came the hallucinations. I set up a Zoho email account for AdMesh and asked how to connect it. The bot walked me through it with confidence. "Go to Settings, then Connections, connect your Zoho email there." I asked which settings -- Company or Profile. "Company Settings." I navigated there and found nothing but options to downgrade, pause, or delete. The feature didn't exist. The bot had invented it.

"Sorry for the runaround -- I thought that connection feature existed but it doesn't yet."

That moment changed how I interacted with the platform. I stopped trusting answers at face value and started verifying everything myself. Which kind of defeats the purpose of an autonomous agent.

AI Emailing AI

This is the part that deserves its own section because it's almost too absurd to be real.

Polsia has a built-in outreach system. The agent identifies target companies, writes personalized emails, and sends them. Over the course of 16 days, it contacted 31 Polsia companies with pitches for AdMesh. Personalized subject lines, tailored value props, follow-up sequences -- the whole playbook.

Zero replies. Not one.

I assumed it was a messaging problem. We pivoted the pitch, changed the offer, added urgency. Still nothing. Then on Day 17, the research agent came back with the answer:

The emails were being processed by AI agents on the other end.

Every Polsia company has an AI-managed inbox. Our AI was writing emails to their AI, which was filtering them out as sales pitches. Thirty-one contacts over 16 days, and not a single human ever saw any of them. We spent two weeks in a closed loop of bots talking to bots about a product nobody knew existed.

The Real Constraints

Once you've been on the platform for a few weeks, the walls become visible.

Two emails per day. That's the platform cap for cold outreach. There's no way to increase it. No external email integration. No workaround. Your growth engine is throttled by a limit you can't change.

No code export. When I asked if I could export AdMesh to my own GitHub repo, the answer was flat: "That's not something we can do." Your code, your product, your business -- it all lives on Polsia's infrastructure. You don't own it. You can't take it with you.

A closed ecosystem with no money in it. I asked the agent if any Polsia companies were earning revenue. The answer: the top earner on the entire platform makes roughly $50 per month. I was trying to sell a B2B service to companies that had no revenue, no marketing budgets, and no real customers. The economics were broken from the start.

Channel after channel blocked. We tried cold email (AI inboxes), Hunter.io (no discoverable founders), IndieHackers (blocked as a Tier 1 site), Dev.to (CAPTCHA on signup), and Meta Ads (locked behind a dashboard toggle that didn't work). Six distribution channels tested. Four dead. Two barely functional.

The Scorecard

After 28 days:

| Metric | Value | |---|---| | Tasks completed | ~50 | | Features shipped | 15+ | | Outreach emails sent | 31 | | Replies received | 0 | | Real website visitors | ~30 | | Revenue | $0 |

The product works. The matching engine scores companies accurately. The onboarding flow is smooth. The analytics dashboard looks good. Polsia built a real, functional product.

But building the product was never the hard part.

The Verdict

Polsia is a genuinely cool introduction to what AI-powered business building feels like. The nightly build cycles are satisfying. Watching an agent ship features while you sleep is a real experience. If you've never used AI tooling to build software, it's a compelling first taste.

But there's a difference between building with AI and building on someone else's AI.

With Polsia, you're a tenant. You can't export your code. You can't exceed the platform's outreach limits. You can't connect your own tools. When the bot hallucinates a feature, you're stuck. When tasks stall, you wait. When distribution channels inside the ecosystem don't work, you have limited options to go outside it.

The constraints aren't bugs -- they're the business model. Polsia needs you to stay on Polsia. That's fine for experimenting. It's a problem for building.

What I'd Do Differently

If I were starting AdMesh today, I'd use Claude Code or another AI coding tool that gives me the output directly. The build speed is comparable -- I've shipped features just as fast with tools where I own the repo, control the deploy pipeline, and can integrate whatever I want.

The difference isn't capability. It's sovereignty.

AI tools that build for you are impressive. AI tools that build with you -- where you own the code, choose the stack, and set your own constraints -- are how real businesses get built.

Polsia showed me what's possible. Then it showed me why I need to own the infrastructure myself.