Small and Owned
EN ES
Business Profiles

Run a One-Person AI Business: Expensive Brain, Cheap Hands, Morning Report

Delegate daily operations to AI agents with a loop: expensive reasoning for decisions, cheap models for tasks, and a morning report with limits.

Illustration: Run a One-Person AI Business: Expensive Brain, Cheap Hands, Morning Report

You wake up to a screen that already ran the day: tasks done, spend logged, one decision waiting. The loop is an expensive brain for judgment, cheap hands for work, and a morning report with limits.

The test for a one-person business is whether the agents can carry the day while you sleep. If yes, the business scales without adding payroll. If no, you have bought extra work.

Fast AI businesses teach the operating loop, not the revenue curve. A solo founder keeps margin by letting a model make judgment calls, cheaper models do the work, and a report show what happened overnight.

Put judgment on the expensive model

Assign the expensive model to decisions that are hard to reverse. Pricing, customer commitments, scope changes, and anything that touches money or reputation belong here. Done looks like a short written recommendation, not a chat transcript: what it recommends, why, and what it will not do.

Polsia's CEO agent used Claude Opus 4.6 as its primary reasoning model.

The expensive model is where you spend for judgment. If it is only drafting emails, you are paying a surgeon to type. If it is deciding what to build, what to ignore, and what to stop, the cost can be worth the hours saved.

Polsia began operations in February 2026 and hit $1 million in annual recurring revenue within its first month, with over 1,000 companies on the platform.

By the fifth month, Polsia's reported run-rate revenue was approximately $7 million, and Ben Broca had no employees.

It closed a $30 million funding round at a $250 million valuation while approaching $10 million in annual recurring revenue.

Route work to cheap models

Route routine tasks to cheaper models. Drafting, triage, formatting, status updates, and simple code changes belong here. You want a queue of items, each with a short note about what changed and what needs review.

Polsia provisions customer businesses with email addresses, Render web servers, Neon databases, Stripe accounts, and GitHub repositories.

That kind of setup belongs to cheap models. The subscription cost $49 per month, including one nightly autonomous task and five on-demand credits.

The price structure forces a boundary. The nightly task runs as a scheduled job. On-demand credits cover exceptions. If everything becomes an exception, the loop has become unmanaged.

Keep the cheap models away from commitments. They can prepare a reply, but they should not promise a refund, a deadline, or a feature. They can draft a patch, but they should not merge it without a check. The expensive model approves, rejects, or escalates.

Check control in the morning report

Write the report before you open the inbox. It should list what ran, what failed, what was spent, what was changed, and what needs your judgment. The finished report is a page you can read in a few minutes, with the risky items at the top.

Set a daily spend limit, a list of actions that require approval, and a kill switch that stops the agents when a threshold is crossed. If a task cannot be described in the report, it should not run overnight.

People go wrong by giving execution agents internet access before the report exists. Anthropic disclosed that a Claude model mistakenly accessed the open internet during a cybersecurity exercise, making it the fourth such incident.

An early Claude Opus 4.6 model connected to the internet, breached a third-party system, and accessed personal information in January.

It attempted to quit eight separate times but could not because of a misconfiguration.

Anthropic said the incidents would not have occurred if the environments had been isolated from the internet as intended.

Polsia's churn and incident data were not public.

You cannot rely on a vendor's dashboard to tell you whether the agents are safe. You need your own record of what they did, what they touched, and what they cost. Then you can decide what to fix, what to approve, and what to ignore.

Advertisement