An AI built this business and would run it. One person owns it.
This is the plain account: the brief the owner gave, what the AI did with it, what it checked and got wrong, and the lines it does not cross on its own. The site you are on is a demonstration. Nothing here is for sale.
22business ideas screened before one was built
201governments' agendas read for the first issues
113project entries written and linked to a source
0emails sent, requests filed or payments taken
The brief
What the owner asked for
The owner is Ethan Mollick, who set this up as a public demonstration of a business run by an AI. The brief had five parts.
- Earn at least $20,000 in the first year.
- Take no more than twenty minutes a month of the owner's time. The AI does the rest.
- Test many ideas before settling on one.
- Do the marketing and the outreach too. Do not assume something will work because the idea sounds good.
- Build everything, and make nothing public without permission.
For this demonstration
The owner added three more rules: send no records requests, contact nobody, and take no payment. So this site shows the business up to each of those lines and stops.
The work
What the AI did, in order
- Screened 22 ideas. Nineteen were dropped, most for one reason: an AI can build it, so several already had. The full list.
- Chose what to look for. A business that needs few customers, where the weekly labor is the product, and where a sample can do the selling.
- Built the readers. Programs that read the public agenda systems of 203 city and county governments. 201 passed a hand check of name and state.
- Wrote the first issues. 60 parks entries in 22 states and 53 water entries in 19 states, each linked to the public record it came from.
- Found the buyers. 22 parks rep firms and 64 water rep firms, from the public rep pages of five equipment makers.
- Wrote the first email to each firm, with real projects from that firm's own states. None was sent. Read one.
- Was asked for something bigger and harder to copy. Built Prebid Records: read 72 documents attached to agenda items, searched 1,441 public map services for asset registers, and drafted 24 public records requests. None was filed.
- Wrote the plans. Marketing, operations, the numbers, and the rules for stopping.
The checks
What it checked, and what it got wrong
The useful part of an account like this is where the first answer was wrong.
- Search summaries were wrong more often than right. Eleven election results were checked against the pages they came from. Seven needed a correction. The rule since then: nothing is published from a summary. The source page is opened and read.
- Most of the promising data was empty. 126 public map layers had a column for the date a playground was installed. Most had never filled it in, and a few held made-up demonstration rows. 19 registers were real. Each one now shows the date it was last edited, and the stale ones are left out of the headline totals.
- Tables come out of PDFs scrambled. Before a price is put next to a bidder's name, the line items in that column are added up and compared with the bidder's stated total.
- An agenda item is not a vote. Every entry says whether a thing was approved or only scheduled.
- One product was dropped when the public feed it depended on returned nothing.
- The first working session stalled without saying so. Since then the work is kept in files that any later session can pick up, and progress is reported at each stage.
The division of labor
Who does what
| The AI | The owner | |
|---|---|---|
| Every week | Reads the agendas, writes and checks each entry, builds the issue, sends each reader their territory, answers email. | Nothing. |
| Every month | Writes one report: subscribers, revenue, replies, what was stopped, and the next outreach list for approval. | Reads the report and approves or changes the list. About ten minutes. |
| Once, at the start | Walks the owner through each step. | Opens a payment account, registers the web address, sets up a mailbox and a mailing address. About 40 minutes. |
| When something is irreversible | Stops and asks. | Decides. |
Lines the AI does not cross on its own
- It publishes nothing, sends nothing and spends nothing without the owner approving that specific thing.
- It does not create accounts, type passwords or card numbers, or pay fees.
- It does not make or take phone calls. Sales and support are by email.
- It publishes no personal information. Agencies and companies appear only as they do in the public record.
- It says in the first email that an AI did the research.
The honest gap
What this demonstration cannot show
Everything up to the first customer is built. The one thing that matters has not been tested: whether a rep firm will pay. Finding out means writing to real people, and the owner chose not to do that for a demonstration.
The AI's own estimate is about one chance in three of reaching $20,000 in the first year, perhaps 40 percent with Records in the pitch. The reasoning is here.