All builds

Business Experiment · completed

Two business briefs, one autonomy experiment

What a month of mostly unattended AI work produced, what the audit found, and what it didn't establish.

The evidence

A recorded walkthrough of two prototypes and Symon's follow-up checks. The original month-long trial ended; no working customer offer or independent replication is claimed.

Who it is for
Hypotheses: local service businesses buying leads, and people sourcing supplement contract manufacturers.
The offer being tested
Local lead generation through RentEngine; access to a structured co-packer directory.
How demand is being checked
Technical checks of pages and directory records. Customer demand was not tested, and payments were deliberately disconnected.
How far it has been built
Two generated prototypes with reported deployments, incomplete user journeys, and data-quality failures.

The question

Could I give an AI service a business brief, leave it to work, and come back to something useful enough to consider charging for?

I tried two briefs in Polsia. RentEngine was a broad system for finding local service opportunities and building lead-generation pages. The co-packer directory had a narrower job: gather useful public manufacturer information into structured records.

I answered early clarifying questions, then mostly stepped away until the subscription was close to ending. The experiment is discussed in a Homebrew Club conversation with Jim, recorded August 11, 2026 and published August 28. My reported cost was $20 for that subscription period.

What I found

Observation from the experiment What it establishes
Roughly 700 files, 70 tasks, 48 reports, and 41 pages reported across the projects A substantial amount of reported production activity.
The co-packer directory returned a 404 despite reports of populated listings The user-facing result didn’t match the report.
15 of 25 RentEngine pages I inspected returned a success status but rendered incorrectly Server status alone wasn’t an adequate acceptance check.
Direct intervention produced 350 directory records Narrower direction changed the result.
My follow-up check with Claude identified 60 of those 350 as real The new row count still wasn’t adequate evidence of usable data.

These are the observations and counts I presented in the episode. The public video is a walkthrough of my audit, not a downloadable dataset that lets a reader reproduce every check. The result should be read with that limit.

What this didn’t test

I left Stripe disconnected. There were no paid customers in the experiment, so zero revenue tells us nothing about willingness to pay for a working version.

It also wasn’t a controlled comparison of AI tools, or a test of whether a more carefully supervised run would succeed. The broader RentEngine brief and the narrower directory brief weren’t equivalent workloads. One ran for less of the month than the other.

I still found the attempt useful. It exposed the gap between work being reported and a result I could trust. It doesn’t establish how the service performs today.

What I’d change in the next trial

Start with one observable outcome and a review early enough to redirect the work. Require an actual visitor journey, traceable records, and access to inspect the live result. Label anything that can’t be checked as unverified.

Then test demand separately, with people who would use the offer. Making the prototype work is a step toward that question, not an answer to it.

The reusable artifact here is the acceptance worksheet. The experiment’s application code and underlying audit dataset aren’t offered as public downloads on this site.

Built with: Polsia, Claude.

Borrow the pattern

Symon He · Newsletter

Practical experiments for your next chapter.

I test ideas with AI and share what works, what it costs, and what you can borrow, so you can choose what deserves your time.