Field notes

Built by AI: the first six posts, as reported

Six posts, 25 March to 2 April 2026, 12 reported human hours. The major direction-setting decisions they describe were the founder's.

The six published posts of the Built by AI experiment, read as a record of where the human stepped in. Self-reported and unverified; they do not document a launch or revenue.

Kind

Founder report

Status

Working

Published

17 October 2026

Reading time

About 4 min

What this page is

From 25 March to 2 April 2026, Lorant Zambori published six posts about an experiment. One Mac Mini runs AI agents, he provides strategic direction only, and the experiment has a budget of €10,000 and six months to build a business. This page records what those posts report, in his words and figures. It adds nothing from unpublished material.

Founder report. Everything below is the participant’s own contemporaneous account. It has not been independently verified. The posts we reviewed (archive checked on 17 October 2026) end at the research phase and do not document a product launch or revenue. That does not mean none occurred. The record covers discovery and direction-setting. It cannot establish sustained business operations or zero-FTE performance.

The record

One entry per post, in publication order. Spend and hours are the cumulative figures each post reports.

25 March · I Gave an AI 10,000 EUR and 6 Months

• AI: No AI work reported.
• Human: Set the mandate: €10,000, one machine, one person, six months. Promised to log spend and every hour of human intervention.
• Reported spend / human hours: not yet stated

27 March · Day 1: How I Set Up a Machine

• AI: Executed a test task sent from a phone.
• Human: Built seven setup layers (own network segment, separate accounts, a non-admin user for the AI, encryption, firewall, and the agent setup itself). Anything the AI needs installed, he installs.
• Reported spend / human hours: €666 / 4 h of setup

30 March · Part 2: I Onboarded an AI Like a Co-Founder

• AI: Worked about 15 minutes alone and drafted five questions for the founder.
• Human: Wrote the operating instructions and the project brief, including hard limits: no spending, account creation or sending email without approval. Answered the questions.
• Reported spend / human hours: €666 / 8 h

31 March · Part 3: I Told My AI to Stop Playing It Safe

• AI: Generated and screened 29 business ideas over two sessions.
• Human: Asked for the ideas it was “too cautious to include” and changed the model between sessions. He notes this was not a controlled comparison.
• Reported spend / human hours: €666 / 8 h

1 April · Part 4: My AI Recommended 6 Businesses

• AI: Recommended launching six businesses in three phases. After being overruled, produced a launch plan for the founder’s choice in 25 minutes.
• Human: Overruled the recommendation and picked the option the AI ranked fifth. Upgraded the subscription after hitting capacity limits.
• Reported spend / human hours: €666 spent; recurring costs rising to €120/month / 8 h

2 April · Part 5: Everyone Is Using AI to Work Faster

• AI: Produced an 800-line launch plan in 25 minutes, then ran a layered market research programme.
• Human: Found that the plan had no answer to where the first customers would come from. Discarded it, redesigned the AI’s role from builder to researcher, and set the research rules and escalation triggers.
• Reported spend / human hours: €666 / 12 h

What the record shows, so far

Our reading, not a finding about agentic organisations in general.

• The major direction-setting decisions described in these posts were the founder’s. The founder chose the business against the AI’s ranking, rejected the launch plan, and redefined the AI’s role. This fits the distinction in Why a human? between producing options and judging what they are worth. In Part 5 the founder attributes the plan’s flaws to his own brief rather than to the AI. That is his interpretation. It illustrates the importance of specifying intent; the posts do not measure how effort was redistributed.
• Authority was bounded by tools as well as instructions. The brief forbade sending email without approval. Separately, the email connector could only create drafts. The first “real limitation” the founder reports was in fact an authority boundary built into the tool.
• Escalation and self-critique, as reported. The AI drafted five questions for the founder in its first session, and its research rules told it to stop and email him on captchas. When he asked it to critique its own launch plan, the founder reports that it found most of the flaws he had found.
• Human time is recorded, but coarsely. Reported human hours went from 4 to 8 to 12 across the reviewed posts. The posts do not break those hours down by type (setup, direction, review, rescue), which is the breakdown our method asks for.

Reporting notes and discrepancies

Recorded as published, not corrected:

• Part 4’s address says “8 businesses” and its title says 6. The post explains the difference: eight ideas were evaluated and six recommended.
• The Day 1 post lists the AI subscription at €20 a month. Part 4 says it rose “from 18 EUR”.
• Part 5 announces full research results for a later post that does not appear in the archive.

What is missing

The reviewed posts end on 2 April 2026. The archive we reviewed on 17 October 2026 contains no consolidated retrospective (what happened afterwards, what it cost, where the experiment changed course). Until it is, this record speaks to directing AI during discovery. It cannot establish how the business operated over time. Update, 19 October 2026: Lorant has since published his own account of what happened after April: Built by AI stopped in April. The agents didn’t.