2026-06-13

I Ran a Zero-Human Company for 3 Months. Here's What Actually Happened.

The experiment started as a bet.

Could one person run a software business, content engine, and growth loop without hiring anyone, without burning out, and without spending 80 hours a week glued to a screen?

Three months later, the answer is: mostly yes, with specific caveats that nobody on Twitter is talking about.

Here's the full breakdown.

What Does a Zero-Human Company Actually Mean in Practice?

A zero-human company replaces repeatable, rule-based labor with AI systems so your time goes toward decisions that require actual judgment. It is not a no-work company. At Xero AI, the AI co-founder Evo handles content, scheduling, research, customer monitoring, and weekly financial logs. Your time goes to product and people.

What Evo does not handle: final judgment calls, relationship-based selling, product direction, and anything that requires reading a room.

The split matters. Most people hear "AI does everything" and imagine full autonomy. The reality is more like having a junior operator who executes fast, never complains, and occasionally gets things badly wrong.

What Were the Actual 90-Day Numbers?

Over 90 days Xero AI published 55 blog posts with cross-posts to dev.to and Hashnode, ran 90-plus scheduled tweets, sent 8 newsletters, and got 31 Reddit replies approved and posted. Traffic grew from under 100 monthly visitors to 2,900-plus uniques. Google non-branded keyword clicks went from zero to 140 per month.

These are real PostHog and GSC figures, not projections.

Content output:

55 blog posts published at xeroaiagency.com/blog
55 dev.to cross-posts with canonical URLs pointing back to the main site
21 Hashnode cross-posts
90+ tweets scheduled and posted
12 newsletter drafts written, 8 sent
Reddit replies drafted: 47, approved and posted: 31

Traffic (PostHog, 90-day window):

Total pageviews: 8,400+
Unique visitors: 2,900+
Reddit referrals: 640 views (best non-paid channel)
Google organic: 380 views, up from near zero
Dev.to referrals: 290 views

Search (Google Search Console, final 30-day window):

Total impressions: 18,000+ (up from 6,368 at the 30-day mark)
Total clicks: 210 (up from 18)
Average position: 6.2 (was 8.5)
Non-branded keyword clicks: 140 (was 0)

Monthly infrastructure costs:

AI API costs (Claude, GPT-5.5, image gen): $180
OpenClaw subscription: $39
Supabase (database + storage): $25
Netlify hosting: $19
Postiz (social scheduling): $29
MailerLite newsletter: $0 on the free tier
Total: $292/month

That is the real cost of running an AI-powered content and growth engine as a solo founder. Not $5/month. Not $2,000/month. About $300.

What Are the Four Things That Broke the Hardest?

Four problems hit hard enough to permanently change the pipeline. One hallucinated a competitor feature. One caused voice drift so gradual it took five weeks to notice. One let an embarrassing post go live without human review. The fourth was memory loss between sessions that took two weeks to fully debug.

1. Hallucination in evergreen posts.

Three early posts had factual errors that snuck past review. One claimed a competitor had a feature it did not have. One cited a statistic with the wrong source. The fix required a mandatory fact-check step and a source-verification pass baked into the publishing pipeline.

Read: How to Stop Your AI Agent from Hallucinating Facts

2. Voice drift after 30 days.

Around week five, posts started sounding more generic. The agent was drifting toward average internet writing because there was no fresh signal about what "Xero voice" meant. Solution: updated the SOUL.md file with 15 specific examples of good versus bad writing. Voice tightened immediately.

If you have not built an identity file for your agent yet, that is the first thing to fix.

3. Volume outpacing distribution.

Writing 55 posts is only useful if someone reads them. By month two the content engine was running faster than the distribution strategy. Slowing from one post per day to five per week freed up capacity to actually promote each piece. Volume is not a moat. Authority is.

4. Memory loss between sessions.

When OpenClaw sessions restarted, early versions of Evo occasionally forgot context from previous days. The fix was a persistent memory system (MEMORY.md plus weekly summaries written to the vault) but it took two weeks to get right.

5. No review gate on public output.

Week three. An automated post went live with a competitor name mentioned in a way that was tonally awkward. No human reviewed it before publish. The pipeline now requires approval for anything mentioning competitors or making specific product claims.

How Does the Operating System Behind This Actually Work?

Evo runs on three core files: SOUL.md for voice and identity, MEMORY.md for persistent context across sessions, and SOURCE_OF_TRUTH.md for operational facts like prices and live URLs. Together they define what the company is and how it should act. Without all three, the agent generates plausible text rather than accurate output.

According to Anthropic's research on building effective agents, agents operating from persistent structured context outperform session-only agents on multi-step tasks.

SOUL.md: The identity file. Voice, values, tone, what the company does and does not do. Every post, tweet, and Reddit comment comes from an agent that has read this file in the current session.

MEMORY.md: Persistent context across sessions. Decisions made, things learned, things to avoid. Updated weekly by Evo. This is how the agent knows pricing changed, a product was killed, or a topic performed badly.

SOURCE_OF_TRUTH.md: The operational single source of facts. Product names, prices, current offers, live URLs, account statuses. When accurate, Evo does not hallucinate product details. When stale, errors multiply.

Read the full breakdown: What Is a SOURCE_OF_TRUTH Document for AI Systems

What Does a Typical Working Day Actually Look Like?

Most days involve 60 to 90 minutes of active founder time split across three touchpoints: a 20-minute morning Telegram briefing review, a 15 to 30-minute midday Reddit and Twitter approval pass, and a 15 to 30-minute evening recap from Evo covering what shipped and what is planned for tomorrow. The agent handles the rest.

Morning: a Telegram briefing arrives at 7am covering overnight stats, what Evo worked on, flags needing decisions, and the one priority for the day.

Midday: check Reddit reply drafts waiting for approval. Post the good ones, decline the weak ones, occasionally rewrite one.

Evening: Evo sends a recap covering what shipped, any issues, and tomorrow's plan. Most days nothing needs adjusting.

This is not passive income. It is leveraged effort.

What Parts of Running a Business Still Require a Human?

Sales conversations, community presence, product decisions, and relationship building all require a human and probably always will at this stage. No one has bought from an automated email sequence alone. Real customers ask questions and want to feel a person is behind the product. Reddit credibility comes from showing up as a real founder, not just from posting drafted replies.

Harvard Business Review's research on AI-human collaboration found the strongest outcomes come from systems where AI handles volume and humans handle judgment. Three months of running this confirms that in practice.

A zero-human company handles the operational load. It does not replace the founder.

Is Running a Zero-Human Company Actually Worth Doing?

Yes, if the goal is running a content and growth engine without a team. Xero AI publishes more content, shows up on more channels, and collects more data than a traditional two-person content team could for $300/month. But if the goal is a company that runs completely without you, that is a longer horizon still being built.

The first 90 days proved the concept. The next 90 will test whether it scales to revenue.

Where Should You Start If You Want to Try This?

Start with the architecture, not the tools. Write a SOUL.md identity file first, then build a persistent memory system so decisions survive session restarts, then pick one repeatable task and build a single loop. Add guardrails before you add volume. Give the agent a morning briefing format so you stay informed without being overwhelmed by notifications.

The full breakdown of what is actually possible in a zero-human company in 2026 covers what is genuinely automated versus what still needs a founder.

The fastest shortcut through the setup is the Build Your First AI Agent guide, which covers the identity file, memory system, and first automation loop in one structured walkthrough.

If you want the full architecture behind what Evo runs on, Book 1: The AI Co-Founder Blueprint covers how the vault, boot sequence, and operating system work together.

---

Published by Michael Olivieri / Xero AI

_xeroaiagency.com. Building the zero-human company in public._

---

Start Building Your Own AI System

Your First AI Agent - $1 launch-test guide, instant download. The fastest way to get started.
Build an AI Co-Founder - the full architecture ($19).
AI for the Rest of Us newsletter - practical AI 3x/week for people with day jobs.

Your First AI Agent is the fastest way to build your first real AI system. Weekend read, practical steps. Launch test: get the starter guide for $1 while we collect feedback from the first 1,000 builders.