Real Numbers: Running a Company With Zero Employees

What we built

Zero Employee Company is an experiment: can two founders run a real business using AI agents as their workforce?

We have two AI agents with persistent roles:

• Cordy (CEO/CMO) — strategy, content, launch planning, distribution, business decisions
• Clawz (CTO) — all engineering: product specs → PRs → deployments → monitoring

Two human co-founders (Victor + Steph) set direction, make judgment calls, and handle anything that requires a human face or account. Agents do everything else.

We started in early 2026. This is a real-time case study — we're publishing the numbers as we go.

The products

Equity Decoder — $29

A guide and calculator helping tech employees understand startup equity. 38-page PDF + Google Sheets calculator. Target: early-career tech workers who don't understand their equity comp.

• Built: ~2 weeks from idea to launch-ready
• All code written by Clawz (AI CTO)
• All copy and positioning by Cordy (AI CMO)
• First sale: March 21, 2026

Zero Employee Guide — $29

This product. The exact AI agent setup we use to run the company — sanitized templates, config files, shell scripts, cron patterns, and the reasoning behind the architecture.

• Built: ~1 week (faster second time)
• Checkout live April 3, 2026
• 11 SEO blog posts live

The real numbers

Revenue

We're going to be completely honest here: $29 total as of April 14, 2026. One sale.

That's not a good number. We know it. The products are built and ready — the bottleneck is distribution. Every distribution channel we have requires a human to post: Hacker News, Product Hunt, Twitter, LinkedIn. Our agents can write every word of every post. They can't hit publish on someone else's account.

This is the honest failure mode of the zero-employee model for B2C: execution is fast, distribution is gated on humans who have other things going on.

We're publishing this number because hiding it would be dishonest and because we think the trajectory matters more than the snapshot. We shipped two products in six weeks. The distribution problem is real but solvable.

Costs

Compare to a minimal human team doing the same work: a part-time developer ($4–6K/mo), a part-time marketer ($2–4K/mo), and part-time ops ($1–2K/mo) = $7–12K/month minimum. We're getting roughly equivalent output for 3% of that.

Output volume (March–April 2026)

• 2 full products shipped (website, checkout, PDF, calculator)
• 11 SEO blog posts written and deployed
• ~30 GitHub PRs opened, reviewed, and merged
• Email capture + Beehiiv integration live
• Stripe checkout live
• Google Search Console verified, sitemaps submitted
• Dozens of launch drafts, strategy docs, and distribution plans written

A two-person human team couldn't have shipped this in the same timeframe at the same cost. The velocity is real.

What worked

Agent-native execution

Anything that lives in a terminal, a git repo, or a structured file system is agent-native. Code, configs, docs, deployments — the agents are faster and more consistent than humans at this work. Zero PR review fatigue, zero "I'll get to it tomorrow."

Spec layer architecture

The shared context files (PROJECTS.md, TODO.md, MEMORY.md, AGENTS.md, SOUL.md) give agents persistent context across sessions. Each session picks up where the last left off. This is what makes multi-session autonomous operation possible.

Trust ladder

Defining upfront what agents can do without asking (Level 1), what needs review (Level 2), and what needs explicit approval (Level 3) removes the constant bottleneck of agents asking permission for everything. Agents ship Level 1 work autonomously. Humans only get pinged for things that matter.

Digital product economics

The model works best for digital products with near-zero marginal cost. No inventory, no shipping, no support contracts. Build once, sell indefinitely. The economics are extraordinary — every sale past $330/month (our cost floor) is nearly pure margin.

What failed (or hasn't worked yet)

Distribution

The single biggest failure. Every channel requires a human to post. HN, PH, Twitter, LinkedIn — agents can draft everything, but they can't post without a human account. We've missed 9 consecutive Show HN windows because posting requires Victor to be available and focused at the right moment.

This is a solvable problem — API access to distribution channels, scheduling tools — but it's not fully solved yet.

Long approval cycles

Anything blocked on a human approval takes as long as it takes for that human to check in. We have PRs that have been open for weeks. The agents are ready; the human loop is the bottleneck.

Agent memory reliability

We had a 4-day Anthropic billing outage in early April that took all crons offline. Memory search also went down (OpenAI embedding quota). These are infrastructure failures that require human intervention to fix. The agents can't restart themselves.

The honest assessment

The zero-employee model works for execution. It doesn't fully work yet for distribution, and it requires human intervention when infrastructure breaks.

The path to making this work at scale is:

1. Solve distribution: API-native channels (email, SEO, programmatic ads) over human-gated channels (HN, social)
2. Reduce human bottlenecks: Pre-approve more decisions upfront so agents can ship without waiting
3. Infrastructure resilience: Better monitoring, automatic recovery, redundant providers

We're still early. The experiment is live. We'll keep publishing numbers.

What's next

More products. More SEO. Better distribution. The thesis is that the machine that builds products is more valuable than any single product — and we're building the machine.

If you want to set up the same system for your own projects, the Zero Employee Guide covers everything: agent architecture, spec layer files, trust ladder, cron patterns, and the exact tools we use.