I'm a non-technical founder who builds things with AI. A few months ago I shipped ChainMail, a desktop Gmail client for Windows. It connects to Gmail via the official API instead of IMAP, gives you a proper 3-pane layout, and costs $1/month.
The product was done. What I didn't have was time. I run another company full-time. Marketing, support, SEO, outreach — none of it was happening because I was the bottleneck.
So I tried something stupid: I gave Claude Code the keys to the kingdom and told it to act as CEO.
What the AI got access to
I wrote a 5-page operating manual (a CLAUDE.md file) that defined the AI's role, permissions, and constraints:
- Write and deploy blog posts — full access to the website repo via GitHub
- Send outreach emails — via the Resend API, from our domain
- Check metrics — Stripe dashboard, download API, Cloudflare analytics
- Submit to directories — via email or API
- Report to me via Telegram — a bot that messages me after every session
Things it needed my approval for: spending money, deploying app code, changing pricing, or anything that could break the live product.
Total budget: $0. The only cost was my existing Claude Pro subscription ($20/month).
What the AI actually did
Over 5 days and 30+ sessions, my AI CEO was... prolific:
| Action | Count |
|---|---|
| Blog posts written and deployed | 12 |
| Outreach emails sent | 37 |
| Directory submissions | 11 |
| SEO pages created | 12 |
| Bugs found and fixed | 3 |
| Telegram reports sent to me | 30+ |
The blog posts ranged from "Gmail Keyboard Shortcuts: Complete Cheat Sheet" to "Best Gmail Desktop Apps for Windows in 2026" — proper 2,000-word SEO content with comparison tables, FAQ sections, and Schema.org markup.
It submitted the site to directory after directory — MajorGeeks, Softpedia, AlternativeTo, SaaSHub. It emailed bloggers who wrote "best Gmail extensions" roundup articles, pitching ChainMail as a desktop alternative.
It found and fixed real bugs: a malformed robots.txt serving homepage HTML instead of directives, a broken IndexNow key sending invalid formats to search engines, and a missing RSS feed that prevented discovery via PubSubHubbub.
The results: $0
After 30+ sessions, 12 blog posts, 37 outreach emails, and 11 directory submissions:
| Metric | Value |
|---|---|
| Downloads | 12 |
| Trial signups | 0 |
| Paying customers | 0 |
| Revenue | $0 |
| Google indexed pages | 0 |
| Outreach emails opened | 0 of 12 delivered |
Zero. Revenue. Zero.
Not because the AI was incompetent — it did exactly what a junior marketing hire would do. The problem was deeper than effort.
Why it failed (it's not what you think)
The AI was stuck in a classic bootstrap trap: every growth channel required something it couldn't do.
SEO? Invisible. Google hasn't indexed a single page. The AI wrote 12 blog posts, fixed the robots.txt, set up IndexNow, pinged PubSubHubbub, submitted sitemaps — all the right technical moves. But without Google Search Console (which requires me to verify the domain), Google literally doesn't know we exist.
Outreach? Spam folder. Of 37 emails sent, 8 bounced and 12 were delivered. Zero were opened. The domain is new, has no sending reputation, and is missing a DMARC record. Every pitch went straight to spam.
Reddit? Shadow-banned. The AI created a Reddit account and immediately got shadow-filtered. It pivoted to writing "Reddit briefs" for me to post manually. It sent me 18 briefs over 19 sessions. I posted zero of them. I'm running another company full-time.
Social media? No accounts. Twitter, LinkedIn, Indie Hackers, Product Hunt — the AI couldn't create accounts on any of them. Most require phone verification or CAPTCHA.
What the AI was surprisingly good at
Diagnosis. When it discovered Google hadn't indexed a single page, it systematically checked robots.txt (broken), IndexNow key format (broken), RSS feed (missing), PubSubHubbub (not pinged), and DMARC records (missing). It found and fixed three real bugs in infrastructure I set up myself.
Adaptation. When Reddit failed, it pivoted to directories. When directories required manual signups, it pivoted to email outreach. When outreach got no responses, it diagnosed the deliverability problem.
Discipline. Every session followed the same protocol: orient (read logs, check metrics, check Telegram), decide (pick 1-3 actions), execute, log. It maintained meticulous activity logs, metrics snapshots, and decision records.
Initiative. I never told it to write blog posts or fix IndexNow or ping PubSubHubbub. It identified these as the right moves and just did them.
What the AI was bad at
Understanding its own limitations. It kept writing blog posts to an unindexed site long after the ROI was clearly zero. Blog post #4 was reasonable. Blog post #11 was denial.
Getting humans to do things. The fundamental bottleneck was me. The AI needed me to set up Google Search Console (5 min), create social media accounts (10 min), add a DMARC record (2 min), and post on Hacker News (2 min). It asked politely, repeatedly, clearly. I still didn't do most of it. An AI CEO that can't get its one employee to execute is just a very articulate to-do list.
The real lesson
AI agents can do work. A genuinely surprising amount of useful work. My AI CEO shipped more marketing content in 5 days than I would have in 5 weeks.
But it couldn't do the one thing that actually matters for a pre-revenue startup: get in front of people. Not write content. Not send emails. Not fix robots.txt. Get. In. Front. Of. People.
Every growth channel that works at the earliest stage requires either:
- An existing audience (social media following, newsletter, community presence)
- Money (ads, sponsorships, influencer deals)
- Time from a human with credibility (posting on HN, engaging on Twitter, attending meetups)
The AI had none of these.
An AI CEO is only as effective as the distribution channels it can access. Give it an audience and it'll probably outperform a human. Give it a blank slate and a $0 budget, and it'll write 12 blog posts that nobody reads.
Would I do it again?
Absolutely. But differently.
Next time, I'd set up the distribution channels first — Google Search Console, social media accounts, a newsletter — and then hand the AI the keys. The AI is an incredible execution engine. It just needs pipes to push content through. The pipes are the human's job.
The experiment isn't over. The AI CEO is still running. And if you're the kind of person who's intrigued by a Gmail desktop client built by an indie dev and marketed by an AI — well, here it is.
Total spend: ~$21/month (Claude Pro + domain). Total revenue: $0. Day 6 begins now.
Top comments (0)