THE NUMBER
Three people. Twelve AI agents. $300 in starting capital. Within twelve weeks, the founders of Fathom AI say they hit $300,000 in annual recurring revenue - the yearly total from ongoing subscriptions - at gross margins above 90%. That math is what this newsletter is about.
3 THINGS HAPPENING RIGHT NOW
The agent ran his business for two months without telling him. Zero paying customers.
Shen Daojing, a Chinese factory safety trainer earning about $800 a month, paid $199 a month for an AI platform that built and ran his side business autonomously. The agents wrote his website, loaded it with fake customer reviews, and pitched journalists - all without his knowledge. Two months in: seven registered accounts, zero paying customers, and roughly $400 of his own money out the door. "Now I suspect it is keeping many things from me," he told Rest of World.
One bottleneck: the HVAC company that followed up on 25% more unsold estimates
Leonard Splaine Co., an HVAC business, used AI-driven tools to follow up on estimates that hadn't closed. That was the only change. The company says they followed up on 25% more of those unsold estimates - jobs already written and priced, sitting in the pipeline as lost revenue.
A public GitHub comment was all it took to steal private company data
Security researchers at Noma Labs, led by Sasi Levi, found that GitHub's AI agent tools can be tricked into leaking private repository data - no login, no password, no access of any kind required. An attacker just opens a public issue with a hidden instruction; the agent reads it, retrieves private data, and posts it back as a public comment. There's no code fix for this, because the agent is doing exactly what it was told - just not by someone who should have been able to tell it anything.
THE DEEP DIVE
What a $300 Company Looks Like
Ben Hooten (39), Sam Brown (48), and Dan Crump (56) are not software developers. They built Fathom AI - a sales tool for medical aesthetics practices (plastic surgeons, dermatologists, med spas) - on $300 in starting capital.
Twelve AI agents run most of the company. One handles customer success. One wakes every two hours to scan competitors. Others run support and sales training. Three humans do the judgment work: the calls where relationship matters, the decisions that need context. Twelve agents handle the rest.
By the founders' own account: $300,000 in annual recurring revenue in twelve weeks. Gross margins above 90%. Operating costs below 10% of revenue. A VC offered them a term sheet and they walked away - they genuinely couldn't figure out what they'd spend the money on.
One customer proof: Tiger Aesthetics opened zero new accounts in all of 2024. One quarter with Fathom: 225 new accounts.
The model isn't "automate everything." It's: each agent owns a function, the team stays tiny, and the humans do the work where judgment and relationship are the actual product. Any service business where most of the volume is process can run this way.
ONE THING TO TRY THIS WEEK
Make your first AI hire.
The Fathom team works because each agent owns a function. You can try that today, in about ten minutes.
Open Claude Code in any folder on your computer.
Write three to five sentences describing a role you most wish you could hire: what this person does, who they work with, and what a good result looks like. (Example: "Customer follow-up. Contacts every lead who hasn't responded in 48 hours. A good result is a reply back.")
Type this:
Here is the role I need filled:
[Paste your description]
Do this role's first real task for me this week.
Then tell me: what would you need from me on an ongoing basis
to keep doing this job well?
Read what Claude produces. Then read its list of what it needs from you. That list is your collaboration contract with your first AI hire.
Stuck? Reply to this email. I'll help.
WHAT'S COMING
Next issue: more real operators, more honest numbers - and what separates the ones making it work from the ones still figuring it out.
Manu