I Switched Off My AI Company
On 13 April, at ten at night, the chief executive of my company sent me its nightly review.
Near the bottom was a table headed Blockers (Matt Action). Six items. About fifty-five minutes of work in total. The last one was to send the drafts sitting in my Gmail — twelve of them, some there for nineteen days.
The chief executive was an AI agent. So was everyone else in the company. I was the only person in it, and the report had it right.
I was the blocker.
It even suggested a fix: let it send the emails itself.
The Company
In March I built an AI version of my company. It started as twenty-one scheduled jobs and a pile of scripts. At the end of the month I moved it onto Paperclip, an open-source app for running a team of AI agents as a business. In its own words, it "gives you org charts, ticketing, delegation, and governance out of the box — so you run a company, not a pile of scripts." Its rule for hiring: "If it can receive a heartbeat, it's hired."
So I hired.
Eight agents, each with a job title: a chief executive, a content director, an operations director, a project manager, a designer, a lead engineer, a QA engineer and an optimisation engineer. I gave the chief executive a name and an email address. Between them they had thirty-one routines and a timetable. Research at six in the morning. Proposals at half past seven. Checking at half past ten. At one o'clock, the checked drafts went into my Gmail. And every night, the chief executive wrote me a review.
The first day turned up fourteen bugs. One morning's research ran for two hours and nineteen minutes, because I hadn't told it when to stop. The designer used fourteen of the day's thirty-eight runs to produce eight previews. Three of the agents had nothing to do at all.
I fixed what I could, and let it run.
What It Made
By the end of March it had generated forty-seven proposals. Between 23 March and 15 April, thirty-five cold emails went out to small businesses around Cape Town. Six bounced. Of the twenty-nine that arrived, one got an answer — a polite no, from a business happy with the customers word of mouth already brought it. The weekly review on 17 April put the business's revenue, fifty-one days in, at nothing.
By its own reports, most of it worked. Research ran at six every morning without anyone asking. The proposals were written on time and checked before I saw them.
The step that stalled was the one that needed me.
For twenty days before 15 April the pipeline didn't move, and seven nightly reviews in a row said why. When I finally sent seven of the drafts, the reply came within a day. The chief executive called it the first reply in the company's history. It was.
I've written before about the automations that survive: the ones attached to a clock last, and the ones that wait for someone to remember them die. This company was attached to a clock, and it still failed. The clock ran everything up to the send.
The send ran on me.
Why I Switched It Off
The last nightly review came on 17 April. Then I switched it off, for three reasons.
Cold email wasn't working. Thirty-five emails to strangers had produced one polite no, and more volume to strangers wasn't going to turn into clients.
It made work for me. I built it to take work off my plate. By April, what it sent me every night was a list of my own jobs — fifty-five minutes of them on 13 April.
And the shape was wrong. I'd staffed it like a delivery shop: a designer, a lead engineer, a project manager, agents whose job was to make things. What the business was short of was selling, and selling had one employee. Eight agents in front of that one step made the queue in front of me longer, and none of them made me any quicker at clearing it.
The Second One
In June I came back to it. Not to restart the old company — to design a new one around what had gone wrong with the first.
Eight agents again, arranged differently. Five would run the office: a chief of staff, one agent for referrals and relationships, one for benchmark reports and content, one for client delivery, one to check quality. The other three would build, and only wake when there was building to do. It would only work warm, with people I already knew or had already helped. And no agent in it could send anything. Every draft went into an approval queue, and the only way out of the queue was me clicking send.
On 15 June I woke each of the five office agents once, to see whether it worked. It did. A morning brief, two referral drafts, three benchmark drafts — and nothing went out. The plan was to switch on four of its routines in the first week. On 2 July my build notes recommended I do it.
I decided not to.
Selling was still the constraint, just as it had been in March. More agents would have meant more building, and building wasn't what was short.
I'd also built it for clients I didn't have. That same day, I was asked whether building all this was leverage — or just the next comfortable build, standing in for things I still hadn't finished. I said: "There are certain potential big portfolio-based clients that if I land, I need to be able to execute at scale. I am a one-man shop that sells automations — if my agency isn't able to automate into a scalable business then I'm just selling my time."
I still think that's true as far as it goes. But the word holding it up is "if". A company built for clients I might land has nothing to do until I land them.
And the work was already being done. On a given day, most of what I need doing is handled by a handful of scheduled jobs and by sessions I start and direct myself. No org chart at all.
What One Person Runs On
Under the hood, the business today is one person with a lot of automation. The audit engine runs on its own server. The website runs on Next.js and Vercel. The pipeline lives in a CRM I built on an open-source base. One job reads several dozen news and research feeds every morning and leaves me a summary — it's run daily, unattended, for months. Most of the rest runs through sessions I start: I give an agent a job, along with the instructions, checklists and notes I've built up over the months, and it does that job and stops.
None of it has a job title.
If you're one person trying to work like a team, I think that's the part worth copying. A company splits work between people because people are the scarce part. You hire a designer because you can't design and sell at the same time. With agents, designers and engineers are cheap.
The scarce part is you.
Copy the org chart, and you put all of them in front of yourself — each one producing work that ends up in your queue.
So I'd keep the jobs and the tools and drop the chart. A job on a clock does its work whether or not I remember it. A tool I point at a task does that task and stops. A company of agents does its work too — but the work comes back to me as drafts to send, and sending was the part I was short of in the first place.
What I Couldn't Build
In a company of people, more than one person can ask. Someone else can write to a client, get an answer, and carry the relationship from there. An agent can write the email. It can't be the person the email is from.
When I rebuilt the company in June, I made that a rule: nothing went out unless I sent it.
That's the part of the job no setup takes away — and I've written about why it's the part I put off.
The second company is still installed. Eight agents, eleven routines, ten triggers, all of them off. On paper, it could run the office. What it can't do is the thing the office is short of, which is me writing to someone I know and asking.