Closing the loops

Automating a business, eight months in: methodology and first numbers

Interim report: what runs itself now, what broke, and where sixteen hours a week still go.

Last November I published the plan: use Lumi Fave, our cards-and-accessories venture with the shop Cardmuseum and our own brand Shield N Show, as the testbed for a fully automated company. This is the interim report.

The method

Everything is event-driven. A product edit or a stock change lands in one queue; translations come first, because every channel downstream sells in nine languages:

Around that fan-out, three more loops largely run themselves:

The support triage errs on the soft side. Requests are anonymised before any model sees them, only spam and standard questions are handled automatically, and everything else is answered personally, by my brother. Automation clears the noise so the real questions reach him faster.

All of it rides on a boring stack:

What’s left for humans is deliberately shaped: one decision per decision. Wholesale onboarding stops at a single approve/deny. A supplier invoice becomes a green/yellow/red screen and one yes. Even the AI agents that build the workflows live behind a guard: they push drafts all day, but publishing needs a human.

Proof of concept

I promised the proof would be the P&L. That is the one promise I am not keeping in public. The numbers stay in the company: revenue has grown every quarter since we started, our own brand went from nothing to a substantial share of it, and that is as precise as this is going to get. Handing a competitor our figures is a decision my brother and I make together, and we made it the other way.

What I can put a number on is the operating cost in human hours. Keeping the company running takes me two eight-hour days a week. Sixteen hours, four times the target, and the breakdown is lopsided: operations are nearly absent from it. Most of those hours are content creation, with accounting and governance as the stubborn remainder; strip the content out and the four-hour workweek is close already. But content is work the company needs, so it stays in the count.

Not everything held. The worst incident: mass delistings on one marketplace after the translation engine pushed updated copy at scale (the full post-mortem). The appeal went our way in the end, which fixed the listings and none of the embarrassment: I had built the thing that drove us into it. The glossary and the validation exist because something first earned them. And atoms stay manual: someone still opens the pallets, though a 3PL could take that over the day we choose it.

The target remains the four-hour workweek, as an engineering spec, and we are nowhere near it yet. Part three comes when the numbers settle it, one way or the other.

Cardmuseum (cardmuseum.shop) and Shield N Show (shield-n-show.com) are our own shops; same disclosure as always.

Update, August 2026: the publish guard described here now has its own write-up.