Case Studies
Building from a van while the factory ships
What AI operations looks like when it is how the work actually gets done, including the day it stopped nine times.
If you run a small business or a consultancy, you might suspect AI could carry more of your work, and still be wondering: what would that actually look like for me? I can only answer from my own record, so that's what follows. The bits that went wrong are left in, because they're the useful bits.
I'm building a business from a van while on parental leave. Yes, really. Most of the building is done by a set of AI agents I call the factory. It picks up tickets, builds, checks its own work and hands it back for me to review. I make the decisions; it does the carrying out. Here's a dated account of how that's gone, good days and bad.
14 to 15 September: a night nobody was watching
On the night of 14 to 15 September the factory ran unattended and landed four tickets. I woke up to four pieces of finished work waiting to be looked at.
That's the bit worth noticing if you run a business. The gain wasn't speed in the abstract. It was that work I'd already described clearly carried on without anyone at a desk.
8 October: the good half
On 8 October the factory went from idle to five slots working at once. It landed four pieces of work: tickets #348, #336, #328 and #318.
One of them, #318, was an accuracy, consent and About pass on this website, the one you're reading. Two others were improvements to the factory itself: #348 let the stage that reads a new ticket run on any model, and #336 let the cloud lane take on product work once its plan had been checked.
That evening the model provider ran out of credit part-way through. The factory switched provider and carried on.
8 October: the other half
The same day, the factory stopped nine times on one recurring fault. An orchestrator (one of the agents that supervises the factory) had to repin fifteen tickets by hand to get them moving again.
I'm telling you this because it's the real shape of the thing. An agent factory doesn't simply run. It runs, fails in a repeating way, and needs somebody to notice the pattern and fix the cause. Nine stops on one fault is a fault to fix, not a success to dress up. By the next morning the factory had committed a fix, ticket #354, for one repeating stoppage it had hit that day: a ticket's working session fell out of step, and the ticket sat stuck until somebody cleared it by hand.
If someone shows you AI doing business work and there's no account of where it broke, ask them what they left out.
What would this actually look like for me?
That's your question, not mine, so here's my honest answer. You don't need a factory to take something from this. Three habits carry over from my setup to any small business:
- Write the work down so it can be handed over. A ticket for the factory starts with a plain-English description of what's wanted and how to tell it's done. If you can brief a new hire on a task in a paragraph, an agent can make a start on it.
- Keep the decisions with a person. The factory doesn't decide what the business does. I answer the questions and review the results.
- Plan for the stoppages. Something will fail, and a provider running out of credit or a repeating fault will happen at an inconvenient hour. Decide in advance who looks, how often, and what they can safely reset.
The next test
From around 13 or 14 November I'll be in Costa Rica for four to five weeks. Same setup, run from further away and across a time difference (and, I hope, with a few surf breaks). It'll be judged the same way: on what lands, and on what stops.
If you'd like to talk through what this could look like for your own work, get in touch. It's a conversation, not a pitch.