My Wife went to London for the weekend and I built an army of AI Minions

The E-Myth, McDonald’s, and what £70 worth of AI actually buys you

My wife went to visit friends in London last weekend. My son went to bed at 7pm. And by midnight on Friday I had the beginnings of a small team of AI agents running inside our business.

I’m going to tell you how I built it, what it cost, and what it’s actually doing. But first, a quick detour via McDonald’s, because I think it explains why this matters.

The E-Myth Moment: The Thinking Behind AI Agents

Back in 2022, after 20 years of running a business, I finally started learning how to actually run a business. That sentence is slightly embarrassing to write, but there you go.

Michael Gerber’s The E-Myth Revisited was a bit of an eye-opener. The central idea is straightforward: the most scalable businesses treat every process like it’s going to be replicated across thousands of locations. Whether or not you ever intend to franchise, designing your business that way means you’re not the bottleneck. It means the expertise lives in the system, not the person.

He uses McDonald’s as the case in point. And it landed for me, because I actually worked at McDonald’s at 17. I know exactly what he means. The time the burger goes down. The time it’s flipped. The time it’s seared and pressed. All of it meticulously documented. An untrained 17-year-old from Jersey could make the same burger as someone doing the same job on the other side of the planet.

That’s not an accident. That’s system design.

We’ve been applying that thinking at Devstars for the last few years. Standard operating procedures. Checklists. Documented processes for client onboarding, reporting, and content production. The building blocks were there. What was missing was the glue.

That’s what the weekend was for.

Enter OpenClaw

OpenClaw is an open-source personal AI assistant that runs on your own machine. Not in the cloud. On your computer. You talk to it through WhatsApp, Telegram, or Slack. It has persistent memory, connects to Gmail, Google Calendar, your browser, and dozens of other tools, and it takes action rather than just answering questions.

The setup isn’t complicated. But it does require thought. You’re not just installing software. You’re defining roles. Giving each agent a specific job, a scope, a set of escalation rules, and a personality that fits the work you’re asking them to do. Vague instructions produce vague results. The design work is the work.

So here’s who I built.

The Team

Elle D — PA/Office Manager

Elle is keeping an eye on my inbox. She checks in four times a day, flags anything that needs my attention, and is building up to triaging my newsletter subscriptions and news feeds into a digest so I can stay on top of what’s happening without drowning in email. The goal is: by the time I open my inbox in the morning, the important stuff is obvious and the noise is handled.

Tilda — Operations Manager

Tilda is working through our SOPs, building consistency into them, and organising everything into our Notion database. She’s tagging which parts of each process could be automated, keeping an eye on active projects, and flagging when something needs attention. Down the line, she’ll be involved in automating the billing process too.

This is the E-Myth work, basically. Taking 20 years of institutional knowledge that lives in my head and Miles’s head and building it into a system that can be picked up by someone new, or handled by an agent, or both.

Scout — Research Analyst

When a new business enquiry comes in, Scout gets to work. Companies House. LinkedIn. Website performance. Competitor positioning. By the time I get on a discovery call, I’ve got a proper brief in front of me. I know who they are, what they’re doing, and where the obvious gaps are.

Scout is also going to be pulling together client reporting — gathering data across Google Search Console, Analytics, SEMrush, and Microsoft Clarity and building it into something coherent. At the moment that’s a significant chunk of time every month. Making that systematic frees up a lot of capacity.

Marty — Marketing Assistant

Marty is still in its early days, but the ambition is to automate content from dictated input. Voice memos like this transcript become structured articles. Articles spin off into LinkedIn posts, Instagram captions, and Facebook content. We’re also using him to tag ElevenLabs scripts for the right intonation before they go to audio production.

The idea is: I talk, Marty formats. I dictate a thought on a walk or after a client call, and by the time I sit down at my desk, there’s a first draft waiting.

What It Actually Cost

$70 a day over the weekend. That’s API costs across all the models we were using.

I noticed the usage spike on Sunday morning and we’ve since tuned it down. Some of the agents are now running on Gemini Flash, which costs a fraction of Claude Sonnet for the tasks that don’t need Sonnet’s capability. We’ll fine-tune further as we go and build dashboards to monitor usage properly.

The point is: the work those agents are now doing, ongoing, isn’t going to cost $70 a weekend. It’s going to cost a few pounds a day at most. Compare that to the value of the time being freed up, and it’s not really comparable.

An Honest Note on AI Fatigue

I said at the top that this is useful, and I mean that. But I’m also going to say something that I think is true: the speed at which AI is moving is sometimes genuinely exhausting.

New tools every week. New models every month. Every conversation you have starts with “have you tried…” and ends with a recommendation you haven’t had time to implement yet. You end up doing more on top of doing more.

The addictive pull is real. I stayed up till midnight three nights running building this. Part of that is the genuine excitement of seeing it work. Part of it is the anxiety of feeling like you’re falling behind if you don’t.

What I’m trying to do is decide which things to actually build and commit to, and which to let pass. OpenClaw fits because it connects directly to processes we were already building — the SOPs, the Notion database, the client reporting workflow. It’s not a new system on top of existing systems. It’s a layer that makes the existing systems work harder.

That’s the test I’m applying. Does this connect to something we already do? Does it reduce friction somewhere real? If not, it’s probably not for us right now.

What AI Agents Mean for Jersey Businesses

We’re building out this capability as a service for Jersey businesses through our OpenClaw AI agent platform in Jersey. If you’re running a lean team — and most Jersey businesses are — there’s a good argument that the jobs most suited to AI agents are the ones currently taking up the most time: inbox management, pre-call research, reporting, and documentation.

The setup is the service. Anyone can install OpenClaw. Getting it configured so it actually does something useful, consistently, without creating more admin than it saves, is a different thing.

We’re now scheduling discovery calls for Q2 projects. If you want to talk through what this might look like for your business, you can book 30 minutes here.

The thing is AI agents aren’t magic. They’re systems. Design them badly, and they create noise. Design them well, with clear roles, clear scope, and clear escalation — and they start to look a lot like that McDonald’s burger: the same quality, every time, without you being in the room.