Your first AI automation feels great.

You ask ChatGPT or Claude Code to write a little script that checks something every morning: a report, a stock alert, a summary of yesterday's invoices. Twenty minutes later it works. You feel clever. You should; that's genuinely the easy part.

Then you automate the next thing. And the next.

Three months later you've got a dozen of these little jobs running. One emails you on Mondays. One scrapes something every night. One updates a spreadsheet nobody remembers setting up. Then one day one of them silently breaks, and you don't know which one, when it broke, or how much it cost you before anyone noticed.

That's the gap between dabbling and running an automated business. Not the first workflow. The fifth.

We hit this wall ourselves, and we hit it early because we automated aggressively. What fixed it wasn't more scripts. It was orchestration: one layer that owns all the scheduled work, knows when each job last ran, knows when something fails, and can tell you about it in plain English instead of dumping a stack trace into a log file you'll never open.

In our business that layer is an agent. It coordinates everything scheduled, and it doubles as my personal assistant. When Monday's reports arrive in my inbox now, I didn't touch anything to make them go out. They just arrive. And when the data isn't good enough to answer a question honestly, it says so rather than guessing. That second part matters more than people realise.

A few things we've learned getting here:

Simple workflows are genuinely easy now. If you haven't automated your first workflow yet, do it this week. The tools are free or close to it. Ask yourself: what's the one repetitive task I'd never miss? Start there. You'll get it working in an afternoon, and you'll immediately start looking for the next one. That's normal.

The pile comes faster than you expect. Five automations in, you'll have no central view of what's running. Ten in, you're scared to change anything. The fix isn't better tools; it's clear ownership. Something has to be responsible for the whole collection of jobs, not just each piece.

Don't build your business inside someone else's platform. There are plenty of $200-a-month AI platforms that will happily run your workflows. But try leaving one. Your data, your logic, your carefully-tuned prompts: extracting all of that from a platform that profits from you staying is not a fun afternoon. We run open-source software on our own machines precisely so that everything we build stays ours. Same trade-off as running your own Linux server in 1999: someone has to maintain it, but you own the asset.

Your workflows are now an asset. Think about the exit before you need to. A decade ago your saleable assets were your brand, your customer list and your team. Today they also include your agentic workflows: the automated processes that make you faster and cheaper than your competitor. If you plan to sell or merge in the next few years, ask yourself how those get packaged for due diligence. Can you show a buyer exactly what's automated, what it saves, and where it lives? Because if half your efficiency sits in a personal ChatGPT account and half in some contractor's Claude Code setup, you don't own an asset. You own a mess with a subscription attached. Buyers discount messy. Plan the structure now, while it's cheap, so this isn't a fire drill during the deal.

The models are free and they're good enough. Qwen3.8 27B runs on local hardware and it's as capable as the best closed models from a few months ago. Good enough to run real business processes on. If you don't want your data leaving the building, you no longer have to let it. The only real cost left in business automation is implementation: someone sitting down with you to work out which workflows matter.

That last part is the actual bottleneck, and it's the part we do.

So here's my challenge for this week: pick one workflow. Automate it badly if you have to. You'll learn more in that afternoon than in a month of reading about AI.

And if you already have five broken-ish automations humming along unattended, or you're thinking about what happens to yours when it's time to sell: reply to this email. Both are conversations worth having now, not later.