I don't do predictions. People who do mostly get paid for being confident, not for being right. But I do run a hosting company's infrastructure by day and a handful of automated businesses by night, and from that seat a few things have become clear enough that I'll build on them.
What has changed
Models got good enough that the model is no longer the bottleneck. Two years ago the question was "can it do this?" Now it's "can I feed it the right information and catch it when it's wrong?" That shift moves the hard part from research labs to people like you and me, wiring things together.
The unit of work moved from a chat to an agent. A chat answers a question. An agent takes a goal, uses tools, and comes back with a result. Everything on this channel — n8n workflows, MCP, Claude with connectors — is about that move. It is also exactly where the risk lives, which is why the pause-for-approval rule keeps coming back in the lessons.
Plumbing is being standardized. MCP and similar standards mean a tool you connect once works across models. Last year I rebuilt the same Gmail connection three times. This year I built it once. Boring, and it's the thing that makes small-business automation actually affordable.
Local models became usable. You can run a capable model on a machine you own, with no data leaving the room. For anything sensitive this changes what's possible, and it's why the Local AI series runs on hardware I actually have on my desk.
The scams arrived before the safeguards. Voice cloning and deepfakes are already in the wild, aimed at families and finance teams. The defense is not technical; it's a code word and a habit of calling back.
What I think holds
- Judgment doesn't get automated; it gets more valuable. The person who knows what a good customer reply looks like, what a fair price is, what a correct invoice contains — that person now reviews fifty of them an hour instead of writing five. The skill is knowing what "right" looks like.
- Whoever owns the data and the process wins, not whoever owns the model. Models are becoming interchangeable. Your customer history, your checklists, your way of doing things are not.
- Trust becomes the product. As generated content gets cheaper, "a real person checked this" becomes something people pay for. Build that in and say so.
- The gap is between people who have tried and people who haven't. Not between experts and beginners. Three weekends of building something small puts you ahead of most of the room.
What I'd learn, in order, if I were starting today
1. One automation tool, properly. n8n is my pick because it's visual, self-hostable and honest about what each step does. Learn triggers, one API call, one branch, one wait-for-approval. Then build something you'll actually use, like the lead reply or booking reminder in the small-business series. Two weeks.
2. How to talk to a model so it stays inside the lines. System prompt sets the rules; user content is data; the model gets the real documents instead of guessing; the output has a fixed shape you can check. This is 80% of "prompt engineering" and it's an afternoon of practice, not a course.
3. The security basics from the last post. Scoped credentials, spending caps, what never goes in, where the off switch is. One hour to set up, and it's the difference between a tool and a liability.
After that, follow your actual work. If you sell things, learn how models write and check listings. If you run a team, learn standup bots and summaries. If you have parents online, teach them the code word this week.
What I would not spend time on first: the math behind neural networks, fine-tuning your own model, or chasing whichever model topped a benchmark this month. All of that is real, and none of it is where a small business or a learner gets value in year one.
The thing I keep coming back to
Every lesson I make follows the same shape: show the real result, build it step by step, leave the mistake in, add the human check. I used to think that was a teaching style. I now think it's the whole philosophy. The future I'm building toward isn't one where AI runs my businesses. It's one where AI does the repetitive ninety percent and I make the calls that matter, with a report in my inbox instead of a fire.
If that's the version you want too, start with the free lessons, build one small thing, and email me when it breaks. That's when it gets interesting.