AI for the skeptical business owner: how to actually start.

1:22 · The full post is below.

You don't need to learn AI. You need to learn to direct it.

Every transformative technology that's come along — the telephone, the internet, GPS dispatch — had the same shape to it. Nobody who won from it built the technology. They were the ones who directed it best, while their competitors were still arguing about whether it was real.

AI is that technology now. Everyone is going to use it — that part is not optional, and the businesses that survive the next decade will be the ones that dominate it. But "dominate" doesn't mean write code. It means three things:

• Know what to ask — the questions you can put in front of a machine, and the ones you can't.
• Know what to check — a machine that's fast and occasionally wrong is only useful to someone who verifies.
• Own the output — the work it produces should be yours, in a form you can keep, move, and fix yourself.

The mental model I give every owner: think of it as extending your own brain, not hiring a robot. A skill is a procedure you write down once — how you answer a review, how you write a quote, how you follow up on a lead — that the machine then repeats exactly, every time, without being told. A knowledge base is the raw material it works from: your services, your pricing, your past work, your customers. Skills are what it can do. The knowledge base is what it knows about you. Build those two, and you haven't adopted a tool — you've extended yourself.

The objection I take seriously: "I'm not going to depend on a machine"

This is the objection that's usually dismissed with a smile. It shouldn't be. You already depend on machines you don't understand — your payment processor, your email, your bank, the map that tells your crews where to drive. Dependence stops being a problem only when three things stay true:

• Your data stays visible to you — in plain files you own, not locked inside someone's app.
• You can leave — month-to-month, not a contract that traps you.
• You still hold the judgment — the machine drafts, you approve. You depend on the draft, never on the decision.

Here's the part most people miss: with AI, all three of those are within your control. You can run it month-to-month through an API and switch providers in an afternoon. Or you can run it on a machine in your own office, where nothing ever leaves the building. The dependence is real either way — but it's your dependence, on your terms, and it's reversible. That's as safe as "depending on a machine" gets for anyone, owner or not.

The privacy and IP angle

The thing that keeps skeptical owners skeptical — for good reason — is this: your customer list, your P&L, your contracts, your pricing. That's the crown jewels of the business, and the default way "AI" gets pitched is to paste it all into a chat window on a stranger's server. If that's your only option, your skepticism is correct.

It isn't your only option. Open-source models — free to run, free to modify, free of anyone's terms of service — are now genuinely good. The one I run day to day is Qwen3.8-27B, from Alibaba: released under the Apache 2.0 license, 256,000-token context, reads images and text, and a 4-bit build fits in about 17 GB of memory. On the right machine, that model — or bigger ones — runs entirely inside your office. Your data never leaves. There is no chat window, no terms of service, no one to leak it to.

Two ways to run it — with real numbers

For a local service business there are exactly two sensible setups:

	Path A — Cheap cloud	Path B — Local machine
What it is	You call the model through an API (OpenRouter and similar). Nothing installed, nothing maintained.	The model runs on a Mac Studio in your office. Nothing leaves the building.
Cost	Qwen3.8-27B runs roughly $0.15–$0.45 per million input tokens. A heavy month for a small business costs under $10.	A 256GB M3 Ultra Mac Studio is roughly $5,600–$7,500 depending on configuration; the new M5 Ultra at 256GB is about $9,499 (a 512GB version lands at the end of October). One-time, no per-use fees.
Your data	Goes to the provider per request. Fine for marketing content, quotes, scheduling — not for client P&Ls or contracts.	Never leaves the machine. The right call when confidentiality is part of the service — legal, medical, financial, or just the way you run things.
Switching cost	Zero. Month-to-month, and your workflows are plain text you keep.	Low. It's hardware in your office; the models are open weights, not subscriptions.
Start here if…	You want to start this week with zero hardware decisions. This is where 9 out of 10 owners should begin.	Your clients' data can't leave your hands, or you're running enough AI that the API bill stops being trivial.

Honest version of the advice: start on the cloud, keep everything in plain files, and only buy the machine when the data demands it. Most owners never need to — and the ones who do will have already built the workflows that make the hardware worth it.

How to actually start — this week, not next quarter

You do not take a course. There is no curriculum, and anyone selling you one is selling you time. Learning to direct AI is a loop you run on your own business:

• Day 1 — Pick one real task. Something you already do five or more times a month: answering reviews, writing quotes, following up on leads, writing your website content. Not a demo task. A task you'd rather not do.
• Day 2 — Do it once with the machine, and check every line. This is the actual learning. Not the prompting — the checking. Every time you catch something wrong, you've added a rule you didn't know you had.
• Day 3 — Write down what you did. What you gave it, the steps, the rules, what "done" looks like. That document is your first skill. The machine now does the task the same way every time, without being told.
• Week two — Run it five times. Refine the skill after each run. By the end of the week it's doing the task better than you do it on a bad day.
• Month one — Repeat for three or four tasks. Add a knowledge file: your services, your prices, your best work, your customers' names. That's the brain.

Four weeks later you have three to five skills and a knowledge base that encodes how your business actually works. That's more AI literacy than most of your competitors will ever have — and it cost you a few hours, not a course.

Where the machine will still disappoint you

So nobody's surprised later:

• It doesn't do the physical work. It doesn't install your HVAC or run your line. It handles everything around the work — and that "around" is where most of your month actually goes. On that one, see the line-item breakdown of where AI cuts your month.
• It is confidently wrong sometimes. It will invent a detail, smooth over a number, write something that's 90% right. That's exactly why the Day-2 rule exists: you stay the one who approves. The machine that runs your business is a machine you supervise — permanently.
• It doesn't make the judgment calls. Pricing a tough job, turning someone away, reading a client. Those stay yours. That's a feature. A business where the machine makes the calls is a business for sale.