Is "AI Skills" Actually a Skill, or Just a Subscription You're Paying For?

I left banking because the bonus math stopped working, started freelancing because I had no other plan, and within four months discovered that nobody had ever taught me how to make money when the money arrives irregularly and then doesn't arrive at all for six weeks.

Nobody teaches this. Not school. Not your parents, unless your parents were freelancers, which statistically they weren't. You learn it by being broke in public.

So when everyone started saying "learn AI skills to stay relevant," I paid attention. And then I actually did it, and tracked what it cost and what it earned, because tracking things is the only professional habit banking gave me that transfers cleanly.

Here's the accounting, without the inspiration.

What I spent in eighteen months

Courses: about $370. Three of them were good. Two were a person reading a PDF aloud over a screenshot.

Subscriptions: about $1,100 across the period, mostly two tools I used daily and one I kept forgetting to cancel.

Time: roughly 200 hours, which is the number that actually matters and the number nobody in "upskill now" content ever tells you to budget.

What I earned directly attributable to it: a $6,800 contract I would not have gotten otherwise, plus about $2,400 in retained work because I could deliver something a client wanted in two days instead of two weeks.

Net: around $7,700 over eighteen months, on about 200 hours. That's $38 an hour of study time, which is fine and not transformative, and I want to be honest that the "AI skills will double your rate" claims I read did not happen to me and I have not met anyone they happened to.

The thing that actually made money

It wasn't prompt engineering. That term was already going stale when I learned it.

The thing that made money was that I could take a client's ugly, manual, recurring process and automate the boring middle of it. Not the whole thing. The middle. The part where someone copies numbers from a PDF into a spreadsheet four times a week and no one has ever asked why.

That skill isn't "AI." It's process decomposition with an AI tool at the end. And the reason it pays is that the client can see the hours disappearing, which is the only kind of value they believe in.

If I were starting today with no money, here's the sequence I'd actually follow, which is not the sequence I followed.

Week one. Pick three tasks you personally do that are repetitive and text-based. Not your industry's tasks. Yours. Be specific. "Responding to inbound enquiries" is not specific. "Reading a client's form and writing the first paragraph of an assessment" is specific.

Week two. Automate one of them badly. Badly is fine. The point is to find the friction, and the friction is never where the tutorial says it'll be. Expect to be annoyed by API keys and file formats and the moment a model decides to return JSON with a preamble.

Week three. Charge someone for it. This is the step people skip indefinitely. You will not learn what's actually valuable by reading. You'll learn it when a paying client says "no, not that, the other thing" three times.

That's it. Three weeks, one task, one client. Cheaper than any course.

On the tools

Everyone asks which tools. It doesn't matter as much as the marketing implies, and it changes every four months, which means any specific recommendation I give you is depreciating while you read it.

What matters is that you can move between them without relearning your whole approach. If your entire capability lives inside one vendor's interface, you don't have a skill. You have a subscription with a personality.

I keep one rule: everything I build has to survive my tool changing. Data in a format I control. Prompts in a file on my own disk. If the company raises prices tomorrow, I should be annoyed and not paralyzed.

The part where it goes badly for freelancers

Two specific traps, both of which I hit.

The first is pricing like the AI does the work. It doesn't. It does maybe 60% of the keystrokes and 0% of the judgment. If you price at "it took me twenty minutes," you'll earn twenty-minute money and the client will wonder why they need you at all next quarter. Price the outcome, which is what you were always selling.

The second is replacing yourself. I built a system for a client that automated a weekly report a junior analyst used to spend four hours on. Three months later, they let the analyst go. I felt terrible and I also understood, because that was the assignment.

The version of this that's sustainable is building the thing that lets the person do more valuable work, not the thing that makes the person unnecessary. Same effort, different framing, and it's the difference between a client who trusts you and a client who's quietly calculating your own redundancy.

What I'd tell the version of me who just quit

The AI part is the easy part. It's a weekend. Everything on the internet says otherwise because the internet sells the weekend.

The hard part is knowing what to charge, which you learn by charging wrong for a while. And knowing which work is worth doing at all, which you learn by doing a lot of work that isn't.

The tool will not make you money. The tool will make your judgment cheaper to execute, and judgment is the only thing you're actually selling. Which means the return on learning AI is capped by how good your judgment is, and that's a much slower thing to improve.

Which is annoying, and also the good news, because it means the people flooding into this don't have a shortcut.

Code doesn't care about your feelings.