What a Sabbatical to Learn AI Actually Looks Like

I took six months out to learn AI properly. Not to read about it, but to use it every day on real work until I actually understood what it could and couldn’t do. Apply it across my life and, above all, across my businesses. See what held up. That was the plan, written down on day one.

I’m going to write these up as I go: a quarterly report on where the time went and what it bought me. This is the first one.

So I did the obvious thing and pulled up my own task history to see what I’d been doing. A few thousand tasks. I expected it to remind me of all the clever ways I’d bent AI to my work. Instead it reminded me of something a bit different. I’d set out to apply AI, and what I’d mostly done was build with it.

I’m not mad about it. But it’s not the story I’d have told you if you’d asked me at the time.

What was the plan?

The short version: six months to go deep on applied AI, framed to myself as “use it on the real stuff, the businesses, the admin, the decisions, my own mess, and learn by doing.” Very reasonable. Very of-the-moment.

There was a second thing underneath it, quieter, that I only half-admitted at the time. I’ve spent years advising other people on building software and ventures without much room to build my own. I had an itch. Six months with AI doing the heavy lifting was, if I’m honest, a licence to finally scratch it: to prototype, to spin things up, to make.

The first thing I actually did, before a single line of anything else, was write a strategy document. Not “let’s try AI on my inbox.” A definition of what done-enough looked like, a success metric, and a first task that read: identify the highest-leverage things an assistant like this could do. With hindsight that’s the tell. I didn’t sit down to use a tool. I sat down to build one.

And what actually happened?

I have the receipts, because the whole point was to work in the open and keep the record. In these first months I closed 1,325 tasks.

One caveat on that number, because it matters: that’s not 1,325 pieces of AI wizardry. It’s my entire list. Every idea, every to-do, business and personal and home, finally out of my head and into one system. Some of the real wins of the six months weren’t clever uses of AI at all. They were just getting the contents of my brain into a workflow, even if that workflow was only a holding pen. You cannot act on what you can’t see.

When I sort those tasks by what they were actually about, the biggest category by a distance is building the assistant itself: the skills, the workflows, the plumbing, the thing that decides what to do next. Nearly half of everything I tracked was building the machine, not using it.

The rest is a mix I recognise. A run of prototypes, product ideas, and experiments, some scaffolded into working shape in days, most captured and parked for another time. The businesses got proper attention. The life and home admin, for once, actually got written down.

I set out to apply AI to my work, and instead built an increasingly elaborate apparatus for applying AI to my work, plus somewhere to finally put every idea I’d been carrying for years.

Isn’t that a polite word for failure?

I don’t think so, and this is the bit most writing about AI and productivity quietly skips.

If you actually try to put these tools to work, you hit the same wall I did. The raw model is clever but it has no idea who you are, what your business is, or what you decided last Tuesday. To get real leverage you have to build the context around it: the memory, the conventions, the little machine that hands it the right information at the right moment. That building is the work. Pretending you can skip it and go straight to “AI runs my business” is the fantasy being sold.

And the itch? Properly scratched. I got to be a builder again for six months, at a speed I’ve never had before. Some of it will go nowhere. Some already earns its keep, and all of it taught me something I couldn’t have picked up from the outside.

So no, I didn’t fail to use AI. I found out what using it properly actually costs, and I got a lot of building done while I paid.

Why quarterly reports?

Because I’m treating this like a proper piece of work, with a proper end. I’m aiming to wrap it up and return to full engagement in autumn 2026, if the world decides to settle down. When I do, I intend to walk back in with a stack of skills I didn’t have six months ago and the receipts to prove they’re real, not LinkedIn-real.

These posts are the running log of earning them. Out loud, mistakes included, so that when I say “I can do this now” there’s a paper trail behind it.

So what’s the series?

Over the next few posts I’ll go through, as straight as I can manage:

  • What I built: the portfolio of machines and half-machines, and why I spin up a new prototype most fortnights.
  • The shape of it: month by month, the frantic stretches and the week I did precisely nothing.
  • What I dropped and rebuilt: the least flattering post, and the most useful.
  • The two things I didn’t see coming: the privacy deal I signed without reading the terms, and the afternoon I found my own system had quietly erased the first three weeks of its own history.

I’m not writing this to sell you a workflow. There’s enough of that already.

This isn’t a victory lap. It’s a field report, including the parts that should worry you.

Posts in this series

Working LifeIdeas and Adventures