The Best Use of AI Isn't Doing. It's Deciding.

In the first week of this sabbatical I had the model open and ready to build. I closed it again without writing a line of code, and sat down to argue with it instead: what did I actually want out of the next six months, and was I about to spend them well? That conversation, not any of the tens of thousands of lines I went on to write afterwards, turned out to be the best use of AI I made all year.

Everyone reaches for the doing first

Watch anyone get their hands on a genuinely capable model for the first time and the pattern is the same. Write the post. Draft the reply. Summarise the document. Chase the actions from last week’s meeting. It’s satisfying, because you can watch it happen in real time, and it looks exactly like the productivity the marketing promised. But task execution was never the scarce thing. Doing has been getting cheaper for years, model or no model. Deciding what’s actually worth doing hasn’t got any cheaper at all.

The session that wasn’t a build

The most valuable hour I spent with AI all year wasn’t spent building anything. It was being asked, properly and repeatedly, what I actually wanted, and being made to answer instead of deflect. No consultant I’ve ever paid asked harder questions or let me get away with less. That’s the use of the tool nobody puts in a demo, because “I sat still and changed my mind” doesn’t make good footage.

Why leaders skip straight past it

Hand a capable model to someone running a business and watch what they ask it to do: summarise the market, draft the strategy deck, tidy up the roadmap. All doing. The actual judgement calls — what to stop, what to commit the whole team to, which of the twelve live initiatives deserves the attention — stay exactly where they always were: made under time pressure, by someone who never sat down long enough to think it through properly. Automate the busywork and leave the thinking to hurried instinct, and you haven’t made the organisation smarter. You’ve made it faster at the wrong things.

The bit I didn’t expect

Here’s what surprised me: the harder the question, the more useful the thing turned out to be, not less. It doesn’t get tired of being disagreed with. It doesn’t need managing, or a coffee, or careful handling of its ego. It will sit with you through the eleventh version of the same hard question and not flinch when the honest answer means undoing three months of work. Ask it to write your report and you’ll get competent prose. Ask it to argue with your plan until only the true part is left standing, and you’ll get something no report ever gave you.

I don’t know yet how far that generalises past my own sabbatical. But six months in, it’s the one claim in this whole series I’d make without hedging.

So point it at the thing that’s actually scarce

The task queue was always going to get automated, model or no model. What’s still genuinely scarce is someone willing to sit still long enough to decide, out loud, what’s worth doing at all. Point the most capable tool you own at your to-do list and you’ll get through it faster. Point it at the decision you’ve been avoiding, and you might actually change what’s on the list.

Which leaves me with a question I keep circling and haven’t answered. The thing I’ve described here — sitting down with someone and working out what’s actually worth doing — is most of what I know how to do, and I’ve spent the sabbatical building the scaffolding that makes me better at it. Should I be productising that? I honestly don’t know yet.

If you want to see what I’ve been building out here, get in touch. I’d rather talk it through with someone than keep turning it over on my own.

Companion essay to The shape of it so far — the AI Sabbatical, post 2.