Why I stopped using AI for the part everyone uses it for

The default way to use AI for knowledge work is now so obvious it barely needs stating: you have a thing to write — a memo, an analysis, a recommendation — and you ask the machine for a first draft, then you edit. Draft fast, refine after. It feels like the whole point. It’s in every productivity guide. For about a year, it was how I worked too.

I stopped. Not because of some purist objection to machine-written prose, and not because the drafts were bad — they were fine, often better than fine. I stopped because I noticed the first draft was never really the work, and by letting the machine do it, I was skipping the part where the actual thinking happened. The time I “saved” was being quietly debited from somewhere I couldn’t see until later.

This is an essay about that hidden debit, and about the one category of work where I now refuse to start with AI no matter how tempting the time saving looks.

The first draft was never the deliverable

Here’s the thing nobody tells you about writing a recommendation for a client: the document is a by-product. The real product is the thinking that the act of writing forces you to do. When you write “therefore we should,” you are compelled to have actually worked out the therefore — to have felt the weak joints in your own logic, the place where the argument almost doesn’t hold, the objection a smart skeptic would raise.

That struggle — the staring, the deleting, the rewriting of one sentence five times — isn’t friction in the way to the deliverable. It is the deliverable, happening in real time inside your head. The polished document is just the residue it leaves behind.

When AI writes the first draft, it produces the residue without the reaction. You get the shape of a worked-out argument without having worked it out. And here is the trap: it reads as finished, so you edit it as if it were finished. You smooth the prose, fix the transitions, and ship something that looks like thinking and is actually a very good impression of thinking — including to you.

A first draft you didn’t write is a conclusion you didn’t earn. It will hold up right until someone pushes on it in the room — which is the only moment that ever mattered.

How I noticed the bill

I didn’t reason my way to this. I noticed it in meetings.

There was a stretch where I was shipping more, faster, with AI drafting underneath me — and where I was, more often than before, getting caught flat-footed when a client pushed on a specific claim. Not on the big thesis, which I’d thought about, but on the connective tissue — the why exactly behind a step, the second-order implication, the assumption I’d have noticed myself making if I’d written the sentence by hand. The model had written those parts smoothly enough that I’d edited right past the places where I didn’t actually understand my own argument.

The time AI saved me on the draft, I was paying back with interest in the room, in the currency that matters most: looking like I hadn’t fully thought it through. Because I hadn’t. I’d outsourced the thinking-by-writing and kept only the typing.

That’s the hidden debit. The savings are immediate and visible; the cost is delayed and shows up somewhere you don’t connect back to the cause. Which is exactly the structure of every bad trade.

The line I now hold

So I drew a line, and it’s specific. I do not let AI produce the first articulation of anything where my own reasoning is the product. Recommendations, strategic arguments, anything where a client is buying my judgement — I write the first version myself, by hand, struggling through it. The struggle is non-negotiable because the struggle is the thinking.

This is not a blanket rejection of AI for writing, and I want to be precise about where the line falls:

• First articulation of my own reasoning → I write it. No AI. This is the protected category.
• Everything around that core → AI freely. Restructuring something I’ve already reasoned through. Tightening prose whose logic I already own. Generating counter-arguments so I can pressure-test a position I arrived at. Summarising material as input to my thinking. Drafting the parts that are genuinely mechanical.

The test is a single question: am I about to use AI to skip the part where I figure out what I think? If yes, I stop and write it myself. If the thinking is already done and what remains is execution, the machine is welcome to all of it.

Notice this is the same boundary from every other workflow I run: machine proposes, human disposes — and you never let the proposing quietly replace the disposing. In the research workflow it’s “don’t automate the judgement about which tension matters.” Here it’s “don’t automate the reasoning that the writing is supposed to generate.” Same line, different altitude.

The uncomfortable part

The reason this is hard — the reason almost nobody holds this line even after they’ve felt the bill — is that writing your own first draft is slower, and the slowness is visible and immediate while the cost of not doing it is hidden and deferred. Every incentive in a busy professional’s day pushes toward the fast, finished-looking residue.

I’m not going to pretend I resolved that with willpower. I resolved it by reframing what the slowness is for. The forty minutes I spend writing a recommendation badly by hand isn’t forty minutes of typing I could have automated. It’s forty minutes of the only activity a client is actually paying a senior person for — and the document at the end is proof the thinking happened, not the thinking itself.

The machine is extraordinary at producing the residue of thought. Be very careful before you let it, because residue is convincing, and the only place the difference shows up is the one place you can’t afford it: in the room, when someone who’s paying you pushes back, and you find out whether the therefore was ever really yours.

Next in Systems → Building a second brain that actually pays off.