What I Delegate to AI and What I Refuse To

I handed a first draft of a release summary to an agent last week without a second thought, and an hour later I spent twenty minutes rewriting a single sentence of a client note by hand, refusing to let the same agent anywhere near it. Same tool, same day, opposite instincts. I had never written the rule down, but my hands clearly knew it. One kind of work went out the door the moment it occurred to me to delegate it. The other kind I guarded without even pausing to ask why.

So I went looking for what the rule actually was, because an instinct that strong is usually pointing at something real. It turns out I have a line, a fairly consistent one, and I had never once said it out loud. Most of us are in that position now. We delegate some things reflexively and refuse others just as reflexively, and almost nobody has put words to the sorting rule underneath. That is worth fixing, because a line you cannot name is a line you cannot defend. When someone asks why you did a thing by hand that the tool could have done in seconds, "it felt wrong to automate" is not an answer you can stand behind, even when it happens to be correct.

First drafts of almost anything. Status updates, release notes, the skeleton of a plan, the boilerplate scaffolding around a document that has to exist but carries no judgment. The only real cost in that work is the blank page, the friction of starting, and that is exactly the cost an agent removes cleanly without removing anything I actually needed to keep. I am not handing off the thinking. I am handing off the part before the thinking, the staring at nothing.

Summaries of things I am still going to read myself. A long thread, a dense PDF, an hour of meeting transcript. The summary buys me a fast way in, a map of where the important parts are. What it never does is replace the reading, and I never let it pretend to. The moment a summary becomes the only version of a document I have in my head is the moment I have handed off understanding I am still on the hook for, and that is a different list entirely.

First-pass code on a reversible path with tests around it. If the agent gets it wrong, the tests shout, the blast radius is contained, and the cost of the mistake is a few minutes. That shape of work, where being wrong is cheap and catchable, is precisely the shape I am happy to check rather than produce from scratch.

The pattern under all three is the same once you look for it. I delegate the work whose output I can verify faster than I could have produced it myself. If checking is cheaper than doing, the agent does it and I check. That is the whole first list.

The decision itself. What we build, what we cut, what "on track" honestly means this week. An agent can lay the options out for me beautifully, better than I would, and I use it for exactly that. But it cannot own the call, because when the call turns out wrong it is me in the room with the client, not the model. Accountability does not delegate, so the decision it attaches to does not either.

Anything with my name on it gets my own eyes, line by line, before it goes anywhere near another person. I learned that one the hard way, presenting something I had only skimmed because it looked finished, and I am not interested in learning it a second time. The polish an agent produces is real, and it is exactly what makes skimming feel safe, which is exactly why I do not let myself.

The first real conversation with a client who is unhappy, or a teammate who has gone quiet. The judgment about a person in a room, what they are actually worried about underneath what they said, is not something I will hand to a tool that was not in the room and cannot read the thing that was never written down.

Delegate production. Keep the decisions. And never delegate the understanding you remain accountable for, because handing off the understanding while keeping the accountability is the one trade that always comes back around. It is the worst of both: you own the outcome and you have outsourced the one thing that would have let you steer it.

The uncomfortable part is that the line will not hold still. Things I refused to delegate a year ago sit on the first list now, not because I got more trusting but because the tools got better and my ability to check their output got sharper. As checking gets cheaper, more work crosses from the refuse column to the delegate column, and the border keeps migrating under me. So this is not a settled map I can laminate and stop thinking about. It is a thing I redraw every few months, which is its own quiet admission about the pace of the year, and about how much of what felt like principle was really just a snapshot of what the tools could do in a given month.

So, genuinely, not rhetorically: do you know where your own line is, and is it actually where you think it is, or just where it happens to be convenient this quarter?