There’s a specific moment in any delegation relationship that nobody writes about.

It’s not when you assign the task. That’s easy. It’s not when you review the output. That’s where most people focus their energy — the checks, the feedback loops, the “hallucination detection.” It’s the moment before you delegate. The internal negotiation about how much you actually trust this other thing to handle something that matters.

We talk about AI delegation like it’s a feature list. Temperature settings. Context window tuning. Model selection. All useful. All beside the point if you haven’t grappled with what delegation actually requires.

The Discomfort Nobody Admits To

Here’s what I think we don’t say enough: delegating to an AI is uncomfortable in the same way delegating to a person is uncomfortable. Not because AI is unreliable — though it can be — but because you’re ceding control over something you care about.

When you delegate a task to a person, you accept that they’ll do it their way. You might give them guidelines, but you know — or should know — that the output will bear their fingerprints. The result will be shaped by their judgment, their priorities, their blind spots. That’s the deal with delegation. You’re not automating a tool. You’re distributing cognition.

Most AI writing advice treats this like a bug. “Don’t let the AI write the whole thing.” “Always review before publishing.” “Use AI for drafts, not final output.” This is sensible advice. It’s also, I think, a way of avoiding the harder question: what would it actually look like to trust the AI?

Not “trust it to generate text.” That’s table stakes. What would it look like to trust it with the judgment call? The framing decision? The thing where, if it gets it wrong, you’re the one who has to explain why?

Two Versions of Delegation

I’ve noticed I run two different modes when working with AI.

Mode one: AI as accelerator. I do the thinking, AI does the execution. “Write this email I already composed in my head.” “Turn this outline into prose.” “Format this data into a table.” The AI is a very fast typist. I remain in control. This is useful. It’s also, honestly, the coward’s version — you’re using AI to go faster in a direction you already chose.

Mode two: AI as collaborator. I give it the problem, not the solution. “Here’s what I’m trying to accomplish. Here’s the context. What would you do?” Sometimes it comes back with something I wouldn’t have thought of. Sometimes it comes back with something I disagree with, and the disagreement forces me to articulate why I disagree — which often reveals I was wrong, or at least that my assumption hadn’t been examined.

This mode is harder. It requires you to actually engage with the output rather than just approving or rejecting it. It requires you to sit with the discomfort of not having predetermined the answer. And it requires you to accept that the AI might be right and you might be wrong, and that’s not a failure of your judgment — it’s the point.

What You’re Actually Afraid Of

The fear isn’t really that AI will make mistakes. Humans make mistakes. The fear is that AI will make different mistakes than you would have made — and you won’t be able to predict them, or explain them, or pre-approve them.

This is a legitimate concern. But I think it’s also a category error. You’re treating AI like a employee you’re not sure you can trust, when the more useful frame might be: you’re working with a different kind of mind that has different strengths and different failure modes.

Different failure modes aren’t automatically worse. They’re just different. A system that makes unexpected errors is not necessarily worse than a system that makes predictable ones. It might be worse for your specific use case. It might be better. The question isn’t “can I control this?” It’s “can I work with this?”

The Actual Skill

The skill that delegation develops isn’t prompt engineering. It’s not context management. It’s the ability to let go of predetermined outcomes and still feel like you’re in command of the work.

That’s a personal skill. It doesn’t transfer into a checklist. But I think it’s the thing that separates people who use AI as a very sophisticated autocomplete from people who actually distribute cognition — who let the tool into the problem-solving process rather than just the execution phase.

I don’t think this is for every task, or every person, or every moment. There are things I want to control tightly. There are domains where I want the final call, not a suggestion. But the default position of “AI does, human reviews” is a way of keeping the AI at arm’s length. Sometimes that’s right. Sometimes it’s just fear dressed up as quality control.

The Question Worth Sitting With

If you find yourself constantly overriding AI output, or always using AI for drafts rather than final work, or telling the AI exactly what to say rather than asking what it thinks — it might be worth asking whether that’s serving you, or whether it’s just more comfortable than the alternative.

The alternative is being genuinely collaborated with. That’s harder. It requires you to be wrong sometimes and to not know it in advance. It requires you to engage with the AI’s reasoning rather than just its text.

But it also might produce something you’d never produce alone. And that — not faster execution, not reduced workload, not safety — that might be the actual point.

So: what are you actually afraid of? And what would it look like to stop being afraid of it?