When It’s Faster to Do It Myself

I recently completed a task with AI that would be suitable for an intern.

I understood the context and the main problem. I wasn’t sure of the solution, but I had a clear goal. With recent LLM models, I could investigate the code, make changes, and test them without doing the implementation myself. I was still deciding what to try and whether the results were good enough.

AI helps both of us work faster. But if doing the task myself now takes half the time, while explaining and reviewing it takes almost as long as before, I save less time by delegating. For example:

That makes doing it myself tempting. But if I keep making the decisions, when does someone else get to learn how?

Part of what I want to help an intern learn is to bring an approach, explain why they chose it, and show what they checked. I want to review their reasoning instead of supplying it. That takes time together, even when doing the current task myself would be quicker.

AI can reduce that support time too: the engineer may use it to find context, investigate alternatives, and catch mistakes before review. That changes how much time I save by delegating. It doesn’t settle whether the assignment is worthwhile. An intern can learn something valuable even if it never saves me time during the internship.

The work can get done, and someone can learn from it, while I still need to direct the next task. That made me wonder what changes this over time.

The same task, four settings

Imagine an integration with an external service: someone needs to investigate, implement, and test it. Here are four combinations of support, time together, and room to make decisions, with AI available to everyone.

Who is directing the investigation?

A task can be completed correctly while the lead still chooses what to investigate, what to try, and what evidence is enough. The work has moved to someone else; directing it has not.

I may still review every significant decision, especially when the consequences demand it. That alone does not tell me whether delegation is working. What matters is whether I am reviewing someone else's reasoning or still doing the reasoning for them. I remain accountable in either case.

Knowing when to ask for help is part of that judgment. Fewer questions are useful only when the person can recognize what needs escalation.

A review I remember

In one review, an intern brought examples of incorrect answers that our matching rule had counted as correct. That made our earlier accuracy results look less reliable. I asked what that meant for our evaluation and which earlier conclusions we needed to revisit.

Finding the problem was useful. The next step was understanding which conclusions might no longer hold and how to test them. I could help by asking questions without supplying the answers. What I wanted them to bring next time was their view of the “so what?”: what the findings changed, what they would do next, and where they were unsure.

I would still expect questions, reviews, and disagreements. The change I am looking for is having an approach to discuss, rather than having to supply one. That can happen in one familiar part of a system long before it happens elsewhere.

What I would do differently

What we need to consider is how easily I could make the same decision on the next task, and the next, because each looks cheaper to do myself.

I would now ask what I am leaving for someone else to learn to decide. If I keep the unfamiliar problems and delegate only the steps I have already worked out, I should not be surprised when I am still needed to work out the next one.

I would look for work where someone can make real decisions, but a mistake or delay is manageable. I would ask for their approach before breaking the problem into steps myself. They could use AI to explore options and check their work before we talk.

This may take more of my attention today, not less. In a small team, that investment competes with work that needs to get done now. Doing the task myself may be faster today; delegating it may create more capacity later. The difficulty is knowing which investment I am making.