The Delegation Ladder
Delegating work to an AI system is not one decision about whether to trust it. It is a ladder of five rungs, from the machine typing while you think to the machine owning the outcome and reporting only exceptions. No rung is the correct one. The work sets the rung, and two questions decide which.
Framework 8 of 10 · Chapter 8: The Room Nobody Booked
Most engineers treat delegation as a switch. Trust the tool or don't. Run the agent or write it yourself. Held that way, the question has no good answer, because the honest answer changes with every piece of work that crosses your desk.
It is not a switch. It is a ladder.
The five rungs
- Typing. The machine types while you think. Every decision is still yours.
- Drafting. It drafts, and you read every line before anything moves.
- Task. It completes the task, and you check the result against something you already trust.
- Direction. It runs a whole direction, and you meet it at checkpoints you set in advance.
- Outcome. It owns the outcome, and you hear about the exceptions.
When it applies
Any time you are about to hand something over, or are refusing to and cannot say why. It applies with most force on work you have done by hand for years, because that is the work whose reasons are least likely to be written down anywhere.
How to run it
Two questions set the rung:
- How much of the judgment this work needs is written down somewhere the machine can read it?
- How expensive is wrong, and how fast would you find out?
Answer both in writing before you start. The answers give you the rung, not your comfort with the tool, and not the rung that worked on the last piece of work.
If the answers say the work could live a rung higher than where you are running it, the gap is an artifact you have not built yet: a fixture set that holds the reason a threshold is lopsided. An invariant written at the top of the context file. An alert that fires on the way out instead of a review that blocks on the way in. Name the one artifact that would move the work up a rung, and build that instead of doing the work by hand again.
Moving work up the ladder has a price, and the price is always the same two things. You write the judgment down, and you make wrong cheap enough to survive.
Where it fails
Too high. You hand over work whose judgment lives only in your head, and the output is excellent, and it is wrong in a way nobody reviewing it could see. This failure is loud and arrives late.
Too low. You read every line of everything, you become the slowest component in your own system, and you call it rigor. This failure is quiet and can run for years without anyone naming it.
Practice
Take one item you are about to delegate, or have been refusing to delegate. Write the two answers down, not as a note to yourself but as two sentences someone else could read. Then write the rung the answers actually justify. If it is higher than where you have been running the work, you have just written your own next task, and it is not the work itself.