Stop chatting with AI and start handing it real work

Most people use AI the way they message a coworker: type a request, read the reply, type a refinement, read the next reply, and keep going until something usable falls out. That works for quick questions. For real work, and especially for a small team trying to get hours back, delegating tasks to AI this way is slow, and it keeps you involved in every tiny decision when the whole point was to get some of your attention back.

The better posture is delegation. You hand a scoped task to an agent, give it what it needs, let it run, and then review the result the way you would review work from a junior teammate. The shift sounds small. In practice it changes how much you finish in a morning, because you stop babysitting the middle of the task and only spend your judgment at the start and the end.

Why delegation beats conversation

Conversation makes you the bottleneck on every step. The agent does a little, stops, waits for you, does a little more. You are present for all of it, which means you are not doing anything else, which means you have automated nothing. You just moved your typing into a different box.

Delegation flips it. The agent owns the execution. You own the definition and the review. Those are the two parts where human judgment actually matters, and they are also the two parts you can do quickly. Defining a task well takes a few minutes. Reviewing a finished draft takes a few minutes. The forty minutes of grinding in between is the part you wanted to hand off, and in a delegation loop that is exactly the part you are not in.

The loop, step by step

Five moves, and they repeat.

Define the task. Be specific about the output you want, not the steps to get there. "Draft a reply to this customer that acknowledges the delay, offers the refund, and stays under four sentences" is a task. "Help me with this email" is a conversation starter that will cost you ten rounds.

Give context. The agent cannot read your mind or your files unless you hand them over. Paste the customer history, the relevant policy, the example of a reply you liked. Most bad agent output is not the model being dumb. It is the model working blind because nobody told it what good looks like.

Let it work. Resist the urge to interrupt halfway through and steer. If the task was defined well, let the agent finish a complete attempt. You learn more from reviewing a full draft than from micromanaging a partial one, and you stay out of the middle, which was the goal.

Review. Read the output with a critical eye. Is it correct? Would you put your name on it? This is where your expertise earns its keep. The agent produced a candidate. You decide whether it ships.

Correct and repeat. When something is off, do not just fix it silently. Tell the agent what was wrong and why, then let it try again or carry that lesson into the next task. The correction is how the loop gets tighter over time.

Keep the human in the loop on purpose

The review step is not a formality you drop once the agent gets good. It is the load-bearing part of the whole arrangement. An agent will produce confident, fluent, completely wrong output and hand it to you with no signal that anything is off. The only thing standing between that output and your customer is you reading it.

Scope the tasks so review stays fast. A task small enough to check in two minutes is one you can actually delegate at volume. A task so sprawling that reviewing it takes as long as doing it yourself was never a good candidate for delegation in the first place. When you feel review getting heavy, that is the signal to cut the task smaller, not to skip the review.

Start with one workflow you already understand cold, where you can spot a bad output instantly. Run the loop on it for a week. You will quickly learn which of your tasks are a clean fit for delegation and which still need you in the chair the whole time. That sorting is the real skill, and you only build it by running the loop on real work, not by reading about it.

Related reading

- [Treat AI output as a first draft, never a finished product](08-ai-output-first-draft.md)

- [A simple framework for deciding how to handle any AI task](22-deciding-how-to-handle-ai-work.md)

- [What daily AI use taught me about my own workflow](09-lessons-from-daily-ai-use.md)