Not every task is an AI task
The first decision in any AI collaboration happens before you type anything: should this task go to an AI system at all, and if so, which part of it? AI models are strong at tasks with a lot of precedent in their training data and a fair amount of tolerance for error — drafting, summarizing, explaining a concept, generating boilerplate, brainstorming options. They're weak at tasks that require verified up-to-the-minute facts, precise arithmetic over many steps, access to information that lives only in your head or your private systems, or a single correct answer where being subtly wrong is worse than not answering. Delegation is the habit of sorting a task into one of these buckets before you invest time in it.
A useful test: if a smart intern with broad general knowledge but zero access to your specific context could plausibly produce a decent first draft, it's probably a good delegation candidate. If the task depends on information only you have — your team's unwritten conventions, a fact you haven't told anyone, a judgment call that carries real consequences — either you need to supply that context explicitly first, or the task isn't a good fit for delegation at all.