Delegation: Deciding What to Hand Off to Claude
Before you ever type a prompt, there is a decision that shapes everything after it: should this task go to Claude at all.
//
Search across all documentation pages
Before you ever type a prompt, there is a decision that shapes everything after it: should this task go to Claude at all.
That decision is Delegation, the first of the four practices in Anthropic Academy's AI Fluency Framework.
It sounds simple, but getting it wrong in either direction costs something.
Hand off too little and you waste time on work Claude could reliably help with.
Hand off too much and you risk shipping something that needed human judgment it never got.
This page is about building the judgment to tell the two apart.
Delegation, in the AI Fluency sense, is not about whether Claude is "capable" of doing something.
Claude can attempt almost any task you describe.
The question Delegation asks is different: should this particular task, at this particular moment, be handed to Claude, and if so, how much of it.
Three properties make a task a strong delegation candidate.
The first is that it is bounded: there is a clear input and a recognizable "done," like summarizing a set of documents or drafting a first version of something.
The second is reversibility: if Claude's output is wrong or off-target, the cost of catching and fixing that is low, not a mistake that ships to a customer or triggers an action that cannot be undone.
The third is that it still ends with a human checkpoint: someone reviews the output before it goes anywhere that matters.
Tasks that fail one or more of these tests lean toward staying human-led.
A useful analogy is delegating to a capable new hire on their first week.
You would hand them a first draft of a report, background research, or a summary of a long document.
You would not hand them the final call on a layoff, a legal filing, or which vendor to fire, even if they could produce a confident-sounding recommendation.
The same instinct applies to Claude, and it is worth noticing that this is a judgment about the task, not a judgment about Claude's ability.
Delegation decisions typically fall into three categories rather than a strict yes or no.
Full delegation covers tasks that are bounded, low-stakes if wrong, and reviewable: drafting an email, summarizing a document, generating a first pass at an outline, doing repetitive research synthesis across many similar sources.
Human-led work covers final judgment calls, decisions involving another person's livelihood or wellbeing, legal or financial commitments, and anything where accountability cannot be transferred, meaning someone must be able to say "I decided this" and mean it.
Partial delegation is the category most people underuse.
It means splitting a task: keep the decision-making human, but delegate the drafting, formatting, or first-pass research that follows the decision.
A layoff decision itself is human-led, but drafting the announcement email once the decision, timeline, and terms are already set is a reasonable thing to delegate.
This split matters because many real requests arrive as a single bundled ask, "handle this for me," when they actually contain both a decision and a drafting task glued together.
Recognizing the seam between the two is most of the skill.
Bundled request: "Handle the vendor termination."
|
v
Split at the seam
/ \
Decision Drafting
(human-led) (delegate)
"terminate, "write the
effective date, termination
terms" notice given
these facts"Discernment, the third D, is what catches it when a Delegation call turns out to be wrong after the fact: if Claude's output on a supposedly "full delegation" task keeps needing heavy correction, that is a signal the task actually belonged in the human-led or partial category.
Stakes are not fixed properties of a task type, they shift with context.
Drafting a social media post is usually a safe full-delegation task, but drafting a public statement during an active crisis, on the same platform, with the same word count, is not, because the reversibility and accountability profile has changed even though the task "looks" similar.
This is why Delegation is a per-instance judgment, not a lookup table of task categories.
Audience matters too.
A first-pass literature summary for your own reading is a very different delegation decision than the same summary going directly into a report a client will act on without further review, because the human checkpoint that makes delegation safe is only present in the first case.
Volume changes the calculation as well.
Repetitive, similar tasks at scale, like triaging hundreds of support tickets into categories, are often better delegation candidates precisely because the stakes per item are low and errors are easy to catch in aggregate, whereas a single high-stakes document deserves more human involvement per unit of output.
| Task Shape | Delegation Fit | Why |
|---|---|---|
| Repetitive drafting or formatting | Full delegation | Bounded, low stakes per instance, easy to review |
| Research synthesis across many sources | Full delegation | Pattern-finding at volume is exactly where Claude helps most |
| First-pass analysis with a human review step | Full delegation | The checkpoint absorbs the risk of an imperfect first pass |
| Final decisions affecting people's livelihood, safety, or legal standing | Human-led | Accountability cannot be transferred to an output Claude generated |
| Communications about a human-led decision, once the decision is made | Partial delegation | Split the judgment call from the drafting that follows it |
Delegation judgment also compounds with experience: the more you notice where a "full delegation" call went wrong, the sharper your instinct gets for the next similar task, which is the practical reason this page exists as a skill to build rather than a rule to memorize.
The practice of deciding which parts of a task to hand to Claude and which parts need to stay human-led, based on the task's stakes, reversibility, and who is accountable for the outcome.
Final judgment calls, decisions affecting someone's livelihood or wellbeing, legal or financial commitments, and anything where accountability for the outcome cannot be transferred to an AI-generated answer.
Splitting a bundled request at the seam between a decision and the work that follows it, keeping the decision human-led while delegating the drafting, research, or formatting that comes after.
Ask what part of the request is a judgment call someone must own, versus what part is producing a document, summary, or draft based on a judgment call already made; the first stays human, the second can often be delegated.
No. The same task, like drafting a public statement, can be a safe full-delegation candidate in a routine situation and a human-led task during a crisis, because stakes and reversibility change with context, not just task type.
Yes. Keeping Claude away from bounded, low-stakes, repetitive work out of general caution wastes the tool's value; under-delegation is treated as its own failure mode alongside over-delegation.
At high volume, like triaging hundreds of similar tickets, the stakes per individual item are usually low and errors are easy to catch in aggregate, which often makes repetitive-at-scale work a stronger delegation candidate than a single high-stakes document.
Treat repeated heavy correction as a signal that the Delegation call, not just the prompt, may have been wrong, and reconsider whether the task belongs in the human-led or partial category instead.
No. Delegation only decides what gets handed off; a human checkpoint reviewing the output is one of the three properties that makes a delegation candidate safe in the first place.
Delegation comes first and determines how much Description the task will need; Discernment later checks whether the output was good, and if it repeatedly wasn't, that feeds back into future Delegation judgment as part of ongoing Diligence.
No. Accountability is about who owns the outcome and can explain why a decision was made; typing the prompt does not transfer that ownership, which is exactly why some decisions cannot be delegated no matter how the request is worded.
Stack versions: Written against the Claude model lineup current as of ~June 2026 - Claude Fable 5, Claude Opus 4.8, Claude Sonnet 5 (the default), and Claude Haiku 4.5. Model names, pricing, and product features move quickly - verify current specifics at platform.claude.com/docs before relying on them.
Reviewed by Chris St. John·Last updated Jul 16, 2026