What you will be able to do
- Break an existing business workflow into separate steps and record what happens when each step goes wrong
- Rate each step on reversibility, stakes, need for human judgment or empathy, and accountability
- Label each step AI-executed, human-retained or human-reviewed, and say why the label fits
- Put a redesigned workflow in an order where human decisions and checks come before anything that depends on them
- Recognize the signs that a workflow has been over-delegated
Key concept
Delegation — The decision, made one step at a time, about which parts of a piece of work AI carries out and which stay with a person. It starts with breaking the work into steps, not with the question of whether to use AI at all.
1.Delegation is decided step by step
Delegation is about where AI fits inside a workflow. Whether to use AI at all is the wrong question. Delegation is one of the four AI Fluency competencies, and its job is deciding what AI handles and what stays with you. That decision makes sense only at the level of a single step. A refund dispute, a donor thank-you campaign or a quarterly report each holds several steps. Some of them suit Claude well, and some must never leave a person's hands.
Delegation therefore starts by breaking the work into steps. How to split work into parts is covered in task decomposition. Here, decomposition is the first move of a redesign: once a workflow is split apart, each step can get its own owner. A team that labels the whole workflow "AI-assisted" has made no delegation decision. It has only stopped looking.
Each step gets one of three labels:
- AI-executed: Claude does the step, and the result moves on without a person approving that specific output.
- Human-reviewed: Claude produces the output, and a named person checks it before anything downstream relies on it.
- Human-retained: a person does the step, and AI may at most supply inputs to it.
These labels are not ranked, and a good redesign usually uses all three. One illustrator's workflow shows why. Applying the same delegation analysis led to different outcomes for different pieces of work. The goal is never to maximise the AI-executed column. Each label has to be one you could defend to a peer who challenged it.
Checkpoint 1 of 8· Check yourself
A support team redesigning its refund-dispute workflow reports that "the whole process is now AI-assisted." What should the team have done to make a real delegation decision?
Delegation is decided one step at a time: first break the work into steps, then give each step its own label. A single label for the whole workflow, or an attempt to fill the AI-executed column, means no delegation decision has been made.
“Delegation means decomposing the problem first, then deciding what AI handles at each step.”Source: academy.claude.com
2.Four criteria that decide the label
Four properties of a step decide its label. Effort, length and boredom are not among them. A label that follows from how unpleasant a task is will put the wrong steps with Claude.
Reversibility comes first. If a step's mistake can be caught and undone before it matters, the cost of an AI error is a quick edit. A draft that colleagues can freely rewrite before anyone outside the team sees it is fully reversible. A step that sends money, makes a commitment to a customer or deletes a record is not. Even guidance for highly capable agents keeps this line: retain your own confirmation step for risky or irreversible actions.
Stakes are the size of the harm if the step goes wrong. Stakes decide how hard a human-reviewed output gets checked, and the rule is to verify in proportion to the stakes. How deep that review goes is covered in review depth and escalation. For delegation, the point is simpler: when stakes are high, at least review, and often retention.
Need for human judgment or empathy covers steps where the right answer depends on values, relationships or reading a person in the moment. Comforting a distressed customer and reading their emotional state as a live conversation unfolds is the clearest case. Teachers drafting their own AI commitments land in the same place, refusing to delegate final judgment, relational feedback and high-stakes evaluation.
Accountability asks who answers for the step's result. If a named person must stand behind a decision, that person has to actually make it. Owning the outcome regardless of AI's role is the competency called diligence, covered in policy, ethics and diligence. In delegation terms, a step where someone must answer for an irreversible decision is the strongest case for human-retained.
| Criterion | Points toward AI-executed | Points toward human-retained |
|---|---|---|
| Reversibility | Output can be edited or discarded before anyone relies on it | Commits the organisation to something that cannot be undone |
| Stakes | An error costs a quick correction | An error harms a customer, a donor or the organisation |
| Judgment or empathy | Follows a fixed checklist applied the same way every time | Requires weighing values or reading a person's emotional state |
| Accountability | No one has to answer personally for the output | A named person must answer for the decision |
Steps in the middle of the table, which are reversible but still consequential, usually become human-reviewed. A fixed checklist that has been applied the same way for years needs no fresh judgment, so it leans toward AI-executed. A long, repetitive step leans the same way.
Checkpoint 2 of 8· Check yourself
In a charity's donor thank-you workflow, which of these steps has the strongest case for being labelled human-retained?
Returning a gift sends money and can't be undone, and a named person has to answer for it. When irreversibility and personal accountability come together, that is the strongest case for human-retained. The draft can be reversed, the template follows a fixed checklist, and tedium is not a reason to keep a step human, so all three lean toward delegation.
“You own the outcome, not the output. “AI wrote it” explains nothing and excuses nothing.”Source: academy.claude.com
Checkpoint 3 of 8· Exam question
Which TWO characteristics of a workflow step most strongly justify keeping it human-retained? Choose two.(Select 2)
Correct answers: A, C — The step requires comforting a distressed customer and reading their emotional state during a live conversation.; The step commits the company to a decision that cannot be undone, and a named person must answer for it.
- A. Correct. Empathy and live emotional judgment are core reasons to keep a step with a person, since Claude cannot truly share responsibility for the relationship.
- B. Incorrect. Time-consuming repetitive text is a classic reason to delegate to Claude, not to retain; effort alone is not a Delegation criterion for keeping work human.
- C. Correct. Irreversibility combined with a need for a named accountable person is a strong reason to retain the step, since errors cannot be corrected afterwards.
- D. Incorrect. A freely editable internal draft is reversible and low-risk, which points toward delegating with a normal review rather than retaining.
- E. Incorrect. A stable checklist is rule-based and predictable, which makes the step a good candidate for delegation rather than a reason to retain it.
3.Redesigning a workflow in the right order
A redesign runs as a chain in which each stage uses the output of the one before it. Understand each step before you judge it, judge it before you assign an owner, and test the design before you rely on it. Running a stage early leaves it with nothing to work from.
The five stages, in order:
- Inventory. List every step of the current workflow separately, including what happens if that step goes wrong. The failure consequence is the raw material for the reversibility and stakes ratings.
- Rate. Score each step on reversibility, stakes, need for judgment or empathy, and accountability.
- Label. From those ratings, mark each step AI-executed, human-reviewed or human-retained.
- Assign owners. Name one accountable person for every human-retained and human-reviewed step. Until the labels exist, you do not know which steps need an owner.
- Pilot. Run the redesign on a small batch, check the results against a written definition of done, and confirm that reviewers actually catch problems. Move any step that proves over-delegated back toward human control.
The pilot is where a design on paper meets real work. Piloting cautiously is a sensible default whenever a team is unsure an AI use is appropriate. Concrete acceptance statements give the pilot something to measure against. A pass or fail that a stranger could judge beats "the workflow feels smoother." The aim is a repeatable AI-assisted workflow with appropriate human review. Making that workflow reusable across a team is the subject of repeatable workflow design.
Checkpoint 4 of 8· Put it in order
An operations lead is redesigning the team's monthly vendor-invoice workflow so Claude can take on part of it. Put the redesign stages in the order they should run.
- 1.Mark each step AI-executed, human-reviewed or human-retained based on those scores
- 2.List every step of the current workflow separately, noting what happens if each one goes wrong
- 3.Run the design on a small batch, check the results against a written definition of done, and move over-delegated steps back toward human control
- 4.Score each step on reversibility, stakes, need for judgment or empathy, and accountability
- 5.Name one accountable person for every human-retained and human-reviewed step
Each stage uses the output of the one before it. Failure consequences feed the ratings, and the ratings decide the labels. The labels show which steps need an owner, so owners can't be named earlier. The pilot then tests the finished design before anyone relies on it.
“Delegation means decomposing the problem first, then deciding what AI handles at each step.”Source: academy.claude.com
Checkpoint 5 of 8· Exam question
A team is redesigning an existing workflow using delegation criteria to decide which parts Claude executes and which people keep. Which sequence should the team follow?
Correct answer: B — List each step and its failure consequences → Rate steps on stakes, reversibility, judgment, accountability → Label steps AI-executed, human-retained, or human-reviewed → Name accountable owners for redesigned steps → Pilot on small batch with seeded errors
- A. Naming owners before labeling is misplaced: until each step is labeled AI-executed, human-retained or human-reviewed, the team does not know which steps need a human owner, so it would assign owners by guesswork and may have to redo them.
- B. The team lists steps and their failure consequences, rates them against the criteria, labels each step from those ratings, names owners for the labeled human steps, and pilots the finished design to check that reviewers catch seeded errors.
- C. Naming owners right after listing is misplaced: it happens before the ratings on stakes, judgment and accountability exist, so owners are chosen from habit rather than evidence and before the labels show which steps stay with people.
- D. Piloting before owners are named is misplaced: the pilot is meant to check that reviewers catch seeded errors, but without accountable owners in place there is no real reviewer to test, so the pilot results would not reflect the finished design.
4.Worked example: a refund-dispute workflow
A support team redesigns its refund-dispute handling under two stated rules. The agent makes the remedy decision before any reply is drafted, and nothing reaches the customer until the agent has approved it. Both rules come straight from the criteria. A remedy commits the company and the agent answers for it, so it is human-retained. The customer reply is consequential but editable before sending, so it is human-reviewed.
Ordering follows one principle: evaluate what AI gives back before acting on it. Each human check sits right after the AI output it checks and before anything that relies on it. The remedy decision depends on the facts of the complaint, so a verified summary must come before the decision. The reply must reflect the remedy, so the decision must come before the draft.
| Order | Step | Label |
|---|---|---|
| 1 | Claude summarizes the complaint thread into the customer's main claims and dates | Human-reviewed (output checked in step 2) |
| 2 | The agent compares the summary against the original thread to confirm it is accurate | Human check |
| 3 | The agent decides the remedy, such as a refund, credit or replacement | Human-retained |
| 4 | Claude drafts a reply that reflects the remedy the agent chose | Human-reviewed (output checked in step 5) |
| 5 | The agent edits the draft, approves it and sends it to the customer | Human check and send |
Two orderings that look efficient both break the rules. If Claude drafts the reply before the agent decides, the draft quietly suggests a remedy, and the agent ends up approving a decision instead of making one. If the agent decides from an unchecked summary, a misread date or a missing claim feeds straight into an irreversible commitment.
The remedy is the irreversible, accountable step, and it is made from the summary. An error found after drafting would mean the decision had already been made on bad facts. Checking the summary first means the human-retained decision rests on verified inputs.
Checkpoint 6 of 8· Check yourself
To speed up refund disputes, a manager suggests changing the order of the redesigned workflow. Which sequence still keeps the remedy human-retained and every human check in the right place?
Each human check comes right after the AI output it checks and before anything that relies on it. The verified summary feeds the remedy decision, and the agent's decision feeds the draft. If the draft or a Claude-proposed remedy comes first, the agent ends up approving a remedy instead of choosing it. If the summary is checked late, an irreversible decision rests on facts nobody verified.
“Discernment means evaluating what AI gives back before acting on it.”Source: academy.claude.com
Checkpoint 7 of 8· Exam question
A support team redesigns its refund-dispute workflow. The remedy decision stays with the agent and is made before any reply is drafted. Nothing reaches the customer until the agent has approved it. Which sequence of steps follows these rules?
Correct answer: C — Claude summarizes the complaint thread → Agent checks summary against original thread → Agent decides the remedy to offer → Claude drafts reply reflecting the remedy → Agent edits, approves and sends reply
- A. The agent decides the remedy before checking Claude's summary against the original thread. The decision could then rest on an inaccurate summary, and the later check comes too late to protect it.
- B. Claude drafts the reply before the agent has decided the remedy. This breaks the stated rule that the decision comes before any drafting, and the draft could promise the wrong remedy.
- C. Claude summarizes the thread first, and the agent verifies that summary against the original so the remedy decision rests on checked facts. Claude drafts only after the agent has chosen the remedy, and the agent edits, approves and sends the reply last.
- D. The agent checks the summary only after the remedy is decided and the reply is drafted. The remedy and the draft would both be built on an unverified summary, so any error in it is baked in before it is caught.
5.Recognizing over-delegation
Over-delegation means a step that meets the criteria for human control has drifted to AI, or a human check has become a formality. It rarely announces itself. It shows up months later in how people work, so the useful skill is knowing the observable signs.
Signs that a workflow has been over-delegated:
- Skill atrophy. Staff can no longer do a step without Claude, so nobody could take it back or judge Claude's output well.
- Convergence. Outputs that should be distinctive, such as appeals, proposals or customer replies, have started to sound alike.
- Reaching for AI before knowing what you want. People open Claude to work out what a step should produce, so the judgment that belonged in the step never happens.
- No one who answers for the result. When something goes wrong, nobody can say who owned the decision.
- "AI wrote it" as an explanation. That phrase explains nothing and excuses nothing, and hearing it means accountability has slipped out of the workflow.
The remedy is to re-run the criteria, not to ban AI. The creative-practice version of the method makes this concrete: start from no AI and let the analysis decide whether and how to opt in. A step pulled back to human-retained can move forward again once the reason for pulling it back is gone. That only happens if someone revisits the labels. A delegation policy without revision triggers goes stale silently.
Checkpoint 8 of 8· Check yourself
Six months after a fundraising team handed appeal drafting to Claude, the director audits the workflow. Which finding is evidence that the work has been over-delegated?
"AI wrote it" offered as an explanation means nobody owns the result. That shows accountability has slipped out of the workflow, which is a sign of over-delegation. Time saved and higher volume are not evidence either way, and merging names into a fixed template is a sensible step to delegate.
“You own the outcome, not the output. “AI wrote it” explains nothing and excuses nothing.”Source: academy.claude.com
Sources
(11)
Every claim above is drawn from one of these pages, quoted as it was written on the date shown.
- 1.
“Delegation means deciding what AI handles and what stays with you.”
↩︎ Delegation is decided step by step“Diligence means owning the final result fully, regardless of AI’s role.”
↩︎ Four criteria that decide the label“Discernment means evaluating what AI gives back before acting on it.”
↩︎ Worked example: a refund-dispute workflow“Discernment means evaluating what AI gives back before acting on it.”
↩︎ Exam trap 2 - 2.https://academy.claude.com/courses/ai-fluency-for-builders/delegation-the-builder-s-toolkitOfficial docs
“Delegation means decomposing the problem first, then deciding what AI handles at each step.”
↩︎ Delegation is decided step by step“Write acceptance tests before code. They give you and AI a shared definition of done.”
↩︎ Redesigning a workflow in the right order“Delegation means decomposing the problem first, then deciding what AI handles at each step.”
↩︎ Key concept“Delegation means decomposing the problem first, then deciding what AI handles at each step.”
↩︎ Exam trap 3“Delegation means decomposing the problem first, then deciding what AI handles at each step.”
↩︎ Checkpoint - 3.https://academy.claude.com/courses/ai-fluency-for-creative-work/putting-it-all-togetherOfficial docs
“There is no single position on AI usage; it is decided at the level of mode and task.”
↩︎ Delegation is decided step by step“A policy without revision triggers goes stale silently.”
↩︎ Recognizing over-delegation - 4.https://academy.claude.com/courses/ai-fluency-for-creative-work/delegation-and-diligenceOfficial docs
“If a peer challenged a Delegation call, could you defend it?”
↩︎ Delegation is decided step by step“Convergence and skill atrophy are the long-term pitfalls; reaching for AI when you do not know what you want is the immediate one.”
↩︎ Recognizing over-delegation“The starting position is no AI; the analysis determines whether and how you opt in.”
↩︎ Recognizing over-delegation - 5.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5-2Official docs
“Keep your own confirmation step for risky or irreversible actions.”
↩︎ Four criteria that decide the label - 6.
“Delegate implementation freely. Keep empathy, judgment, and shipping in your hands.”
↩︎ Four criteria that decide the label“Delegate implementation freely. Keep empathy, judgment, and shipping in your hands.”
↩︎ Exam trap 1 - 7.
“verify in proportion to the stakes”
↩︎ Four criteria that decide the label - 8.https://academy.claude.com/courses/ai-fluency-for-k-12-educators/ethics-responsible-useOfficial docs
“I will not delegate decisions related to: (e.g., final judgment, relational feedback, high-stakes evaluation)”
↩︎ Four criteria that decide the label“pause, consult a colleague, review district guidance, pilot cautiously”
↩︎ Redesigning a workflow in the right order - 9.
“design a repeatable AI-assisted workflow with appropriate human review”
↩︎ Redesigning a workflow in the right order - 10.https://claude.com/blog/building-effective-human-agent-teamsSecondary source
“Humans often work in the same threads the agents do, but they hold the roles only humans can hold.”
↩︎ Worked example: a refund-dispute workflow - 11.
“You own the outcome, not the output. “AI wrote it” explains nothing and excuses nothing.”
↩︎ Recognizing over-delegation“You own the outcome, not the output. “AI wrote it” explains nothing and excuses nothing.”
↩︎ Exam trap 4“You own the outcome, not the output. “AI wrote it” explains nothing and excuses nothing.”
↩︎ Checkpoint