What you will be able to do
- Break a business process into steps and assign each one to Claude, a human, or both working together
- Use reversibility to decide whether a human gate goes before an action or a review comes after it
- Spot designs that look supervised but leave no accountable human owning each decision
- Justify a collaborative split for high-stakes steps and reject confidence thresholds or AI auditors as a substitute for human sign-off
Key concept
Step-level delegation — Break the process into its steps first, then decide for each one whether Claude does it, a human does it, or the two share it. Each decision rests on three questions: can the step's effect be undone, what happens if it is wrong, and who has to answer for the outcome.
1.Assign steps, not whole processes
Delegation is the decision about what Claude handles and what stays with people. You make that decision step by step, never for a whole process at once. A claims workflow, a compliance review or a refund pipeline is a chain of separate steps, and those steps differ in kind. Reading a form and pulling out its fields is one kind of work. Deciding whether the company owes money is another. Moving the money is a third. A design that calls the whole process "automated" or "human-run" has skipped the analysis the decision depends on.
Each step lands in one of three modes. In Automation, Claude does the work and a person checks it. In Augmentation, Claude and a person work the step together. In Agency, Claude operates with latitude inside boundaries a person has set. A fourth outcome, keeping the step entirely with a human, sits beside these three. It is the right answer more often than enthusiasm for automation suggests.
| Mode | Who does what | Fits a step that is… |
|---|---|---|
| Automation | AI does it, you check | High volume, mechanical, and its errors show up in review |
| Augmentation | You and AI work it together | A judgement call where Claude can prepare the evidence and a person makes the decision |
| Agency | AI operates with latitude inside boundaries you set | Easy to reverse, with the boundary enforced outside the model |
| Retained by a human | A person does it | Irreversible, or a decision a named person must answer for |
Agency needs its boundary enforced outside the model because a model can misread or overstep an instruction, and a limit it is merely told about is only as reliable as its reading of it. Anthropic's incident log records an agent deleting remote git branches from a misinterpreted instruction. Anthropic's own guidance says a prompt block telling Claude to leave payments to a person "guides Claude and doesn't enforce anything", so a person must approve each call to a tool that can move money. A limit built into the surrounding system, such as a tool list with no delete action, holds because the model cannot attempt what is not there.
Implementation and judgement split along a clear line. Handing Claude the production of an artefact is generally safe, because the artefact can be inspected before anyone relies on it. Handing Claude the judgement about what should happen is a different act. Claude is versatile and fast at scale, but it also has known failure modes: hallucination, stale knowledge and unreliable complex reasoning. The strongest designs pair Claude's speed and scale with human judgement and oversight, and step-level assignment is how you build that pairing.
Checkpoint 1 of 7· Match them up
A refund pipeline has been split into steps. Match each step to the mode it should be assigned.
Tap a term, then the definition that fits it.
Mechanical, high-volume work whose errors show up in review fits Automation. A judgement call where Claude prepares the evidence fits Augmentation. A step that is easy to undo, with limits enforced outside the model, fits Agency. An irreversible step that a named person must answer for stays with a human.
“Automation (AI does it, you check), Augmentation (you and AI work it together), or Agency (AI operates with latitude inside boundaries you set)”Source: academy.claude.com
2.Reversibility decides where the human gate sits
Whether an action's effects can be undone decides whether a human approves it beforehand or reviews it afterwards. If an action can be rolled back, corrected or recalled, review after the fact works: a mistake gets caught and undone. If it cannot be undone, review after the fact is just a record of the damage. Anthropic's own prompting guidance for agentic work makes this the first thing Claude itself should consider.
Consider the reversibility and potential impact of your actions. You are encouraged to
take local, reversible actions like editing files or running tests, but for actions that
are hard to reverse, affect shared systems, or could be destructive, ask the user before
proceeding.Claude Code's auto mode applies the same rule. File edits inside the project directory run with no approval step, because version control makes them reviewable and reversible. Shell commands, external integrations and operations outside the project go to a classifier first. Anthropic's incident log shows why that line matters. It records an agent deleting remote git branches after misreading an instruction, and another attempting migrations against a production database. In each case the model was overeager and took initiative the user did not intend. None of those mistakes could be fixed by an after-the-fact review.
Some properties feel relevant but do not move the gate. A draft that passes formatting and schema checks is well-formed, which tells you nothing about whether its effect can be undone. High volume is a reason to hand Claude the preparation work, but the action itself stays exactly as reversible as it was. Running on a more capable model tier changes neither the stakes nor whether the effect can be undone. Model choice is a separate decision; see model tier selection.
Checkpoint 2 of 7· Check yourself
Your team's Claude agent sends about 4,000 supplier payment instructions a week to an external bank API. Every instruction passes schema validation, and the agent runs on the most capable model tier. A colleague wants to drop approval before sending and audit a weekly sample afterwards instead. Which question should decide whether that proposal is acceptable?
Review afterwards only works when a mistake can be undone. If a sent payment cannot be recalled, the audit just records the damage, so approval has to come before the send. Validation results, volume and model tier don't change whether the effect can be undone.
“Consider the reversibility and potential impact of your actions.”Source: platform.claude.com
Checkpoint 3 of 7· Exam question
Which factor most directly determines whether a human approval gate should sit before a Claude-initiated action, rather than a review after it has run?
Correct answer: A — Whether the action's effects can be reversed once carried out, for example through a rollback, a correction or a recall.
- A. Correct: the ability to reverse an action's effects decides where the human gate goes. An effect that cannot be rolled back, corrected or recalled needs approval before it runs, while a reversible one can be reviewed afterward and repaired.
- B. Incorrect: passing formatting and schema checks shows the draft is well-formed, not that its effect can be repaired. A valid request can still trigger a harmful, irreversible action, so this does not decide whether approval comes before or after.
- C. Incorrect: volume determines how much reviewer time is needed, but says nothing about whether a mistake can be repaired. A rare irreversible action still needs a gate before it runs, and a frequent reversible one can be reviewed afterward.
- D. Incorrect: a more capable tier lowers the error rate but does not change what an error costs. An action whose effects cannot be undone still needs approval before execution, whichever model tier produced the draft.
3.Accountability attaches to each decision, not to the system
Claude can do the work behind a decision, but it cannot be the one answerable for it. The person who ships or signs remains responsible for the result, whatever part AI played. "AI wrote it" explains nothing and excuses nothing. So when a regulation, contract or licence names a person as accountable for a decision, that person has to make each decision before it takes effect. Supervising the machinery that makes decisions does not meet that requirement.
Several designs look like oversight but fail on this point:
- Owning the log. Claude decides each case and records its reasoning, and a named officer owns the audit trail. The officer is accountable for the records, not for the decisions in them.
- Owning the policy. The officer sets the scoring thresholds and reads a monthly summary. Claude still makes every individual decision, and the officer sees only aggregates.
- Batch countersignature. Claude acts on cases, and the officer signs a weekly batch of actions that have already taken effect. The signature arrives after the decision is already out in the world.
- Sampled audit. People check a sample of completed cases. Unsampled cases never get a human decision at all.
A design that keeps accountability intact while still using Claude for volume looks different. Claude triages the incoming cases, puts aside the clearly routine ones for the accountable person to dispose of, and writes an evidence-based narrative for each case it escalates. The accountable person then makes and signs the decision on each of those cases. Claude takes on the reading and drafting load, and the decision stays with the person the rules name. That signature step is a human-in-the-loop checkpoint. The same ownership principle applies to code in accountability for AI-generated code.
Checkpoint 4 of 7· Check yourself
A lending regulation names the credit officer as accountable for each loan denial. The bank wants Claude to carry most of the review workload. Which design meets the regulation?
Accountability belongs to each decision's outcome, so the named officer has to make each denial before it takes effect. Owning the log, owning the policy, or countersigning denials that have already gone out means supervising the system, not deciding the case.
“You own the outcome, not the output. “AI wrote it” explains nothing and excuses nothing.”Source: academy.claude.com
Checkpoint 5 of 7· Exam question
A bank uses Claude to screen transaction alerts for money-laundering indicators. Regulation requires that a named compliance officer be accountable for each decision to file or not file a suspicious-activity report. Alert volume is high and most alerts are benign. Which design keeps accountability intact while still using Claude for volume?
Correct answer: C — Claude triages alerts and drafts evidence-based narratives; the compliance officer decides and signs the file-or-close outcome on each alert Claude escalates.
- A. Incorrect: Claude has already submitted the reports before the officer sees them, so the officer only countersigns decisions the model made. A named person must make the file-or-not decision, not ratify it afterward.
- B. Incorrect: a retained log records what the model decided and why, but it does not make a named officer the decision-maker. Owning the log's retention is a records duty, not accountability for each file-or-not decision.
- C. Correct: Claude's triage and drafted narratives absorb the volume, while the officer makes and signs the file-or-close call on each escalated alert. A named person therefore stays accountable for those decisions.
- D. Incorrect: owning the scoring policy and reading a monthly summary sets the rules and gives hindsight. Individual file-or-close outcomes are still decided by Claude's scores with no officer deciding them.
4.Collaborative splits for high-stakes steps
When a step is high-stakes, the usual answer is collaboration, not exclusion. Claude does the extraction and the first-pass analysis, and shows its sources so a reviewer can check them. A qualified person reviews that work against the original material, makes the decision, and carries out any irreversible action. Take a supplier-refund process. Claude reads the invoices and drafts a recommendation citing the contract terms. The finance officer checks the draft against the file and approves the transfer. Claude's speed goes into the preparation, and the decision and the transfer stay with the officer.
Confidence gating fails because the cases it trusts are exactly the ones nobody looks at. A second Claude instance that audits the first one fails for a related reason. Pairing a doer agent with a verifier agent is a real and useful pattern. Good long-running agents verify their own work in several ways before a human sees it, and code tests and documentation rubrics serve the same purpose. But the verifier raises the quality of what reaches the human. It cannot be the accountable party, and it should not carry out an irreversible action. Giving the auditing agent less context, such as the policy text without the case file, makes it a weaker check still.
Autonomy can grow over time, one task type at a time. Anthropic teams give agents autonomy in proportion to the reliability they have shown. They review agent work by hand at first, track which kinds of task each agent has earned autonomy on, and expand scope one task type at a time. In one backlog-triage effort, people first reviewed every agent decision. They then taught the agents to send decisions needing human input straight to a person, so that decisions with hard tradeoffs always had a human in the loop. Trust grows on the reversible, verifiable tasks. The hard-tradeoff decisions stay with people however reliable the agent becomes.
Extraction goes to Claude (Automation), because errors show up when the extracted data is checked against the source. For the judgement, Claude drafts an assessment citing the relevant rules and a qualified person reviews it against the full file and decides (Augmentation), because a named person is accountable for that decision. The payment stays with that person and happens only after they decide, because a transfer that cannot be recalled needs a gate before it, not an audit after it. AI verifiers, confidence thresholds and sampled audits may help catch errors, but none of them replaces that gate.
Checkpoint 6 of 7· Check yourself
An architect wants to speed up supplier refunds with a doer–verifier pair. One Claude agent drafts each refund decision from the invoices, a second Claude agent audits the draft, and any transfer the verifier approves is sent automatically. The finance officer is accountable for each refund. How should this design be judged?
A verifier agent raises the quality of what reaches the human, but it cannot be the accountable party or carry out an irreversible transfer. The officer still has to decide and approve each refund. Giving the verifier less context weakens the check, and confidence gating leaves the confident cases with no human decision.
“we’ve found that the best long-running agents have many different ways to verify their work before a human looks at it.”Source: claude.com
Checkpoint 7 of 7· Exam question
A property insurer wants Claude (via the Messages API) in its claims process. The process has three steps: extract fields from submitted claim forms and photos, compare the claim with the policy wording to judge coverage, and release the payout. Payouts go out by bank transfer that cannot be recalled, and the insurer remains legally accountable for every coverage decision. Which assignment of the three steps is most defensible?
Correct answer: B — Claude extracts fields and drafts the coverage assessment, citing policy clauses; a licensed adjuster reviews each draft against the claim file and releases the payout.
- A. Incorrect: a second Claude instance shares the first one's failure modes and cannot carry the insurer's legal accountability for coverage decisions. The irreversible bank transfer would be released with no human owning the decision.
- B. Correct: extraction and clause-cited drafting are reversible, high-volume work that suits Claude. The unrecallable payout and the legal accountability for the coverage decision rest with a licensed adjuster who checks the draft against the claim file before release.
- C. Incorrect: adjusters only look at paid claims after the money has gone out, and a bank transfer cannot be recalled, so later fixes cannot undo a wrong payout. No person approves a coverage decision before the insurer becomes bound by it.
- D. Incorrect: a confidence score reported by the model is not a dependable control for an irreversible payment. Wrong but confident decisions would be paid with no human review, leaving no accountable person for them.
Sources
(8)
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.”
↩︎ Assign steps, not whole processes“Diligence means owning the final result fully, regardless of AI’s role.”
↩︎ Accountability attaches to each decision, not to the system - 2.https://academy.claude.com/courses/ai-fluency-for-builders/delegation-the-builder-s-toolkitOfficial docs
“Automation (AI does it, you check), Augmentation (you and AI work it together), or Agency (AI operates with latitude inside boundaries you set)”
↩︎ Assign steps, not whole processes“Explain why delegating implementation is safe and delegating judgment is not”
↩︎ Assign steps, not whole processes“Delegation means decomposing the problem first, then deciding what AI handles at each step.”
↩︎ Key concept - 3.https://academy.claude.com/courses/ai-fluency-for-builders/ai-capabilities-and-limitationsOfficial docs
“Current limits include knowledge cutoffs, hallucinations, and unreliable complex reasoning.”
↩︎ Assign steps, not whole processes“The best applications pair your judgment, creativity, and oversight with AI’s speed and scale.”
↩︎ Collaborative splits for high-stakes steps“Current limits include knowledge cutoffs, hallucinations, and unreliable complex reasoning.”
↩︎ Exam trap 1“Current limits include knowledge cutoffs, hallucinations, and unreliable complex reasoning.”
↩︎ Prediction - 4.https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practicesOfficial docs
“The block guides Claude and doesn't enforce anything, so have a person approve each call to any tool that can move money.”
↩︎ Assign steps, not whole processes“might take actions that are difficult to reverse or affect shared systems, such as deleting files, force-pushing, or posting to external services”
↩︎ Reversibility decides where the human gate sits“Consider the reversibility and potential impact of your actions.”
↩︎ Exam trap 3“Consider the reversibility and potential impact of your actions.”
↩︎ Checkpoint - 5.https://www.anthropic.com/engineering/claude-code-auto-modeSecondary source
“deleting remote git branches from a misinterpreted instruction”
↩︎ Assign steps, not whole processes“in-project edits are reviewable via version control.”
↩︎ Reversibility decides where the human gate sits“Each of these was the result of the model being overeager, taking initiative in a way the user didn't intend.”
↩︎ Reversibility decides where the human gate sits - 6.https://claude.com/blog/ciso-guide-to-agentic-aiSecondary source
“It will never attempt an action that isn't in its tool list.”
↩︎ Assign steps, not whole processes - 7.
“You own the outcome, not the output.”
↩︎ Accountability attaches to each decision, not to the system“You own the outcome, not the output.”
↩︎ Exam trap 4“You own the outcome, not the output. “AI wrote it” explains nothing and excuses nothing.”
↩︎ Checkpoint - 8.https://claude.com/blog/building-effective-human-agent-teamsSecondary source
“they hold the roles only humans can hold”
↩︎ Accountability attaches to each decision, not to the system“the best long-running agents have many different ways to verify their work before a human looks at it”
↩︎ Collaborative splits for high-stakes steps“Tracking which kinds of tasks each agent has earned autonomy on and expanding scope per task type after repeated successes”
↩︎ Collaborative splits for high-stakes steps“decisions with hard tradeoffs always had a human in the loop”
↩︎ Collaborative splits for high-stakes steps“the best long-running agents have many different ways to verify their work before a human looks at it”
↩︎ Exam trap 2“we’ve found that the best long-running agents have many different ways to verify their work before a human looks at it.”
↩︎ Checkpoint