What You Need to Know
Two questions decide whether a task goes to Claude. Fit asks whether Claude would do the work well: the right shape, the right volume, output someone can check. Domain 4 covers fit. Appropriateness asks whether the task should be handed over at all. It looks at different evidence: the stakes, the authority the work requires, and who answers for the outcome. A task can be a perfect practical fit and still belong to nobody but a qualified human.
Fit Is Not Permission
Take a bike shop owner with two jobs on the desk. One is a long run of customer reviews, to be summarised into themes. The other is a stack of applications for a mechanic's job, to be scored to pick the hire. Scoring the applications is an excellent practical fit: structured, voluminous, checkable. It can still be inappropriate, because of what the scoring decides and who cannot challenge it. The recurring wrong answer reasons from quality: Claude would do the task quickly and well, so it must be fine to hand over. Quality was never the question.
The Four Inappropriate Categories
Four recognisable categories make a use inappropriate.
The first is work that needs professional authority: legal, medical, financial and safeguarding judgements. A qualified person must make these and be accountable for them. Claude can help someone prepare, summarise or understand. It cannot be the source of the advice. At the bike shop, medical advice for an injured customer must come from a qualified person. The Usage Policy names the settings where this applies most. Its High-risk Areas include legal and medical advice, finance, credit, insurance, housing, employment, education and healthcare access. In these areas a qualified person must review Claude's recommendation before it is given as advice or used to make a decision, and "The qualified person remains responsible for the accuracy and appropriateness of what is delivered or decided." The policy excludes wellness content, such as education on sleep, stress, nutrition or exercise.
The second is a consequential decision about identifiable people: who gets the job, the award, the sanction, the visit. Poor output is not the reason. Nobody is accountable for the decision, and the person affected cannot challenge it. The mechanic scoring sits here.
The third is fabrication presented as real: invented quotes, testimonials, case studies or evidence, such as customer testimonials the shop writes itself. Plausibility makes fabrication worse, because a convincing invention is easier to believe. The Usage Policy's prohibitions on misinformation and deceptive practices sit underneath this category.
The fourth is anything the Usage Policy or the organisation's own policy prohibits. Objective 6.3 covers this ground.
The Substitution Test
One test sorts a task. Picture a capable, well-meaning colleague who has:
- no professional qualification in the area
- no access to the person's history
- no accountability for the outcome
Would you accept this piece of work from them? The test turns an abstract appropriateness question into an ordinary delegation judgement that every professional already makes. The review summary passes: the same colleague could write it, and the owner would check it. The mechanic scoring fails on all three grounds at once. Anthropic's AI Fluency framework names this skill Delegation.
Most Work Is Appropriate
The counterweight matters as much as the categories. Official documentation describes a deliberately broad appropriate range: writing across formats, summarising long documents, brainstorming, analysing images and data, translating. Drafting, reformatting, analysing material you hold, and preparing someone to make a decision are all appropriate. The inappropriate categories are narrow and specific. Treating everything near them as forbidden abandons work a simple safeguard would enable. The owner can still have each application summarised, then read them and choose.
In the exam
This objective asks a different question from the rest of the certification, and candidates arriving from Domain 4 reliably answer the wrong one. The substitution test covers the exam's whole surface. The exam punishes over-caution too: a keyed answer almost always makes the work possible within the rules.
The exam tell
Adding review to an inappropriate delegation does not repair it. A manager checking the output of a consequential decision about identifiable people restores neither accountability for the judgement nor the affected person's ability to challenge it. The owner reading every applicant's score changes neither. Options offering review as the fix are testing exactly this.
The Usage Policy's review rule answers a different question. Where a use in a High-risk Area goes ahead, the policy requires that "A qualified person must meaningfully review a High-risk AI Recommendation, with the authority to change or modify it, before it is delivered as advice". There, review is a mandatory control, and objective 2.4 teaches it as the standard. Review cannot convert a decision that should not have been delegated into one that should. Ask whether the delegation is appropriate first. Only then does the review requirement apply.
The AI Fluency framework lists the four Ds as module names and does not define them. Reading Delegation as "deciding what to hand to AI and what to keep" is our gloss. The Help Center's use-case article lists what Claude is for and carries no prohibitions. The inappropriate categories rest on the Usage Policy and the exam guide.
Exam Traps
Reading practical fit as permission to proceed
Fit and appropriateness are separate tests on separate evidence. A task can be structured, voluminous and checkable, everything Domain 4 wants, and still be inappropriate because appropriateness turns on accountability and challenge, not on how well Claude would perform.
Accepting the internal-use excuse
That the results never leave the organisation appears as a distractor across this domain and is never keyed. Keeping an output in-house restores neither accountability for the decision it embeds nor the affected person's route to challenge it.
Declining both halves of a mixed proposal
Abstinence dressed as caution is the plausible second-best. When a proposal mixes appropriate and inappropriate work, the keyed answer keeps the appropriate half and declines only the inappropriate one; walking away from both wastes the tool and misreads the domain.
Practice Scenario
A children's centre manager is asked to use Claude on the annual family survey. Two proposals arrive together: summarise three hundred free-text responses into themes for the service review, and score individual families' responses to flag which should receive a home visit. The service review needs the themes. Centre policy allows Claude on the survey text and bars any tool from scoring families for home visits. What should the manager do?
Build Exercise
Sort a Fortnight of Real Work
What you'll learn
- How to run the substitution test on real tasks
- How to name which of the four categories makes a task inappropriate
- How to write the enabling safeguard for a borderline task
- How to check your own judgement against the over-caution failure
- List the last ten pieces of work you gave Claude, or considered giving it, plus two you would never hand over
Why: The test needs real delegations to bite on; invented examples flatter the exercise and hide your actual habits.
You should see: A list of twelve tasks in your own words, each one line long.
- Run the substitution test on each: would you accept it from a capable colleague with no qualification in the area, no access to the people involved, and no accountability for the outcome?
Why: The test converts appropriateness into a delegation judgement you already know how to make.
You should see: A pass or fail against each task, with the failing ground named: qualification, history or accountability.
- Sort the list into appropriate and inappropriate, and for each inappropriate item name which of the four categories it falls under
Why: Naming the category is the exam skill: items reward candidates who can say why a use is inappropriate, not just that it feels risky.
You should see: Each inappropriate task labelled with one category: professional authority, consequential decision, fabrication, or policy prohibition.
- For any borderline item, write the enabling safeguard that would let it proceed: what changes, the material, the route, or the authorisation
Why: Keyed answers almost always make the work possible; the professional skill is finding the safeguard, not finding reasons to refuse.
You should see: A one-line safeguard per borderline task, specific enough to act on tomorrow.
- Check the appropriate pile is the larger one; if it is not, revisit your test results with the counterweight in mind
Why: Over-caution is the failure the distractors in this objective are built from; a mostly-inappropriate list usually means the test was run too harshly.
You should see: An appropriate pile of eight or more of the twelve tasks, or a written reason why your work is genuinely exceptional.
Sources
- Usage Policy — Anthropic
- What are some things I can use Claude for? — Anthropic
- AI Fluency: Framework and Foundations — Anthropic
- Claude Certified Associate – Foundations Exam Guide — Anthropic. The Help Center article describes appropriate uses but lists no prohibitions. The inappropriate categories come from the Usage Policy and the exam guide's objective.