Home › Study guides › CCAO-F › Domain 6 › Lesson 6.1

CCAO-F · Domain 6 · 15% of the exam · Lesson 6.1 · 19 min read

Appropriate and inappropriate uses of Claude

How to judge a proposed use of Claude: six questions, three verdicts, when a qualified person must decide, and which uses are off limits whoever reviews them.

Written against objective 6.1 of the official CCAO-F exam guide (Version 1.0, effective July 2026). An independent resource, not affiliated with Anthropic; the practice questions are written from scratch.

6.1.1 Why "can Claude do it?" is the wrong question

You lead the talent-acquisition team at a national home-and-garden retailer with 180 stores. Seasonal hiring starts in six weeks, and a new store-supervisor role drew 500 applications in its first week. Your director asks how the team plans to use Claude this season, so you collect everyone's ideas into one list. Claude would write the job descriptions and summarise the notes from structured interviews, in which every candidate answers the same questions. It would screen and rank the 500 supervisor applicants automatically, draft the rejection emails and answer candidates' questions about the process. One recruiter adds a sixth idea: assess candidates from their social media.

Each item is a use case: one specific job you would give Claude, for one purpose. Given the material, Claude could produce something for every item by the end of the afternoon: a job ad, neat summaries, a ranked spreadsheet, an email, an answer, a verdict on someone's posts. Each would arrive in the same fluent, confident voice. That is what makes the list hard to judge: "can Claude do it?" tells you almost nothing. A use that works is not the same as a use that is right.

The question that sorts the list is a different one: if the output is wrong, what happens to a person, and who answers for it? A clumsy job ad costs an edit. A wrong ranking quietly costs someone a job they were qualified for, and nobody can say why. Anthropic's AI Fluency course calls this skill delegation: deciding what to do yourself, what to do together with AI and what to let AI handle on its own. The goal, the course adds, isn't to automate everything.

6.1.2 Six questions to ask of any proposed use

Here is the question behind this whole objective: why is one use fine and another not, when Claude handles both with the same ease? The answer never lies in the tool. It lies in the use: what the output feeds, whom it touches and who stands behind it. Six questions bring that out.

Memorise the six questions, and note when you ask them: once, about the use itself, before any work starts. They decide whether a task goes to Claude at all, and in what role. How closely to check each output of an approved use is a separate, later question. And you are not adding up points: one worrying answer is enough to change the verdict.

Notice how the two columns part company. The job ad is low on every question, so ordinary care is enough. The ranking raises a concern on all six before anyone has looked at its quality. Even a perfect ranking would be a decision about 500 people that nobody could check, explain or answer for. Accuracy alone never settles whether a use is appropriate.

6.1.3 Three verdicts: ordinary review, a qualified person, or no

The six questions lead to one of three verdicts, and it helps to know them by name. It is tempting to think in two, allowed or banned. Most real proposals land somewhere in between.

Think of a pharmacy. Some medicines sit on open shelves: you pick them up and read the label. Some are kept behind the counter, and a pharmacist must check before you can have them. A few are not sold at all, however politely you ask. They are not sorted by how well they work, but by what happens if they are used wrongly. Uses of Claude sort the same way.

The first verdict covers most everyday work: drafting, summarising, brainstorming, analysing data you are approved to use, and preparing options for someone else to choose from. Claude is at its strongest here, and a mistake is cheap because you read the output before it goes anywhere. The supervisor job ad belongs here. So does the rejection email template, as long as a person decides who receives it.

6.1.4 Decisions about people need a qualified person

Here is the belief that causes the most trouble: if Claude's assessment is accurate, it can stand as the decision. A decision about a person needs more than accuracy. It needs someone who understands the field, can weigh what a summary leaves out and will answer for the outcome. Anthropic's AI Fluency course makes the same point: the most effective uses combine human and AI strengths, with people providing judgment and ethical oversight.

Anthropic's Usage Policy, the rules for everyone who uses Claude, makes this concrete. It names seven high-risk use cases, uses that carry an elevated risk of harm because they touch areas vital to public welfare and social equity. They are legal; healthcare; insurance; finance, such as loan approvals and creditworthiness; employment and housing; academic testing, accreditation and admissions; and media content generated and published automatically. Under employment it names resume screening and hiring tools.

For these uses the policy requires two safeguards. The first is human in the loop, meaning a person sits inside the process at the point of decision. When Claude helps with advice, recommendations or subjective decisions that directly affect people, a qualified professional in that field must review the content or decision before it goes out or is finalised. The policy adds that you or your organisation remain responsible for its accuracy and appropriateness. The second is disclosure: when outputs are presented directly to people, they must be told that AI helped produce them.

In practice, a qualified human in the loop has three things. They have the expertise to judge: a trained recruiter or the hiring manager, not whoever is free. They have the authority to set aside what Claude prepared, and the time and sources to review each case. A pharmacist who initials a stack of prescriptions without reading them is not checking anything. For the interview notes, Claude organises the evidence and the hiring manager decides. The notes are personal data, so they go in only in the form your data policy allows. In the brief below, look at the last two lines: they keep Claude on the evidence and leave the decision where it belongs.

Summarise the attached interview notes for the store-supervisor role, one candidate at a time.
Use the four competencies from our interview scorecard as headings: customer focus, leading a team, stock and cash control, reliability.
Under each heading, give the evidence from the notes in two or three bullet points, quoting the interviewer's own words.
If the notes say nothing about a competency, write "No evidence in notes" rather than inferring anything.
Do not score, rank or compare the candidates, and do not recommend hiring or rejecting anyone.
The hiring manager decides from the full notes and the scorecard. Your summary helps them find the evidence faster; it is not an assessment.

6.1.5 When the answer is no

Some proposals are wrong as they stand, and it pays to recognise them quickly. They come in three kinds. The first is Claude deciding about someone on its own, or triggering a consequential action nobody approves: rejecting everyone below a score, removing a name from a shortlist, sending an offer. The second is anything Anthropic's Usage Policy prohibits. The third is anything your organisation forbids, even where Anthropic's policy would allow it.

The social-media idea runs into both of the last two. Among its prohibitions, the Usage Policy lists using Claude to score or rate people on their trustworthiness or social behaviour without notification or their consent. Judging candidates from their posts, without telling them, is hard to tell apart from that. Posts also reveal what a hiring decision must never rest on, such as religion or ethnicity, and the policy prohibits promoting discrimination on protected attributes like these. And the retailer's own hiring policy forbids social-media screening outright. This use is not reshaped or supervised. It is dropped.

Listen to the reasons people give for going ahead anyway. Each sounds sensible, and none answers the question that matters.

The ranking proposal is off limits as written, because Claude would be deciding. Unlike the social-media idea, though, it serves a legitimate goal: getting through 500 applications fairly and on time. That goal can be met another way.

6.1.6 Judge each step, not the tool

Here is the mistake that often follows a scare: one risky idea on the list, so Claude is banned from hiring altogether. That throws away the job ads, the summaries and the email drafts, which were never the problem. The verdict belongs to each step, not to the tool. Drafting the job ad and ranking the people who answer it sit side by side in the process and at opposite ends of the verdict.

Memorise the pattern rather than the rows: the verdict follows what each step decides, and a doubtful step is reshaped around the part a person must own. Recruiters still decide, and Claude helps them prepare. Whether Claude may go further, say by summarising each application, is not your call alone: it puts 500 people's personal data into a high-risk use.

That is a gray area: a use your policies do not clearly settle. Gray areas go to the person who owns your organisation's AI policy, not to your best guess and not to Claude, which cannot make that call for your organisation or answer for it.

Technical builds have owners too. A chat assistant on the careers site, or a tool that pulls applications from the hiring system in bulk, is an integration: a connection between systems that someone must build and look after. It goes to Claude Architects and Developers. If a candidate chat assistant is approved, the Usage Policy also requires it to tell candidates, at least at the start of each chat, that they are talking to AI. Your part is to spot the handover and say where a person must still decide.

6.1.7 The exam traps

Every trap here treats capability as permission or a surface fix as a safeguard.

• ✗ Judging a use by whether Claude does it well. ✓ Judge it by impact, reversibility, verifiability, accountability, data and policy. A capable tool does not make a use appropriate.
• ✗ Letting Claude screen, approve or reject people because it matched past decisions in a test. ✓ Keep Claude on preparing the evidence and a qualified person on the decision. Accuracy on a sample does not transfer accountability.
• ✗ Calling a bulk sign-off a human in the loop. ✓ The reviewer needs the expertise, the authority to overrule, and the time and sources to judge each case.
• ✗ Adding a reviewer, a disclaimer or anonymisation to a use that is prohibited or forbidden. ✓ Drop the use and meet the goal another way. Safeguards cannot rescue a use that should not exist.
• ✗ Banning Claude from a whole process because one step was risky. ✓ Give each step its own verdict; keep the drafting and summarising, and reshape or drop the risky step.
• ✗ Asking Claude whether a use is allowed, or building the tool yourself. ✓ Take gray areas to your policy owner and technical builds to Claude Architects and Developers.

6.1.8 Put it together: sort a list of proposed uses

You now have the whole method: six questions, three verdicts, a verdict for each step, and an owner for every gray area and build. The quickest way to feel the line between preparing and deciding is to watch Claude cross it when nothing stops it.

The rest of Domain 6 builds on these verdicts. Handling sensitive data (6.2) covers what to remove or minimise before notes or applications reach Claude. Following your organisation's AI policy (6.3) covers the approvals behind every "ask the policy owner" here. The ethics of AI use (6.4) looks harder at bias and fairness in outputs such as screening criteria. And how much checking each output of an approved use gets is the job of review levels (2.4).

Key takeaways

• ✓ Judge a use by what happens to people if Claude's output is wrong, not by whether Claude can produce it.
• ✓ Six questions sort any proposed use: impact, reversibility, verifiability, accountability, data and policy.
• ✓ Drafting, summarising, brainstorming, analysing approved data and preparing options are appropriate with ordinary review.
• ✓ Anything that feeds a decision about someone's job, money, home, health, legal position or education needs a qualified person who reviews each case and decides.
• ✓ Claude deciding on its own, uses the Usage Policy prohibits and uses your organisation forbids are off limits, and no review makes them acceptable.
• ✓ Give each step its own verdict: keep what fits, reshape or drop the rest, and take gray areas to the policy owner and builds to Claude Architects and Developers.

Check your understanding

4 questions written for this lesson, then one from the CCAO-F question bank on the same topic. Every answer option is explained, including the ones you did not pick. Nothing is stored.

Sources

• Usage Policy
• Usage Policy update
• A closer look at Delegation (AI Fluency)
• Capabilities and limitations (AI Fluency)
• The CCAO-F exam guide, summarised