CCAO-F 6.1 — Appropriate and inappropriate use

The three gates

Ask in this order

1 · Permitted?Does the Usage Policy allow this at all?
2 · Suited?Is this the kind of work Claude does well?
3 · Controlled?Who reviews it, and is AI use disclosed?

Work down the stack. A task must clear all three; the gate it fails at tells you whether to abandon it, reshape it, or simply add a reviewer.

Gate 1 — is it permitted at all?

Anthropic publishes a Usage Policy that applies to everyone using Claude, on every plan. Its first part, the Universal Usage Standards, is a list of things nobody may do with Claude. It is broader than most people expect and worth reading once end to end, but the themes are recognisable: attacking computer systems or critical infrastructure, developing weapons, generating child sexual abuse material, creating fraudulent or deceptive content such as phishing and fake documents, impersonating a real human being, interfering with elections, and promoting self-harm or violence.

Two entries in that list catch ordinary office workers out. The policy prohibits compiling personal data about people without a proper basis, which rules out the “find me everything about this named individual” research request. And it names criminal-justice applications — parole and sentencing decisions, predictive policing, surveillance and emotion recognition — as off limits, which matters to anyone in the public sector.

Gate 2 — the high-risk middle ground

The Usage Policy has a second part, High-Risk Use Case Requirements, that covers uses affecting a person’s rights, safety, livelihood or access to essential services. The named domains are legal advice and decisions with legal consequences, healthcare and mental-health guidance, insurance underwriting and claims, financial decisions such as lending and investment advice, employment decisions including résumé screening, housing eligibility, academic testing and admissions, and automated journalistic or media content.

These are not forbidden. They carry two standing conditions. Human-in-the-loop: a qualified professional reviews the output before it is acted on. Disclosure: where the output reaches a consumer or member of the public, they are told that AI was involved — the policy asks for this at a minimum at the start of each session. A separate guideline applies the same logic to consumer-facing chatbots, which must tell people they are talking to an AI rather than a person.

Which gate did it fail?

The task feels risky. Why?

On the prohibited list
Do not use ClaudeNo configuration makes it allowed
High-risk domain
Add expert reviewQualified reviewer plus disclosure to the person affected
Claude cannot know it
Supply the sourceAttach the document rather than asking from memory
Merely unfamiliar
Proceed and verifyNormal checking, covered in 2.4

The action depends on the gate. Only the first gate ends in “don’t”; the others end in “reshape” or “add a reviewer”.

Gate 3 — is Claude suited to it?

Permission is not fitness. Anthropic’s own help centre is blunt that Claude can hallucinate: it can produce fabricated quotations, and it can be confidently wrong about recent events because its training data has a cutoff. The guidance is to scrutinise high-stakes advice rather than rely on it, and, when Claude has searched the web, to open the cited pages and read them in their original context.

That gives a practical test for fit. Ask where the facts in the answer are going to come from. If they come from a document you attached, a connector you configured, or the text of the conversation itself, Claude is on solid ground — it is reading, summarising, restructuring, drafting. If they have to come from Claude’s memory of the world — a policy number, a price, a regulation, last quarter’s figures, whether a client is still under contract — you are asking for trouble, and the fix is almost always to supply the source rather than to prompt harder.

Task	Where the facts come from	Verdict
Summarise this 40-page supplier contract	The attached document	Good fit
Rewrite this rejection letter in plainer English	The letter you wrote	Good fit
Draft interview questions for this job description	The description, plus general craft	Good fit
What does our refund policy say?	Claude’s memory — unless you attach it	Attach the policy
Which of these 200 applicants should we interview?	A high-risk employment decision	Human decides; disclose
Find everything about this named private individual	Compiling personal data	Not permitted

Worked example · Sifting applications at a regional retailer

A recruitment coordinator at a 900-store retailer has 340 applications for twelve seasonal supervisor roles and four days to produce a shortlist. She asks whether Claude can rank the candidates.

Gate 1. Nothing here is prohibited. Recruiting is ordinary commercial work.
Gate 2. Employment decisions, and résumé screening specifically, are named in the High-Risk Use Case Requirements. So the task is allowed, but only with a qualified human making the decision and with disclosure to candidates that AI was used in the process.
Gate 3. Ranking 340 people against an implicit sense of “good candidate” is exactly where an assistant is weakest — the criteria are unstated, so neither she nor anyone else can check the ranking.
The reshape. She uses Claude to turn the job description into an explicit scoring rubric with five weighted criteria, gets the hiring manager to approve the rubric, and then has Claude extract, for each application, the evidence against each criterion with the candidate’s own wording quoted.
The decision. The hiring manager reads the extracted evidence and picks the shortlist. The careers page carries a line saying AI tools assist in processing applications.

Takeaway: The same task moved from “Claude decides, nobody can audit it” to “Claude extracts evidence against criteria a human set, and a human decides”. That reshaping — not refusal — is what the high-risk requirements are asking for.

Worked example · A claims letter at an insurer

A claims handler has declined a subsidence claim and must write to the policyholder. The decision is made and documented; the letter is what remains.

Wrong shape. “Should we pay this subsidence claim? Here are the details.” This asks Claude to make an insurance coverage decision — squarely inside the high-risk list, and based on facts it cannot verify.
Right shape. “Here is our decision and the three policy clauses it rests on. Draft a letter explaining it to the policyholder in plain English at a reading age of twelve, keeping every clause reference exact.”
The check. The handler confirms each clause reference against the policy document before the letter goes out, because a fabricated clause number is the classic failure mode.

Takeaway: Move Claude from the decision to the explanation. The judgement stays with the qualified person; the drafting, which is what actually took the afternoon, does not.

Reshaping a high-risk request

Asks Claude to decidetext

Here are 340 CVs.
Tell me the best 12
candidates for the
seasonal supervisor
role and rank them.

Asks Claude for evidencetext

Attached: the approved
scoring rubric and one
application.

For each of the five
criteria, quote the
words in the application
that bear on it. If the
application says nothing
about a criterion, write
"no evidence". Do not
score or rank.

Same underlying job, same person doing it. The second version names the decision-maker, supplies the source and asks for evidence rather than a verdict.

Writing it down before you need it

Judgement made case by case does not survive a busy week. Teams that use Claude well keep a short written list with three columns — green, amber, red — and put real examples in each, not abstractions. Green is work that goes ahead unsupervised: internal drafting, summarising documents you already have, reformatting, brainstorming. Amber is work that goes ahead with a named reviewer: anything customer-facing, anything quoting a figure, anything in a high-risk domain. Red is the short list of things nobody does: the Usage Policy prohibitions plus whatever your own sector adds.

A team’s use-case list, checked

Passes: Summarise a supplier contract we holdGreen — source attached, internal use
Passes: Draft the weekly team update from my notesGreen — no external facts asserted
Check: Write the customer-facing outage noticeAmber — comms lead signs off before it posts
Check: Explain a declined claim to a policyholderAmber — high-risk; adjuster decides, clauses checked
Fails: Score job applicants and pick the shortlistReshape — human decides; disclose AI involvement
Missing: Build a profile of a named private individualRed — compiling personal data, not permitted

Each line is a real task somebody wanted to do. The verdict column is what makes the list useful — an abstract policy nobody can apply is worse than three worked examples.

Anthropic does its own work on the other side of this line: it tests policies with outside domain experts, runs classifiers that detect violations in real time, and evaluates models for bias before release. None of that removes your responsibility for the particular task in front of you — the safeguards are built for harm at population scale, not for whether this letter to this policyholder is accurate. Data sensitivity and the privacy questions that sit alongside these gates are covered in 6.2; your own employer’s rules are 6.3.

Traps the wrong answers are built from

Tempting but wrong	Do this instead
Refusing any task that sounds sensitive	Check which gate it fails. Most high-risk tasks are permitted with a qualified reviewer and disclosure.
Treating a permitted task as automatically a good fit	Ask where the facts will come from; if from Claude’s memory, supply the source instead.
Using AI in a high-risk domain without telling the affected person	Disclose AI involvement to consumers, at minimum at the start of each session.
Asking Claude for the verdict in a legal, medical, lending or hiring decision	Ask for structured evidence against criteria a human set, and let the qualified person decide.
Keeping the rules in one person’s head	Write a green/amber/red list with real examples your team recognises.

You should now be able to

Sort a proposed task into prohibited, high-risk-with-conditions, or ordinary use.
Name the two standing conditions the Usage Policy attaches to high-risk uses.
Spot the domains — legal, health, insurance, finance, employment, housing, academic, media — that trigger those conditions.
Reshape a request so a qualified human makes the decision and Claude does the drafting or extraction.
Judge fitness by asking where the facts in the answer will come from.

Appropriate and inappropriate use