CCAO-F 2.4 — When a human must review

Who needs to see this first?

What will this output be used for?

Just me; easy to undo
Your own checkaccuracy and completeness pass
Others will rely on it
Verify key factssources checked; second reader
Affects a person’s rights, money, health
Expert sign-offqualified reviewer, before release
Claude will take an action
Approve before it runsespecially if irreversible

Where the output goes and who it affects set the level of review — not how good the draft looks.

What raises the bar

Every output gets the basic evaluation from 2.1. The question here is when that is not enough — when you need additional verification against sources (the techniques in 2.3), or a second person, or a person with specific expertise. A handful of factors decide it, and any one of them can raise the bar on its own.

Factor	Lower bar	Higher bar
Audience	Only you, or your immediate team	Customers, regulators, the public, the board
Reversibility	A draft you can revise tomorrow	A sent email, a signed contract, a payment, a deletion
Who is affected	No one in particular	Specific individuals — their job, credit, care, claim
Domain	Brainstorming, formatting, internal notes	Legal, medical, financial, employment, insurance, housing
Where the content came from	Your own documents, quoted	Claude’s general knowledge, recent events, uncited claims
What your checks found	Nothing unusual	Any hallucination, inconsistency or bias signal (2.2)
Volume	One output, read in full	Hundreds of outputs no one reads individually

The last row is easy to miss. Reviewing one email is simple; a process that generates hundreds of customer replies or candidate assessments needs its own review design, because no single person sees every output. The more automated the flow, the more deliberately the human checkpoints have to be placed.

Levels of oversight

More oversight further down

Self-checkbrief met, claims plausible, nothing missing
Additional verificationkey claims traced to primary sources
Second reviewera colleague reads before it goes out
Qualified professionallegal, clinical, financial or HR sign-off
Human decidesClaude informs; a person makes the call

Move down the stack as the stakes rise. The bottom level is not more checking — it is keeping the decision itself with a person.

Where the rules set the bar for you

In some areas the decision is not yours to make. Anthropic’s Usage Policy lists high-risk use cases — including legal, healthcare, insurance, financial, employment and housing, academic testing, and media uses — where outputs that give advice or recommendations, or feed subjective decisions directly affecting individuals, require two safeguards. First, human-in-the-loop: a qualified professional in that field reviews the content or decision before it is disseminated or finalised. Second, disclosure: people are told that AI was used to help produce it.

Data protection law can reach the same place. Article 22 of the EU’s GDPR gives people the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significant effects on them, with limited exceptions; where such decisions are allowed, safeguards include the right to obtain human intervention and to contest the decision. Anthropic’s own research on discrimination in model decisions makes the same point from the other side: it does not endorse using language models to make automated decisions in high-risk areas such as lending and hiring. Your organisation’s AI policy may add more — following it is covered in 6.3.

Worked example · Two letters from the same clinic

Sam, the practice manager at an outpatient clinic, uses Claude for two drafts in one afternoon: an FAQ about the clinic’s new online booking system, and template letters explaining abnormal blood test results to patients.

The booking FAQ. It is operational, easy to correct, and affects no one’s care. Sam checks it against the booking system’s actual steps, asks a receptionist to try it, and publishes it.
The results letters. They explain individual medical results and suggest next steps. That is healthcare advice to individuals — squarely in the Usage Policy’s high-risk category.
Who reviews. Not Sam, however careful: the reviewer must be qualified in the field. A clinician reviews the template wording, and each individual letter is checked by the responsible clinician before it is sent.
Disclosure. The clinic notes that AI assisted in drafting, in line with the policy and with its own governance rules.

Takeaway: Same tool, same afternoon, two very different levels of review. What changed was the audience, the effect on individuals and the domain — not the quality of the draft.

Review that actually counts

A review is only as good as the conditions it happens under. A busy manager clicking “approve” on forty AI-drafted letters in ten minutes has not reviewed them in any meaningful sense. Anthropic’s documentation is clear that techniques for reducing hallucinations lower the error rate but do not eliminate it, so a reviewer has to be positioned to catch what remains.

Is this review meaningful?

Passes: Reviewer is qualified for the subjectclinician for medical content; lawyer for legal
Passes: Happens before the output is sent or finalnot after complaints arrive
Check: Reviewer has the sources, not just the draftwithout them they can only judge the prose
Fails: Enough time for the volumeforty letters in ten minutes is not review
Passes: Reviewer can reject or change it
Missing: What was checked is recordedneeded if the decision is questioned later

A review that fails any of these is closer to a formality than a safeguard.

When Claude acts, not just writes

Some Claude products can take actions: Cowork, for example, can work with your files, browser and apps, and depending on the permissions you grant it can delete files, send messages or make purchases. Anthropic’s safety guidance for Cowork names two risks: destructive actions, and prompt injection, where instructions hidden in content Claude reads — a web page, a document, an email — try to redirect it. It recommends limiting access to what the task needs, monitoring tasks, using manual approval for high-stakes work, and remembering that you remain responsible for what Claude does on your behalf.

Worked example · Supplier bank details from an inbox

An accounts payable lead asks Cowork to go through a shared inbox, find suppliers who have sent new bank details, and update the supplier spreadsheet used for the next payment run.

Spot the combination. The task reads untrusted content (emails from outside the company) and makes a consequential change (payment details). That pairing is exactly where prompt injection and fraud risk sit.
Narrow the access. Cowork is given only the inbox folder and the spreadsheet, not the finance system.
Approve changes manually. Instead of letting it write directly, she has Claude produce a list of proposed changes, each with the email it came from.
Verify outside the chain. Each change is confirmed by calling the supplier on a number already on file — not one from the email — before the spreadsheet is updated.

Takeaway: The output was a list of edits, and the review had to happen before any of them took effect. When Claude acts on your behalf, approval comes before the action, not after.

Traps the wrong answers are built from

Tempting but wrong	Do this instead
Adding an AI disclaimer instead of reviewing	Disclose and review; disclosure does not make content correct.
Reviewing after the output has gone out	Review before it is sent, published or acted on.
Any available colleague as the reviewer	Where it affects rights, health, money or legal position, use a qualified professional.
The same review for every output	Scale oversight to audience, reversibility, who is affected and domain.
Letting Claude make irreversible changes unattended	Limit access, use manual approval, and verify through an independent channel.

You should now be able to

Identify the factors — audience, reversibility, who is affected, domain, provenance, volume — that raise the level of review.
Choose between self-check, additional verification, second reviewer, qualified sign-off and human decision.
Recognise the Usage Policy’s high-risk categories and its human-in-the-loop and disclosure safeguards.
Explain why solely automated decisions about individuals need meaningful human involvement.
Tell a meaningful review from a formality.
Place approval before Claude takes consequential actions.

When a human must review