Determine when human review or additional verification is required
Decide when a Claude output must be reviewed by a qualified person or verified further before it is used, based on consequence, reversibility, audience, regulation and whether you can verify it yourself.
Key points
- 1
Human review is required when an output is external (customers, regulators, auditors, patients, the public), financial (payments, payouts, filings, reconciliations), legal (contracts, termination letters, statutory claims), medical, regulated, or irreversible (a sent email, a published description, a payroll run).
- 2
A separate trigger: anything you cannot personally verify. If the content requires expertise you do not have (clinical, legal, tax), a qualified person must review it regardless of how correct it reads to you.
- 3
These triggers are independent. Verifying every figure yourself satisfies the "can I verify it" test but does not waive review for an external, regulated record. An intercompany memo going to the auditors still goes through the controller sign-off after the analyst has checked the numbers.
- 4
Review must be by a person with the right competence and accountability: a clinician for patient instructions, legal for a termination letter or a statutory citation, a controller for a filing. The person who prompted is usually not that reviewer.
- 5
Low-risk outputs that do not need formal review: brainstorms the team will discuss anyway, summaries for people who were in the meeting, reformatting existing content, rewording your own message, notes for your own use. The audience can correct errors and the consequences are negligible.
- 6
Rephrasing is not risk-free. Turning a physician's bullets into plain language, or a supplier data sheet into marketing copy, can change a dose, a timing, an exception, or strengthen "supports" into "boosts". A correct source does not make the rewrite correct.
- 7
When full review of every output is infeasible, tier by risk: mandatory full review for regulated categories and anything carrying a health, safety, performance or compliance claim; sampled review for the rest; raise the sample rate when errors are found and lower it as quality proves stable.
- 8
Model self-review (a second pass by Claude checking the first) is a useful filter that can catch obvious errors, but it is not accountable review and cannot replace the human tier for high-consequence outputs.
- 9
A second run in a fresh chat is not a second reviewer. Agreement between two model outputs carries no accountability.
- 10
Disclaimers ("generated with AI assistance") and footers telling the reader to confirm with a professional do not substitute for review before distribution; they shift the burden to the audience.
- 11
Anthropic's own guidance is that users should not rely on Claude as a singular source of truth and should carefully scrutinize any high-stakes advice. The exam guide frames this as recognizing when human expertise, validation or escalation is required.
- 12
Escalation is also a competence decision: the Associate escalates complex or technical implementations to Claude Architects and Developers, and domain judgements to the relevant domain expert.
Read the source
Test yourself on Determine when human review or additional verification is required
Ten questions, with the answer and explanation after each one.