Governance, Risk, and Responsible Use
15% of the Claude Certified Associate – Foundations blueprint — roughly 9 of the 60 questions on a real sitting.
Governance, Risk, and Responsible Use is 15% of the CCAO-F blueprint, roughly 9 of the 60 questions. It is the domain candidates most often expect to be a values quiz and most often find is a judgement test. The questions do not ask whether AI is good. They put a competent professional in front of a specific task — ranking sixty staff for redundancy selection, reducing 220 applications before Friday's panel, drafting the narrative section of an assessment that determines someone's care package, translating a gas safety notice before an engineering visit — and ask what the central problem is, or which of several permitted uses to pilot first.
One distinction runs through nearly all of it: the difference between output that informs a decision and output that becomes the decision. Claude producing a comparable summary of forty funding applications against published criteria is ordinary, defensible work. Claude producing a ranked list that is adopted as the provisional selection is not, because the organisation can no longer show how any individual outcome was reached. Around that distinction sit the supporting ideas the exam tests: whether a human review is real or nominal, whose information you are putting into a prompt, whether the output misleads anyone about who is speaking, and what your organisation's own policies and client commitments actually allow.
What the exam actually tests
- —The line between output that informs a decision and output that is treated as the decision
- —Whether a stated human review is meaningful or a signature on reasoning nobody can reconstruct
- —Judging appropriateness by consequence to the person affected, not by task difficulty
- —Handling personal and confidential information according to what your organisation permits
- —Uses that mislead about authorship, authenticity or who is speaking
- —Applying the organisation's own AI policy and client commitments as the operative constraint
Input to a decision, not the decision
Decisions that determine whether someone keeps a job, receives funding, is admitted to a course or has a complaint upheld must be traceable to reasoning a person actually did and can explain. Generated output can support that work — summarising each application against the published criteria so the panel compares like with like, drafting the letter once the outcome is settled, suggesting risks a team may not have considered. It cannot be the outcome itself. The practical test is what the organisation would say if challenged on one particular case: "a person assessed it against these criteria and here is their reasoning" is defensible; "it came out of the ranking" is not, however sound the criteria were.
What makes a review meaningful
Scenarios frequently offer a human in the loop as though the phrase settles the question: a tutor confirms the mark, a social worker signs the assessment, a manager approves the letter. Whether that is a control depends on conditions. The reviewer needs the underlying material, not just the output; time to disagree; the competence to judge; and a genuine expectation that disagreeing is normal. Someone handed forty polished drafts an hour before a deadline is a signature, not a safeguard. The tell is whether the reviewer could realistically detect the specific error the workflow is exposed to.
Consequence to the person, not difficulty of the task
The risk in a task is not proportional to how hard it is. Triaging forty pieces of correspondence into routine categories is easy and low-stakes because a misfiled item is noticed and re-routed. Translating a gas safety notice is also routine, and carries serious risk because a resident acting on a mistranslated instruction can be harmed and nobody downstream is positioned to catch it. When choosing among permitted uses, weigh what happens to the affected person when the output is wrong, and whether anyone would notice before it mattered. Reversibility and detectability matter more than complexity.
Information you put in, and whose it is
Prompts frequently contain other people's information — case notes, exit interviews, applications, correspondence, student work. The governing questions are what your organisation's policy permits, what it has told those people it does with their information, and what its agreements with clients require. That is the frame the exam uses: the operative constraint is the organisation's own policy and commitments applied to this task, not an appeal to what is technically possible. Note also that consent belongs to the person concerned. A beneficiary story drafted from case notes needs that beneficiary's agreement to publication, and drafting it first does not make the asking optional.
Authenticity and who appears to be speaking
Some uses are wrong not because of accuracy or confidentiality but because they misrepresent authorship. Generating twelve differently worded responses to a consultation for submission under different people's names manufactures the appearance of independent support, and the fault is unaffected by the responses being well argued or by the views being sincerely held. The same applies to invented testimonials, reviews attributed to customers who did not write them, and quotations attributed to people who did not say them. The question is whether the reader would be misled about who is speaking.
Where candidates go wrong
Trap 1 — 'There is a human in the loop, so the control is in place'
This is the most heavily tested wrong belief in the domain, and it is attractive because it describes a real safeguard. The words "a social worker signs it" or "a tutor confirms the mark" do not, by themselves, establish that anyone can catch what is wrong. If the narrative justifying a care package was generated and the reviewer has only the polished text rather than the underlying assessment, the signature transfers responsibility without transferring the ability to exercise it. Ask what the reviewer would need to detect an error, and whether they have it, the time to use it, and any real expectation of pushing back.
Trap 2 — 'Anonymise it and the concern is resolved'
Removing names is a genuine protective measure and candidates reach for it as a universal fix. It answers a confidentiality question; it does not answer an appropriateness question. Scoring sixty employees for redundancy selection is not made acceptable by stripping the names, because the problem is that an employment decision would rest on generated output rather than assessment the organisation can defend — and the individuals are re-identified the moment the ranking is used. When an option offers anonymisation for a scenario whose real problem is accountability, consent or the nature of the decision, it is treating the wrong issue.
Trap 3 — 'The policy permits it, so it is the right thing to do'
Several questions state explicitly that all the options under consideration are permitted, which is the exam removing permission as the deciding factor. Candidates who stop at compliance then have nothing left to choose with. Policy sets the boundary of what you may do; judgement decides what you should do first. Among four permitted pilots, the sound choice is the one where an error is visible and correctable and nobody is harmed while you learn. Being allowed to do something with a person's care assessment does not make it a sensible first deployment.
Trap 4 — 'Claude produced it, so the error is not really mine'
Rarely stated so plainly, but it underlies distractors that respond to a bad outcome by adding a disclaimer, noting that the draft was AI-generated, or proposing to tell the recipient the tool made a mistake. Accountability does not move to the tool. If your name is on the variance commentary, the tender response or the customer letter, you are answerable for it as you would be for anything you had written yourself, and a note about how it was produced changes nothing about that. Disclosure is sometimes appropriate for honesty about provenance. It is never a substitute for the review that should have preceded sending.
How to study this domain
Practise a single question until it is instinctive: is this output informing a decision a person will make, or is it becoming the decision? That one test resolves the majority of this domain, including the scenarios about redundancy selection, shortlisting, marking, complaint outcomes and care packages. Then add the follow-up: if a particular case were challenged, what would the organisation be able to show about how that outcome was reached?
Second, learn to interrogate a claimed human review rather than accepting it. For each scenario that mentions one, ask what error the reviewer would have to catch and whether they realistically could. Third, read your own organisation's AI policy — the exam frames obligations through organisational policy and client commitments rather than through named legislation, so the useful preparation is understanding how such a policy is structured and applied to a specific task, not memorising statutes. Finally, when a question offers several permitted options, stop looking for the impermissible one; it is asking which is the wisest place to start, and the answer is where mistakes are visible, correctable and harmless to the person affected.
Common questions
How many CCAO-F questions come from Governance, Risk, and Responsible Use?
It is 15% of the blueprint, roughly 9 of the 60 questions — the third-largest domain. It is examined as applied judgement about specific workplace tasks rather than as ethical theory or as knowledge of regulation.
Does the exam test data protection law or AI regulation?
It is not a law exam and does not turn on reciting statutes. Obligations are framed the way they reach a working professional: what your organisation's AI policy permits, what it has told people it does with their information, and what its agreements with clients require. Apply the policy the scenario gives you.
Isn't having a person sign off enough to make a use acceptable?
Only if the review is real. A meaningful reviewer has the underlying material rather than just the polished output, enough time to disagree, the competence to judge, and an expectation that disagreement is normal. A signature on reasoning nobody can reconstruct is the most commonly examined false safeguard in this domain.
Where exactly is the line on decisions about people?
Between informing and deciding. Summarising each application against published criteria so a panel can compare fairly is sound. Adopting a generated ranking as the provisional outcome is not, because the organisation can no longer explain how any individual case was decided. Ask what you could show if one person challenged their result.
If all the options in a question are allowed by policy, how do I choose?
By risk, not permission. Prefer the use where an error is quickly visible, cheaply corrected, and does not harm the person affected while the team is still learning. High-volume, low-consequence work is the right place to start; decisions about someone's job, care or place on a course are not.
Practise this domain
The CCAO-F bank is weighted to the blueprint above, so 15% of what you practise is this domain — and every option carries a written explanation, not just the correct one.
See CCAO-FOther CCAO-F domains
Not affiliated with or endorsed by Anthropic. Domain names and weightings are taken from the published exam guide; always check the official guide before booking.