CCAO-F Sample Questions: 10 Scenarios From All 7 Domains

CCAO-F Sample Questions With Explanations: 10 Scenarios From All 7 Domains

Anthropic's Claude Certified Associate – Foundations exam opened for scheduling in July 2026. It is the only credential in the Claude Certification Program aimed at people who don't write code — operations, marketing, project management, education, communications, and general knowledge work.

Below are ten original scenario questions, one set per domain, weighted to match the official blueprint. Every question includes the reasoning for the correct answer and for each distractor, because on a scenario exam the distractors are where the marks are lost.

But before the questions, there is something you should know that changes whether you should study at all this month.

First: check whether you can actually sit the exam

Registration for every Claude certification runs through the Anthropic Partner Academy, and the Academy is open to organisations in the Claude Partner Network. Pearson's own program page states that certification "is open to organizations in the Claude Partner Network" and that preparation training "is available to members of the Claude Partner Network." You register at the Academy first; Pearson scheduling is the second step, not the first.

In practice that means:

• If your employer is a Claude Partner Network member, you go through your organisation's Academy access.
• If it isn't, you cannot currently book a seat by walking up to Pearson with a credit card, the way you can with AWS or CompTIA.

Most CCAO-F guides published this summer skip this entirely and open with the $99 fee, which is accurate but incomplete — the fee is only reachable after the access question is settled. Sort that out before you build a study plan around a date you can't book.

Two related details worth knowing while you're checking:

• CCAO-F does not count toward partner tier standing. The Developer and Architect exams do. If your organisation is sponsoring you specifically to move the partner needle, Associate is not the exam that does it — which is an argument for taking it on your own initiative, not an argument against taking it.
• The credential is valid for 12 months. Renewal is a free, non-proctored assessment on the Partner Academy if you renew on time. Let it lapse and you pay the full fee and sit the full proctored exam again. That renewal path also runs through Academy access, so it is worth confirming you'll still have that access in a year.

Exam facts, from Exam Guide v1.0

Two things to take from that table.

120 minutes for 60 items is two full minutes per question. That is generous by certification standards — AWS gives you about 1.4 minutes per item on the AI Practitioner exam (65 items in 90 minutes), and CompTIA's SecAI+ gives you one minute flat (up to 60 items in 60 minutes). The CCAO-F is not a speed test. It is a reading test: the scenarios are long, and the marks go to people who read the whole stem before looking at the options.

720 out of 1,000 is not 72%. Scaled scoring means the raw-to-scaled conversion is set by item difficulty, not by a fixed percentage. Nobody outside Anthropic can tell you how many of the 60 you need to get right. Anyone who tells you "you need 43 correct" is guessing. Aim comfortably above the line rather than trying to hit a number.

The blueprint, and where the marks actually are

The instinct is to over-prepare Domain 1. Prompting is the thing everybody thinks a Claude exam is about, and it is the smallest technical domain on the paper at 14%.

Domain 2 is the heaviest at 21% — evaluating and validating what Claude produces. Add Domain 6 at 15% and Domain 7 at 10%, and nearly half the exam (46%) is about judging, governing, and repairing output rather than generating it. That is a deliberate signal about what Anthropic thinks a competent Claude user does all day. Weight your study the same way.

The ten questions below follow the blueprint split: one from Domain 1, two from Domain 2, one from Domain 3, two from Domain 4, one from Domain 5, two from Domain 6, one from Domain 7.

Question 1 — Domain 1: Prompting and Task Execution

A programme manager needs a 15-page vendor contract turned into a summary for the executive team. Her first prompt is: "Summarise this contract and tell me if we should sign it." The output is a fluent summary that ends with a confident recommendation to sign, but it doesn't mention the auto-renewal clause she is worried about.

What is the best next step?

• A. Re-run the same prompt with a more capable model.
• B. Decompose the request: ask first for a clause-by-clause inventory with the contract's own language quoted, then ask for a risk assessment of the specific clauses she names.
• C. Add "be thorough and don't miss anything important" to the prompt.
• D. Ask Claude to rate its own confidence in the recommendation.

Correct: B.

The prompt bundled three distinct tasks — extract, summarise, and advise — into one instruction, so the model optimised for a readable answer rather than complete coverage. Splitting it forces the extraction step to happen and be inspected before any judgement is layered on top. This is the blueprint's "apply task decomposition techniques to structure complex requests" objective, and decomposition is the answer to a surprising number of Domain 1 items.

Why the others are wrong:

• A — A stronger model may write a better summary but will make the same scoping decision, because the prompt is what's underspecified. Escalating the model is a Domain 3 move applied to a Domain 1 problem.
• C — Vague intensifiers like "be thorough" don't give the model a structure to be thorough against. Compare "inventory every clause" which does.
• D — Self-reported confidence is not a reliable signal of accuracy, and it doesn't surface the omission. The clause is missing from the summary; asking about confidence in the recommendation won't reveal that.

Question 2 — Domain 2: Output Evaluation and Validation

A market analyst asks Claude to summarise the competitive landscape for a category. The output includes: "Gartner's 2026 report places the segment at $4.2B, growing 18% year over year." The analyst has no Gartner subscription and cannot check the figure.

What should she do before the number goes into a board deck?

• A. Include it with the Gartner attribution, since the source is named and the deck can be corrected later.
• B. Remove the specific figure and describe the trend qualitatively instead.
• C. Treat the citation as unverified, source the figure independently, and only use it if a document she can actually see supports it.
• D. Ask Claude whether it is certain about the figure.

Correct: C.

A named source with a precise number is the highest-risk pattern in generated text, because specificity reads as authority. An unverifiable citation is not a weaker citation — it is not a citation. The blueprint's "apply fact-checking and validation techniques" and "determine when human review or additional verification is required" objectives both land here.

Why the others are wrong:

• A — "Correct it later" is exactly the failure mode the domain exists to prevent. A board deck is a decision artefact; the correction arrives after the decision.
• B — Better than A, but it discards a potentially real and useful figure without trying to confirm it. The exam rewards verification over avoidance.
• D — Asking the model to self-assess produces another unverifiable output. You cannot validate a claim using the same source that made it.

Question 3 — Domain 2: Output Evaluation and Validation

(Multiple response — select TWO.)

A comms lead has Claude draft a customer-facing incident notice from her rough notes. Which two checks should she prioritise before sending?

• A. Confirm every factual detail — timing, scope, affected systems — against the incident record.
• B. Verify the draft's tone and reading level suit customers rather than the internal audience the notes were written for.
• C. Check the word count is within the usual range for these notices.
• D. Ask Claude to rewrite it twice more and pick the best version.
• E. Confirm the draft hasn't introduced commitments — remediation dates, compensation — that weren't in her notes.

Correct: A and E.

Both are accuracy checks against the source material, and E covers the specific hazard of generated text: fluent drafting tends to fill gaps with plausible-sounding content, and in an incident notice an invented remediation date is a commitment the company then owes. A and E are the two that carry real consequence.

Why the others are wrong:

• B — Legitimate and it maps to "edit, adapt, refine... for the intended audience" — but on a select-two item it ranks below two checks that prevent factually wrong or contractually binding statements. Exam items of this shape ask you to prioritise, not to identify everything defensible.
• C — House-style housekeeping, not validation.
• D — Generating more variants doesn't validate any of them. Three unverified drafts are not better than one.

Question 4 — Domain 3: Product and Model Selection

An operations team runs a nightly job that classifies roughly 4,000 support tickets into eight routing categories. The rules are simple and well documented. The team is choosing between Claude's model tiers.

Which reasoning best fits the task?

• A. Use the most capable tier available — accuracy matters most in routing.
• B. Use the fastest, lowest-cost tier, since high-volume classification against clear criteria is where speed and cost dominate; escalate only if measured accuracy falls short.
• C. Use the mid tier as a compromise, since it's a reasonable default for most work.
• D. Alternate tiers across the batch to balance cost and quality.

Correct: B.

The blueprint asks you to "align model selection with task requirements (cost, speed, quality)." The determining features here are high volume, low task complexity, and well-specified criteria — the profile that fits the fast, economical tier. The second clause matters as much as the first: the choice is provisional and validated against measured accuracy, not asserted.

Why the others are wrong:

• A — Reflexively picking the most capable tier ignores cost and latency entirely, which is precisely the tradeoff being tested. Deep reasoning capability is not what eight-way classification against documented rules needs.
• C — "Reasonable default" is not task alignment. The exam wants a reason tied to the workload's characteristics.
• D — Deliberately producing inconsistent classifications across one batch. Introduces variance with no benefit.

Question 5 — Domain 4: Workflow Integration and Solution Design

A department head wants to introduce Claude for first-draft grant reports. In the kickoff meeting, a programme officer says: "If it writes the reports, what stops it inventing a figure and us submitting that to the funder?"

What is the strongest response?

• A. Explain that current models rarely hallucinate and the risk is small in practice.
• B. Acknowledge the risk as real, and describe the workflow control: Claude drafts narrative from figures the team supplies, and every number in the final report is checked against the source system by a named reviewer before submission.
• C. Suggest starting with a small pilot and revisiting the concern with data.
• D. Explain that the funder is unlikely to audit the figures.

Correct: B.

The blueprint objective is "communicate Claude's value and limitations to stakeholders," and the officer has correctly identified a genuine limitation. The strong answer names the limitation as real and answers it with a specific process control — a defined division of labour and a named human accountable for verification. That's a design answer, not a reassurance.

Why the others are wrong:

• A — Minimising a correctly identified risk. It is also the answer that loses the room, because the officer's concern is well founded.
• C — Defers the question rather than answering it. A pilot without a verification control just moves the risk to a smaller sample.
• D — Reframes an accuracy obligation as a detection risk. On the governance-adjacent items, any answer resting on "we probably won't get caught" is wrong.

Question 6 — Domain 4: Workflow Integration and Solution Design

A recruitment team spends about six hours a week on a manual process: read each incoming application, extract eight fields into a tracker, then write a two-line summary for the hiring manager. They want to use Claude to reduce that.

Which approach best fits the blueprint's guidance on integrating Claude into existing workflows?

• A. Replace the whole process — have Claude read applications, populate the tracker, and send summaries to hiring managers directly.
• B. Map the process first, apply Claude to the extraction and summarisation steps where it's strongest, keep the routing and the hiring-manager relationship human, and measure the time actually saved.
• C. Keep the process as it is and use Claude only when the team is behind.
• D. Ask Claude to redesign the recruitment process end to end and implement its recommendation.

Correct: B.

"Integrate Claude into existing workflows to augment or redesign them" is about targeted application with measurement, not wholesale replacement. Extraction and summarisation are exactly where the tool is strong; the judgement and relationship steps are where it isn't. Measuring the saving closes the loop.

Why the others are wrong:

• A — Automates a candidate-facing pipeline end to end with no human checkpoint. Also an employment-decision context, where Domain 6 considerations apply and unreviewed automation is the wrong instinct.
• C — Ad-hoc overflow use captures none of the recurring six hours and produces no measurable outcome.
• D — Delegates the design decision itself to the model and implements it unreviewed. The blueprint asks you to use Claude to support solution design, not to outsource the design authority.

Question 7 — Domain 5: Configuration and Knowledge Management

A finance team sets up a Claude Project for quarterly close questions and uploads the current close checklist, the chart of accounts, and last quarter's workpapers. Six weeks later, answers are increasingly wrong — the checklist changed in week two and two accounts were renamed.

What does this indicate?

• A. The Project's knowledge sources have gone stale; configurations and uploaded knowledge need an owner and a maintenance cadence.
• B. The Project should not have been used for finance work.
• C. Users should add more context in each individual prompt to compensate.
• D. The model has degraded and the Project should be rebuilt from scratch.

Correct: A.

The blueprint has an explicit objective for this: "inform, maintain, and update Claude configurations, knowledge sources, and instructions." A knowledge base is a living artefact. Uploading documents once and treating the Project as finished is the most common configuration failure, and the fix is ownership plus a refresh cadence tied to how often the underlying documents change.

Why the others are wrong:

• B — The use case is sound; the maintenance was missing. Abandoning a working pattern over a fixable process gap.
• C — Pushes the maintenance burden onto every user on every prompt, which defeats the point of a Project and guarantees inconsistency.
• D — Misattributes a data problem to the model. Rebuilding from scratch with the same no-maintenance habit reproduces the failure in another six weeks.

Question 8 — Domain 6: Governance, Risk, and Responsible Use

An HR analyst wants Claude to help summarise themes from an employee engagement survey. The raw export includes free-text comments, employee IDs, department, and tenure. Several comments name individual managers.

What is the appropriate handling?

• A. Upload the full export — engagement data is internal, so no external disclosure occurs.
• B. Check the organisation's AI and data-handling policy first, then upload only what the analysis needs: strip employee IDs, and handle the manager-identifying comments according to whatever confidentiality was promised to respondents.
• C. Upload everything but instruct Claude not to reference individuals by name in its output.
• D. Avoid using Claude for any HR data.

Correct: B.

Two blueprint objectives combine: "apply data sensitivity, regulatory, and privacy considerations" and "follow organisational AI policies and governance standards." Policy is checked first, then data minimisation is applied — only what the analysis requires. The survey's confidentiality promise to respondents is a real constraint that survives the data leaving the survey tool.

Why the others are wrong:

• A — "Internal, so it's fine" conflates external disclosure with the only privacy risk. Respondents were promised confidentiality from colleagues, which internal handling doesn't satisfy.
• C — An output-level instruction doesn't undo an input-level disclosure. The identifying data was still submitted; the instruction only shapes what comes back.
• D — Over-restriction. The blueprint asks you to "identify appropriate and inappropriate use cases," which requires distinguishing them, not refusing the category. Thematic analysis of properly minimised survey data is an appropriate use.

Question 9 — Domain 6: Governance, Risk, and Responsible Use

A manager asks Claude to review six months of a direct report's written output and recommend whether to place them on a performance improvement plan.

Which assessment is correct?

• A. Appropriate — Claude can analyse the writing objectively, without the manager's bias.
• B. Appropriate if the manager reviews the recommendation before acting.
• C. Inappropriate as framed. Claude can help structure the manager's own documented observations, but the employment decision needs human judgement, full context the written output doesn't contain, and whatever process the organisation's HR policy requires.
• D. Appropriate if the employee is told Claude was used.

Correct: C.

This is the clearest "inappropriate use case" shape on the exam: a consequential decision about a person, made from a partial evidence base, delegated to a tool. Note what makes C right — it doesn't ban the tool from the situation. It relocates it to the defensible role of helping organise the manager's own observations, while the decision stays with the human who holds the context and the accountability.

Why the others are wrong:

• A — "Objective, without bias" is the wrong mental model. Model outputs carry their own biases, and six months of written output is a narrow and unrepresentative slice of someone's performance.
• B — Human review is necessary but not sufficient here. A reviewer presented with a generated recommendation tends to ratify it; the anchoring is the problem, so the recommendation shouldn't be generated in that form.
• D — Disclosure is good practice and does not make an inappropriate use appropriate. Telling someone about a flawed process doesn't fix the process.

Question 10 — Domain 7: Troubleshooting and Optimisation

A user reports Claude "isn't good at" writing their weekly project status updates — the outputs are generic and miss what leadership cares about. Their prompt is: "Write my weekly project status update."

What is the correct diagnosis?

• A. The task is a poor fit for Claude and should stay manual.
• B. The prompt supplies no project-specific input — no status, no risks, no audience, no format — so the model can only produce a generic template. The fix is supplying that context, ideally through a Project holding the format and the standing audience, with the week's specifics in the prompt.
• C. A more capable model tier is needed.
• D. The user should ask for several versions and pick the closest.

Correct: B.

"Identify, diagnose, and resolve issues with underperforming prompts or poor outputs." The diagnostic habit the domain is testing: when output is generic, check what input was actually provided before concluding anything about capability. Here the prompt contains no facts about the project, so a generic template is the only correct response to it. The Project-based fix also stops the user re-supplying the same standing context every week.

Why the others are wrong:

• A — Abandons a well-suited task on the basis of one underspecified prompt.
• C — No model tier can infer this week's risks and blockers from a prompt that contains none. Escalating tier is the reflex answer on troubleshooting items and is usually wrong.
• D — Generates several equally generic drafts. Variant-shopping is not diagnosis.

How to study from here

Weight your practice the way the blueprint weights the exam. Domains 2, 6, and 7 together are 46% of the paper — judging, governing, and fixing output. Most freely available CCAO-F material is heavy on Domain 1 prompting technique, which is 14%.

Practise the reading, not the recall. These scenarios are long and the wrong answers are plausible. With two minutes per item you have time to read the full stem; the failure mode is pattern-matching on a keyword in the first line.

Learn the distractor shapes. Across the ten questions above, four recur often enough to be worth naming: escalate the model tier (wrong when the problem is the input), ask the model to check itself (never a validation method), disclose or review after the fact (doesn't cure an unsound design), and refuse the category outright (over-restriction is a wrong answer too, not a safe one).

Use your score report. Pass or fail, the report gives percent-correct by domain. On a retake that tells you exactly where the 90-day wait should be spent.

Confirm your Academy access before you book anything. It remains the step that most often stops people, and it isn't in most study guides.

Sources

All exam facts above come from Anthropic's Claude Certified Associate – Foundations Exam Guide v1.0 (July 2026) and the official Pearson program page. Independent verification:

• Pearson VUE — Claude Certification Program by Anthropic: https://www.pearsonvue.com/us/en/anthropic.html
• CCAO-F exam guide summary, domain weights and format: https://claudecertificationguide.com/ccao-f
• Program timeline and per-exam pricing: https://medium.com/@roanmonteiro/the-complete-guide-to-anthropics-claude-certifications-the-4-exams-the-prerequisite-that-blocks-4d1f743bc5c4

Comparison figures in the timing paragraph: AWS AIF-C01 is 65 items in 90 minutes; CompTIA SecAI+ (CY0-001) is a maximum of 60 items in 60 minutes per CompTIA's certification page (https://www.comptia.org/en-us/certifications/secai/).