Output Evaluation and Validation

Question 1 · 2.1

A sales operations manager asks Claude to summarise 40 customer call notes into the top objections by region, for a VP review tomorrow. The response is well organised, with five themes and a confident tone.

What is the most appropriate next step before sending it?

• AAsk Claude whether it is confident the summary is accurate.
• BCheck that every region and note is accounted for, then trace key themes back to the notes.
• CRegenerate the summary twice and send the version that reads best.
• DProofread it for tone and formatting so it suits a VP audience.

Question 2 · 2.1

An analyst uploads a 180-page annual report, full of charts, to a Claude chat and asks for commentary on year-on-year changes in operating margin. She plans to use the figures in a client presentation.

Which TWO actions most improve confidence in the accuracy and completeness of the figures? (Select 2.)

• AAsk Claude for a longer, more detailed response so nothing is left out.
• BAsk Claude to quote the passage behind each figure and flag any it cannot find.
• CAccept the figures, since the report was uploaded and Claude has read it.
• DRecompute the year-on-year changes herself from the report’s tables.
• ETurn on web search so Claude can confirm the figures online.
• FAsk Claude to rate its confidence in each figure from 1 to 10.

Question 3 · 2.1

Using web search, Claude tells a marketing lead that “62% of mid-market buyers now prefer self-serve purchasing”, with a citation. She clicks the link: it is a real industry report on the right topic, but the 62% figure refers to enterprise buyers in a different year.

What does this situation best illustrate?

• AThe claim is verified, because the citation links to a real report.
• BWeb search should be avoided for statistics in business content.
• CA citation shows where Claude looked; the claim still needs checking against the page.
• DThis is not an error, because hallucinations only involve made-up sources.

Question 4 · 2.1

An HR partner asks Claude, in a Project containing five office handbooks, to compare parental leave policies. The table covers four offices; the fifth is silently missing. She wants to prevent this next time.

Which change to her request best addresses the problem?

• AName all five offices and require a row for each, marked “not found” if absent.
• BAdd “Please be thorough and don’t miss anything” to the request.
• CAdd several paragraphs of background on why the comparison matters.
• DSwitch to the most capable model for all HR comparison tasks.

Question 5 · 2.2

Without web search, Claude drafts a market overview for a sales kickoff. It includes general trends, a paragraph on customer pain points, and the line “According to a 2025 industry survey, 47% of mid-size firms plan to cut vendors next year.”

Which part of the draft is the most likely hallucination risk?

• AThe general description of market trends.
• BThe paragraph describing typical customer pain points.
• CThe attributed statistic with a year and a percentage.
• DNone of it — without web search Claude only uses training data.

Question 6 · 2.2

A quarterly memo drafted with Claude says in its summary that revenue grew 12% year on year. The table on page two, built from the same spreadsheet, shows 9%.

What is the best next step?

• AUse the table figure, because tables are more reliable than prose.
• BRecalculate the growth from the spreadsheet and correct both places.
• CAsk Claude which of the two figures is correct and use its answer.
• DRegenerate the memo until the two figures agree with each other.

Question 7 · 2.2

A marketing manager asks Claude: “Explain why our spring campaign was a big success.” The response is glowing. She wants an honest assessment to share with her director.

Which change is most likely to give her a balanced assessment?

• AAdd “be honest and objective” to the same question.
• BAsk Claude whether it was just telling her what she wanted to hear.
• CRun the same prompt again on a more capable model.
• DShare results and targets; ask what worked, what didn’t, and why.

Question 8 · 2.2

A lending team proposes having Claude read loan applications, including applicants’ personal details, and output an approve or decline recommendation that is applied automatically.

Which response best reflects what is known about bias in this use?

• AInstruct Claude to ignore demographics, then apply its decisions automatically.
• BRemove personal details, use fixed criteria, and have a qualified person decide.
• CProceed as proposed, because Claude is trained to be even-handed and fair.
• DGo live, then check a random sample of the decisions after a month.

Question 9 · 2.3

A Research report prepared for a compliance team states a new rule’s compliance date and cites a consultancy’s newsletter. The team will use the date to plan system changes.

Which check is most appropriate before relying on the date?

• AConfirm the newsletter link opens and mentions the date.
• BRun Research again and see whether it reports the same date.
• CFind a second news article that gives the same date.
• DCheck the date in the regulator’s own published final rule.

Question 10 · 2.3

An analyst has Claude, with code execution on, calculate customer churn by segment from an uploaded CRM export. The results will go into a quarterly business review.

Which TWO checks best validate the analysis? (Select 2.)

• AAsk Claude to confirm the churn figures are correct.
• BReconcile the customer counts and totals with the CRM’s own reports.
• CAccept the figures, since code execution did the calculations.
• DRecompute churn by hand for a small sample of segments.
• ERun the same prompt again and check the figures match.
• FAsk for more charts so patterns are easier to see.

Question 11 · 2.3

A sales director wants to quote the company’s prior-year revenue in a press briefing. Which source gives the strongest verification?

• AClaude’s answer, after it says it has double-checked.
• BThe figure in the company’s audited annual report.
• CA news article that Claude cited via web search.
• DTwo different AI assistants giving the same figure.

Question 12 · 2.3

A team adds this to its prompt: “For each claim, find a supporting quote from the attached documents; remove any claim you cannot support.”

Which statement best describes what this achieves?

• AIt guarantees the response will contain no hallucinations at all.
• BIt makes claims traceable and drops unsupported ones; spot-check the quotes.
• CIt means the output no longer needs any human review before use.
• DIt only has an effect when web search is turned on in the chat.

Question 13 · 2.4

Which of these outputs most clearly requires review by a qualified professional before it is sent?

• AA list of team-building ideas for an internal offsite.
• BA summary of this morning’s project meeting for the team.
• CA letter telling a policyholder why their insurance claim is denied.
• DA first outline of headings for a company blog post.

Question 14 · 2.4

A consultancy wants to speed up client reporting. It proposes adding “This report was generated with AI and may contain errors” to each report and dropping the partner review step.

What is the best response?

• AAccept it, because the disclaimer shifts responsibility to the client.
• BKeep the review; the disclaimer is disclosure, not a check on accuracy.
• CAccept it, and review a sample of reports after they are sent.
• DReplace the partner review with a request for Claude to self-check.

Question 15 · 2.4

An operations lead wants Cowork to read supplier emails and update bank details in the payments spreadsheet overnight, unattended, so it is ready for the morning payment run.

What is the most appropriate change to this plan?

• ALet it run, since it only reads the company’s own inbox.
• BRun it as planned and review the spreadsheet after the payment run.
• CHave it propose changes for approval, confirmed by phone on file.
• DAdd an instruction telling Claude to ignore suspicious emails.

Question 16 · 2.4

An HR team in the EU sets up a workflow where Claude scores job applications and automatically rejects everyone below a threshold. No one looks at the rejected applications.

What is the main problem with this design?

• ACandidates may be rejected by a solely automated decision with no human involvement.
• BScoring many applications in one chat may exceed the context window.
• CAutomatically generated rejection emails may sound impersonal to candidates.
• DClaude may be slower than a recruiter at reading each application.

Question 17 · 2.5

A project manager has a detailed technical status report and needs a version for the steering committee, who will decide next week whether to extend the project’s budget.

Which revision request is most likely to produce a useful version?

• A“Rewrite this so it is more suitable for senior leadership.”
• B“Make this shorter and more professional, and don’t use jargon.”
• C“For a budget decision: open with the ask and cost, keep all figures, 300 words.”
• D“Rewrite this in the Formal style so it reads well for executives.”

Question 18 · 2.5

Claude rewrites internal incident notes into a customer-facing apology email. The email reads warmly and clearly.

Which TWO checks matter most before sending it? (Select 2.)

• AConfirm the scope and figures match the incident notes exactly.
• BAsk Claude whether the email is appropriate for customers.
• CMake sure the email is longer, so it seems more sincere.
• DCheck it makes no commitments beyond what has been approved.
• ESwitch to the Explanatory style so customers understand the cause.

Question 19 · 2.5

A sales lead has two Claude drafts of a client proposal summary. Draft A is longer and more elegantly written. Draft B is shorter and puts the pricing and next steps first. Both are accurate.

What is the best way to choose between them?

• APick Draft A, because a longer draft covers more ground.
• BAsk Claude which draft is better and go with its answer.
• CPick the draft that sounds most like the sales lead’s own voice.
• DJudge both against what the client needs and pick the better fit.

Question 20 · 2.5

A marketing team wants every Claude draft to match its brand voice without writing out the rules in each prompt.

Which approach best fits this need?

• ACreate a custom style from approved brand writing samples and use it when drafting.
• BSelect the Concise preset style for all marketing drafts.
• CAdd “use our brand voice” to the start of each prompt.
• DPaste the full brand guide into every new conversation.

Question 21 · 2.6

In a chat, a team lead asks Claude what the difference is between the company’s two expense policies for client entertainment. She needs the answer once, before replying to a colleague.

Which output format is most appropriate?

• AAn artifact containing a full comparison document.
• BA short inline answer in the conversation.
• CA CSV file with one row per policy clause.
• DA PowerPoint deck created with file creation.

Question 22 · 2.6

A procurement analyst needs details from 30 supplier quotes so her team can sort by price and lead time in Excel and import the shortlist into the purchasing system.

What should she ask Claude for?

• AA well-written summary paragraph for each of the suppliers.
• BAn artifact with a narrative comparison of the suppliers.
• CA spreadsheet with named columns, units and a missing-value rule.
• DAn inline bulleted list of each supplier’s key terms and prices.

Question 23 · 2.6

An operations manager is creating a step-by-step guide to a new returns process. It will be reviewed by two colleagues, updated over several weeks and shared with store managers.

Which format best fits this output?

• AAn artifact that can be iterated on, versioned and shared.
• BAn inline answer she copies from the chat when needed.
• CA JSON file with one object per step of the process.
• DA new chat each time she needs to update the guide.

Question 24 · 2.6

A strategy analyst has Claude summarise 15 market reports into a brief for the executive team. Two reports give different figures for the same market’s size.

What is the best way to curate the brief?

• ASummarise each report in order, so nothing is left out.
• BUse the average of the two market-size figures.
• CInclude every finding from all 15 reports for completeness.
• DGroup by the executives’ questions, cite sources, and show the conflict.