End-to-end case studies
Real work crosses all seven domains at once. These eight case studies follow a task from start to finish and point out which exam domain each decision belongs to. The organizations and people are fictional; the patterns are the ones the exam tests.
Case study 1: The operations team that rebuilt its weekly report
Background. Harbor Logistics, a 400-person distribution company, produces a weekly operations report for its leadership team. Two analysts spend most of Monday pulling numbers from three spreadsheets, writing commentary and formatting a slide deck. Leaders complain the report arrives late and buries the important points. Step 1: Map the process (D4). The analysts list each step: export data, reconcile three sources, calculate 12 metrics, write commentary, build slides, send. Scoring shows the commentary and slide steps are text heavy, frequent and easy to check. Reconciliation is error-prone and high risk.
Step 2: Choose tools and models (D3). They set up a Project. Calculations go through code execution, not prose. Sonnet is the default; the analysts try Haiku for formatting and find it adequate. Opus is reserved for quarterly deep-dives. Step 3: Configure (D5). Project instructions define the 12 metrics, the variance threshold (flag anything over 10% week on week), the audience (busy executives) and the format (one summary slide, then one slide per area). Knowledge includes the metric definitions and three past reports leadership liked. The shared metrics folder is connected through Google Drive. Step 4: Prompt (D1). The weekly prompt is short because the Project carries the context: "Using this week's files in the metrics folder, calculate the 12 metrics with code, compare with last week, flag variances over 10%, and draft the report in the standard format."
Step 5: Evaluate (D2). An analyst reconciles three headline numbers against the source system every week and reads every flagged variance. In week two, Claude attributes a delivery delay to weather; the analyst knows the real cause was a staffing gap and corrects it. They add an instruction: "Do not state causes for variances; list them as 'Cause: to be confirmed by analyst.'" Step 6: Govern (D6). The data includes no personal information, and the company's Team plan is approved for internal operational data. The report notes that it is AI-assisted, as the company policy requires.
Step 7: Optimize (D7). After a month, the analysts notice Claude sometimes uses an old metric name. The cause is an outdated definitions file; they replace it. They also batch the slide notes into one request instead of 12.
Outcome (illustrative). Report preparation drops from about seven analyst-hours to about two, including checks. The report reaches leadership by 10 a.m. Monday. The key lesson is that the analysts kept ownership of causes and conclusions while Claude handled calculation and drafting.
Case study 2: A marketing team scales content without losing its voice
Background. Fernwood, a software company, has a four-person marketing team that must publish two blog posts, ten social posts and a newsletter every week. Freelance drafts often miss the brand voice.
Decisions.
- Decision Choice
- Do‐
- main
- Where to store brand
- voice
Brand guide and five exemplary posts in a Project; key rules in instructions D5 How to prompt for
blogs
Brief pattern: audience, goal, key message, keyword, length, call to action D1 Research for posts Research mode for industry statistics, with sources D3 Verifying statistics Every statistic opened at its source before publishing; unverifiable ones removed D2 Repurposing One approved blog becomes social posts and a newsletter item via channel-specific instructions
D4
Competitor mentions Instructions forbid claims about competitors; legal reviews any comparison content D6 Fixing a flat tone Added three "before and after" examples showing the desired voice D7 What went wrong once. A blog draft cited "a 2025 study showing 73% of managers waste five hours a week in meetings." The editor could not find the study. The statistic was removed, and the instructions now say: "Include statistics only with a working source link from Research mode."
Outcome. Output doubled with the same team. Editing time per blog post fell, and voice consistency improved. The team still reads every piece before publishing.
Case study 3: A consultant uses Projects for a client engagement
Background. Sofia is an independent consultant hired by a city council to review its customerservice channels. Setup (D5). One Project per client. Instructions capture the council's context, her report style and boundaries: no resident personal data; flag anything legal to the council's counsel. Discovery (D4, D1). She uploads anonymized call-centre statistics and her interview notes. She uses the "interview me" pattern to build the report outline, then asks Claude to extract themes with supporting quotes.
Benchmarking (D3, D2). She uses Research mode to find how comparable councils handle digital channels and checks every cited source. Two sources turn out to be outdated press releases; she removes them. Governance (D6). The council's data-sharing agreement allows aggregated statistics only. Sofia confirms her uploads meet that standard and notes it in her methodology section. Troubleshooting (D7). Early drafts are too long. She adds "Executive summary under 300 words; each finding under 150 words with one recommendation" to the instructions. Escalation. The council asks whether Claude could run a public-facing chatbot. Sofia explains that this is a technical, governance-heavy project for the council's IT team and a qualified developer, and outlines the questions they would need to answer.
Case study 4: An HR team builds a policy assistant, carefully
Background. A 1,200-person manufacturer's HR team answers hundreds of routine policy questions each month. Design (D4). An internal Project, shared with HR staff only, answers policy questions from the employee handbook, leave policy and benefits summary. HR staff use it to draft answers, which they review before replying to employees.
Configuration (D5). Instructions say: answer only from the attached policies; quote the relevant section; if the question involves an individual's circumstances, medical information or a dispute, reply "Refer to an HR advisor" and do not answer. Governance (D6). Employees' personal details are never entered. The assistant is not used for disciplinary, performance or termination decisions. The policy owner reviews the Project quarterly.
Evaluation (D2). For the first month, a senior advisor reviews a random 10% of drafted answers. Accuracy is high, but two answers relied on an outdated parental-leave clause; the file is updated.
End-to-end case studies
Lesson. The design keeps a human between Claude and employees, limits the scope to factual policy questions, and builds in escalation for anything personal.
Case study 5: A project manager rescues a failing pilot
Background. A project management office piloted Claude for status reporting. After six weeks, managers say the reports are "generic" and "sometimes wrong," and usage has dropped. Diagnosis (D7).
Symptom Cause found
Generic reports Each PM wrote a different one-line prompt with no context Wrong dates Reports drew on old plan files pasted weeks earlier
Inconsistent format No template
Low trust No review step; errors reached sponsors
Fix. A shared Project with a status-report template and instructions (D5), connected to the live project folder (D5), a standard prompt (D1), and a rule that PMs verify dates, budget figures and risks before sending (D2). The PMO explains the change to sponsors, including the review step (D4).
Outcome. Usage recovers, and sponsors report fewer corrections. The PMO learns that the pilot failed on setup and review, not on Claude's capability.
Case study 6: A nonprofit writes better grant applications
Background. Riverbend Literacy, a nonprofit with eight staff, applies for about twelve grants a year. Each application takes a program manager roughly a week, and quality varies. Setup (D5). A "Grants" Project holds the mission statement, program descriptions, the latest outcomes data, the budget template and three successful past applications. Instructions describe the organization's voice (warm, evidence-led, no jargon), require every outcome figure to come from the outcomes file, and ask Claude to flag any funder requirement it cannot meet from the sources.
Working the application (D1, D4). For each funder, the manager pastes the funder's guidelines and asks Claude to (1) list every requirement and word limit in a checklist, (2) map each requirement to available evidence, (3) flag gaps, and (4) draft each section within its limit. Validation (D2). The finance lead checks every budget figure. The program lead checks every outcome statistic against the outcomes file. In one draft, Claude wrote that the program "doubled reading scores"; the data showed a 42% average improvement, and the claim was corrected. Model choice (D3). Sonnet handles drafting. For a large, complex federal application with many interlocking requirements, the team uses Opus for the requirements mapping step. Governance (D6). Student names and personal stories are used only with documented consent, and anonymized otherwise. The nonprofit's board-approved AI policy requires human review of all external submissions.
Outcome (illustrative). Drafting time per application drops from about five days to about two, and the team submits more applications. The lesson: grounding in a single outcomes file prevented exaggerated claims from reaching funders.
Case study 7: A university department streamlines course administration
Background. A university department's administrator handles course-change requests, timetable queries and student emails, using a 90-page academic regulations handbook. Design (D4). The administrator creates a Project with the handbook and the department's FAQ. Claude drafts replies to routine queries, quoting the relevant regulation. Anything involving an individual student's grades, appeals, wellbeing or accommodations is routed to the appropriate officer and never drafted by Claude.
Configuration (D5). Instructions: "Quote the regulation number for every rule you cite. If the handbook does not answer the question, say 'Not covered; refer to the registrar.'" The administrator replaces the handbook each academic year and deletes the old one. Validation (D2). For the first month, the administrator checks every quoted regulation number against the handbook. Two errors in the first week came from an outdated appendix; the appendix was removed. Troubleshooting (D7). Replies were too formal for students. Adding two example replies in the desired tone fixed it.
Governance (D6). No student identifiers are pasted; queries are paraphrased. The university's approved Claude workspace is used.
Lesson. Clear escalation boundaries let the administrator save time on routine work while keeping sensitive cases with qualified staff.
Case study 8: A small retailer launches a seasonal campaign
Background. Juniper & Co., a three-store home-goods retailer, has no marketing department. The owner, Sam, wants a holiday campaign across email, social media and in-store signage. Planning (D4, D1). Sam uses the "interview me" pattern to build a campaign brief, then asks for three campaign concepts with pros, cons and estimated effort, and picks one.
Research (D3, D2). Sam uses web search to check typical holiday shopping timelines and confirms key dates on official sources before building the calendar.
End-to-end case studies
Production (D1, D3). Claude drafts five emails, twenty social posts and four signage messages in one batched request using a table of products, prices and promotion dates. An Artifact holds the campaign calendar, which Sam edits as plans change.
Validation (D2). Sam checks every price, date and discount against the point-of-sale system. One draft offered "free shipping on all orders"; the actual offer applied over $75, and Sam corrected it. Governance (D6). Customer email addresses never go into Claude; Sam pastes the approved copy into the email platform. Claims such as "lowest prices in town" are removed because Sam cannot substantiate them.
Optimization (D7). Early social posts sounded generic. Sam adds three past posts that performed well as examples, and later drafts match the store's voice. Lesson. Even a one-person business benefits from structure: a clear brief, batched production, a single source of truth for offers, and careful verification of anything customers will rely on.