Case Studies — Four Complete Worked Examples

Sarah, a freelance UX consultant in the Netherlands, faced quarterly VAT filing with 214 receipts spread across Gmail, Downloads, and phone photos — previously a full resentful weekend of manual sorting. Her new process: collect everything into one AI-Workspace folder (10 min); run an agent session in Claude Code that extracts vendor, date, amount, and VAT from every file, renames files, builds a categorized spreadsheet with a needs-review tab, and shows its plan first (25 min, mostly unattended); review the needs-review tab and spot-check 15 random rows (20 min); create an accountant-formatted summary tab (5 min); and systemize the whole process into a reusable Skill called quarterly-receipts (5 min).

Outcome: about 65 minutes of her time, 20 of which was genuine review, with cleaner input than before for her accountant. What could have gone wrong — the agent misreading a European-format amount, or a personal receipt slipping through — was caught precisely because spot-checking and category review existed. The pipeline is fast because the human checkpoints are cheap and targeted, not because they're skipped.

Marco runs a 4-person marketing agency and wants consistent founder-brand content without hiring. His one-time setup (~90 min): a Claude Project with his audience, positioning, 10 style rules, and banned words; a custom Style built from his top posts; a Canva Brand Kit and carousel template; and a Zapier flow that reformats published posts into newsletter drafts automatically.

His weekly rhythm (Tuesday, 60-75 min): pull a core idea from evidence, like an insight surfaced in Monday's performance review; turn it into five assets — post, carousel, newsletter section, thread, video hooks; run a critique pass asking where a skeptical reader would stop reading and fix those points; design the carousel in Canva via Bulk create; and publish, letting the Zapier flow handle the newsletter draft. Outcome: 5 assets per week from one hour, in a voice that's recognizably his, because the Project and Style did the heavy lifting set up once — and his prompt library kept evolving as he noticed and fixed weak points.

Priya, an operations manager at a 30-person company, handled 80+ emails and 5-7 meetings a day with chronic "I know it's in some thread somewhere" syndrome. Her one-time setup (~45 min) connected Gmail, Calendar, Drive, and Notion to Claude, adopted "draft never send, show before changing," and saved three routines as Skills.

Her day: an 08:30 Morning Brief with meeting prep, ranked emails, top tasks, and "commitments at risk"; a pre-meeting prep pulling contract, negotiation thread, and open issues into one page; a post-meeting recipe extracting decisions and drafting a recap she edits and sends; and a Friday shipping review. Outcome: 6-8 hours a week recovered, but the bigger change was qualitative — nothing falls through cracks. Her trust ladder mattered: week 1 she verified everything, week 3 she spot-checked, by week 6 she trusted reads fully while still approving every send. Trust was earned through verification, not assumed.

Daniel, founder of a DTC brand doing €40k/month across Meta and Google with no agency and 5 hours/week for marketing, runs the Module 11 rhythm concretely: Monday (30 min) analysis using ads connectors and his Ads Project to produce a recommend-only weekly test plan; Tuesday (60 min) production turning the test plan into 10 ad variants via Canva Bulk create, paused and reviewed himself before activating; Friday (15 min) hygiene, including a biweekly negative-keywords ritual that found €410/month of irrelevant search spend in its first month alone; and a monthly memory step appending learnings to his Project's knowledge file so the assistant compounds in usefulness over time.

His guardrails in action: cost caps on every campaign, AI never touching budgets directly, and a "confidence + evidence" line required in every recommendation after one early suggestion turned out to rest on a two-day sample — he caught it by asking what the sample size was, a question you should now ask reflexively.