Putting It Together: Cases and Capstone (part 2 of 2)

This is Unit Nine of Generative AI at Work, part two, the final capstones. Part one ran everyday tasks through the pattern. This part takes on bigger decisions: assessing a tool, documenting a use, designing a high-stakes process, and starting a team well, before drawing the whole course together. Carry one question through this part: what stays true across every use of generative AI at work? The answer arrives with the whole pattern, in eight parts. By the end you will be able to assess a new tool and set its rules. To document a use for the next owner. To design a safe process for a high-stakes goal. To plan a team's first month. And to state the course's whole pattern.

Two capstones on letting a tool in and keeping a use documented.

Capstone three: a vendor tool.

GAI 216 Capstone Three: A Vendor Tool Decision

The third capstone, constructed for this course, follows a team that wants a meeting assistant to record, transcribe and summarize calls. The risk scan shows privacy as likely and severe, because the tool hears client details and personnel matters, and confabulation as likely in summaries. The data terms keep recordings ninety days and use transcripts for training unless an administrator opts out. The model card's caveats warn about overlapping speakers and strong accents, which the team's calls often have. The team's own test of five meetings finds two decisions attributed to the wrong person and one action item missing. The tool is approved for internal meetings only, with the opt-out set, shorter retention, consent from every participant, and summaries checked as drafts.

Capstone four: documenting a use.

GAI 217 Capstone Four: Documenting a Use

The fourth capstone, constructed for this course, hands over the claims-letter process from unit six a year on. The new owner checks the feature gate record and finds a change nobody documented: an update now lets the tool read customer emails, so the privacy review is rerun. A Bronze pack is written for the claim letter. The log shows the rewrite ratio fell, then rose after a template change, with the reason recorded. The current model card shows the vendor changed the underlying model in the spring without a test, so one is scheduled. Both changes arrived gradually, through ordinary updates. Documentation is the habit that makes such drift visible to the next person.

One capstone that applies the course to the hospital case's original goal.

Voice notes to an incident draft.

GAI 218 Voice Notes to Incident Draft

The hospital case begins with a goal its leaders were nervous about: AI drafting incident reports from staff voice notes. This entry is a course application, not something the case reports happening. The gate lets the AI read the reporter's voice note and the incident definition, and write only a draft in the defined fields, never submitting, classifying or assigning severity. Drafts stay on the suggestion side, reviewed and submitted by the reporter. Every number and medication is highlighted for confirmation, because transcription can mishear them. The work runs in the hospital's governed system. The prompt leaves fields empty rather than inferring. Severity and reportability stay with quality and risk staff. Corrections are logged and reviewed monthly with patient safety.

Two capstones: how a team starts, and the whole pattern. The opening question is answered here.

A team's first month.

GAI 219 Your Team's First Month

The sixth capstone, constructed for this course, plans a team's first month. Week one: choose two tasks, one low-stakes and one moderate, write each use in a sentence, and run the risk scan. Week two: write two pages of rules with a named owner. Week three: write criteria, collect real test cases, improve prompts one change at a time, and save the passing ones in a shared library. Week four: run both tasks, log every output as accepted, edited or rejected with a reason, and read the log together. The summary prompt gains a required-points list; the meeting notes keep a light review. Starting small is how a team learns where its own risks actually are.

The whole pattern.

GAI 220 The Whole Pattern

The final capstone draws the course into one pattern, built on the principle it started with: AI proposes, humans decide. First, know the tool: fluent, derived text whose confidence is not evidence. Second, name the risks with NIST's twelve and scan each use. Third, prompt deliberately and test against criteria. Fourth, use safe patterns: use only what is given, invent nothing, ask instead of guessing, and never paste what you would not publish. Fifth, verify claim by claim against real sources. Sixth, keep humans in the loop by design. Seventh, govern the tools. Eighth, document the AI and the work it touches. That answers this part's question: across every use, draft, check and decide. None of the eight makes the tool more accurate; together they make its errors visible, its use accountable, and its benefits safe to keep.

Look at what you can now do. You can assess a new AI tool by its risks, its data terms and your own test, and set its rules of use. You can document a use so the next owner can see what changed. You can design a safe AI-assisted process for a high-stakes goal, with the decisions kept with people. You can plan a team's first month of deliberate use. And you can state the whole pattern in eight parts, built on one principle: AI proposes, humans decide. That completes Generative AI at Work.