Putting It Together: Cases and Capstone (part 1 of 2)

This is Unit Nine of Generative AI at Work, part one, on putting the course together. Unit Eight covered documentation. This final unit puts every unit to work, first through the M-CaReM-AI method's composite hospital case and its domain playbooks, then through capstones built for this course. Carry one question through this part: what does the whole pattern look like on real work, from the first prompt to the final decision? The answer arrives with an extraction checked against an independent source. By the end you will be able to summarize the hospital case. To state the rule that runs through every domain. To test a rule an AI proposes. To apply the domain playbooks. And to run two everyday tasks through the whole pattern.

Five entries on the method's composite hospital case: where AI helped, where it could not lead, and how its errors were caught.

The case the method teaches through.

GAI 201 The Hospital Case

The M-CaReM-AI method teaches through a composite, illustrative case: a regional hospital replacing an old, confusing incident-reporting system. Leadership also wants AI to draft incident reports from staff voice notes, and the team is nervous about accuracy. That nervousness is the starting point: the question is how to use AI safely, not whether.

Where AI helped.

GAI 202 Where AI Helped

In the case, AI helped most with drafting: candidate records from policy text, field lists, first drafts of reference documents, summaries, training material and workshop scenarios, such as a one-page summary checked by an analyst and handed to frontline staff before a validation session. Every one of those helps was a draft, reviewed before use.

Where AI could not lead.

GAI 203 Where AI Could Not Lead

The case is equally specific about limits: AI could not define policy rules unverified, decide which incidents are reportable, define core records, or override human disagreement with generic best practice. A disagreement between nursing and quality staff was settled by the people involved. Each limit is a decision with consequences for patients or compliance.

How the mistakes were caught.

GAI 204 How the Mistakes Were Caught

When AI erred in the case, producing a wrong reporting rule, an invented risk-score field and an oversimplified lifecycle, the team caught each one through its reference documents, structured requirements and human validation sessions, as when frontline staff corrected the lifecycle. None was caught by the AI checking itself.

What the team concluded.

GAI 205 The Team's Conclusion

The case's retrospective concludes that AI was powerful but fallible: it sped up the mechanics of the work and often got policy details wrong. The team's summary is that AI is great at speeding up the method's mechanics, and absolutely not a substitute for domain judgment. Both halves are the lesson; dropping either distorts it.

Two entries on the rule that runs through every domain, and the role the method asks you to take.

The rule across domains.

GAI 206 Draft Around, Never Decide

Across the method's domain playbooks one rule repeats: AI may draft around a consequential decision, but it never makes the decision. In public services, AI may draft a notice or summarize a case for a reviewer, but must not decide eligibility, interpret law or set benefits. The draft is where AI saves time; the decision is where accountability lives.

The role you carry forward.

GAI 207 What You Carry Forward

The method describes a change of role for anyone using AI at work: from AI consumer, who asks for output, to AI framer, who supplies structure, sets limits and catches errors. The framer pastes reference material and forbids invention; the consumer asks for an answer. The method calls the framer's role durable, and this course has been training it.

Six procedures: testing a rule an AI proposes, and the method's playbooks for five domains.

Testing a proposed rule.

GAI 208 The Invented Rule

The case's invented rule, all severe incidents are reportable, calls for a procedure of its own. Write the proposed rule down with every condition it states. Find the authoritative source, with its version and date. Compare conditions and exceptions one by one; in the case, the real rule depended on incident type, harm level and policy version. Confirm the wording with the rule's owners. Record the correct rule so future prompts are grounded in it. Over-general rules are dangerous because they are usually true somewhere.

Public services.

GAI 209 Public Services

The method's public-services playbook sets out where AI is safe. AI may draft notice text from a recorded decision and an approved template, without changing the meaning of any legal phrase. It may flag missing information in an application. It may summarize a case history for the reviewer. It must not decide eligibility, interpret the law, or set benefit levels without validated rules. The case record shows what AI drafted and who decided. The notice is what the applicant sees, so a drafting error lands on a person directly.

Healthcare.

GAI 210 Healthcare

The method's healthcare playbook lets AI support documentation while treatment decisions stay with clinicians. AI may draft clinical notes from structured orders, results and diagnoses. It may suggest next steps for a clinician to review. It may draft patient-friendly care-plan summaries. It must not make final high-risk treatment decisions without a regulated framework, or invent clinical concepts. All of it runs in governed tools. A patient-friendly summary can be wrong in a friendly way, so patient-facing text needs clinical review too.

Lending and fairness.

GAI 211 Lending and Fairness

The method's financial-services playbook treats credit decisions as the highest risk level and builds the documentation to explain them. Map every model input to a defined field of the decision record. Feed the model's output into that record, with rationale and policy reference. Document how model changes alter the decision rules. Be ready to answer regulators: how decisions are made, which factors matter, and how fairness is ensured. Test outcomes across groups, since NIST notes mitigating bias does not by itself make a system fair.

AI features in products.

GAI 212 Product Features

The method's product playbook covers autocomplete, summaries and recommendations, and treats them as features with boundaries, not magic. State exactly which records each feature reads and writes. Keep generated text attached to the record it belongs to. Put each AI feature through the normal design process. Apply the feature gate: what it reads, what it writes, and how output is reviewed. The method's word for the failure is smearing: generated text leaking into records where nobody can tell what it means.

Hiring and people decisions.

GAI 213 Hiring and People Decisions

The method's human-resources playbook puts ethics first: applications, reviews and pay records sit at the highest risk level, and AI may only suggest. AI may parse resumes and suggest a shortlist. Document every place AI influences an outcome and which fields feed the model. Require human review for every sensitive decision. Monitor results across groups, since NIST notes generative AI can amplify biases and perform worse for some groups. Screening errors are silent; monitoring outcomes is the only way to see them.

Two capstones, built for this course, run everyday tasks through the whole pattern. The opening question is answered here.

Capstone one: a policy summary.

GAI 214 Capstone One: A Policy Summary

The first capstone, constructed for this course, follows an operations manager summarizing a revised forklift safety procedure for warehouse staff. Safety makes it high-stakes, so she will check every claim and the safety lead will approve. Her prompt puts the instruction first, names the readers, pastes the full procedure, sets a short plain-language format, and asks for questions on anything unclear. Writing her checklist first, she finds the draft has three of four changes, the missing one folded into a general line. Claim by claim, she finds speed limits reduced everywhere when the procedure reduced them only in loading bays. Both errors softened a safety rule into something more comfortable.

Capstone two: an extraction.

GAI 215 Capstone Two: An Extraction Into Fields

The second capstone, constructed for this course, follows a finance analyst extracting invoice details from a month of supplier emails in a governed tool, with bank numbers excluded. Each field is defined, and missing values are written not stated, never filled in. Checking totals against suppliers' own statements, she finds two mismatches: one invoice extracted twice, and a subtotal read as a total. Eleven missing purchase order numbers become questions for suppliers. That answers this part's question: on real work, the pattern is grounded prompts, independent checks and honest blanks, with every correction recorded.

Look at what you can now do. You can summarize the method's hospital case: where AI helped, where it could not lead, and how its errors were caught. You can state the rule across every domain: AI drafts around decisions and never makes them. You can test a rule an AI proposes against the real one. You can apply the playbooks for public services, healthcare, lending, products and hiring. And you can run a summary and an extraction through the whole pattern. Part two finishes the capstones.