AI Use Case Casebook · Internal Best Practices
Internal AI Use Case Casebook
What is an AI use case casebook?
- Without the original prompt and cautions, no one can replicate it
- Security check before publishing; separate estimated from measured figures
- Without quarterly additions and twice-yearly pruning, a casebook goes stale
What an AI use case casebook changes is the form of what remains after training. Course materials leave behind how to use a tool, while a casebook leaves behind 'where and how we used it at our company.' Its purpose is to let new employees skip waiting for training and take the prompts and procedures that the department next door already uses. When you draw up next year's training plan, the number of cases and their spread across departments become a map showing which job functions need advanced training and where the gaps are.
The format of a single case should be fixed. Work situation, tool used, the original prompt entered, a sample of the output, before-and-after figures, cautions, and the author's contact are enough. Without the original prompt and the cautions, a case makes good reading, but no one can replicate it. For distribution, a searchable, editable format such as an internal wiki or a prompt library lasts longer than a single PDF. Cases usually come from four sources: contest winners, hackathon outputs, entries gathered by AI champions in their departments, and on-the-job application plans submitted after training. Set a cycle of adding cases once a quarter and pruning them once every six months.
A casebook is also a security document. The prompts and outputs in the cases easily pick up client names, unit prices, and personal data, and distributing them company-wide as they are turns the casebook itself into a data leak channel. You need a process in which the information security officer reviews each case before publishing, and figures and proper nouns are pseudonymized. Conversely, adding 'what not to do' cases, such as a time a hallucination put a wrong number into a report, makes for training material far more persuasive than a guideline.
The most common failure in practice is building the casebook once and stopping there. AI services change features every few months, so a prompt from six months ago often no longer works as written, and an outdated casebook creates “I tried it and it didn’t work” experiences that drag down adoption. Another is printing impact figures without verification. “Saves 20 hours a month” is the author’s estimate, but once it’s in the casebook it circulates as fact, gets summed into executive reports, and inflates next year’s expectations. Label figures as author estimates or add numbers measured by someone who reproduced the case, and update the contact and the maintaining owner whenever an author leaves or changes departments.
John KimInstructor John Kim, author of B2B Marketing Design and CEO of Nanum Consulting
Seolyeong ParkGenerative AI and AX training specialist with 17 years of hands-on experience and 200+ lectures, helping non-technical staff put AI to work right awayIt applies best-practice management, the documenting and sharing of strong cases in knowledge management (KM), to generative AI use. As experience showed that tool-usage training alone doesn’t translate into department-level application, it became established in corporate training as a repository for the results of internal contests and hackathons.
Consider a hypothetical case. Company A, a software distributor with 640 employees, bundled 15 award-winning entries from last year-end's contest into a PDF casebook and distributed it. Six months later, a survey showed that only 19% had ever opened it. This October, the training team decided to rebuild the casebook as a wiki.
- Assistant Manager Yoon on the training team fixed the template at seven fields: 'work situation, tool, original prompt, sample result, before-and-after figures, cautions, author.' For the 6 of the existing 15 cases that had no original prompt, Yoon went back to the authors to get them again.
- The information security team's review flagged 4 cases that contained client names and unit prices as they were. Assistant Manager Yoon pseudonymized the figures and proper nouns, then added the note "Please leave this field blank" to the template.
- When Manager Lee of the sales support team shared a hallucination incident, in which AI-invented delivery terms almost made it into a quotation, in the failure-case section, it became the most-viewed case of all.
- They set a cycle in which 8 AI champions each submit two cases from their departments every quarter, and prompts that no longer work are pruned every six months. They also added a field to change the maintenance owner when an author moves to another team.
- Six months later, the casebook had grown from 15 cases to 52, and monthly viewers had risen from 19% to 61% of all employees. In next year's training plan, purchasing and general affairs, two functions with no cases at all, became the top priority for advanced training.