Skip to content
NTPU AI Learn

AD 101 · Module 2

Shape the Use Case

Match the problem to a reasonable AI assistance pattern, review value, data, and risk separately, and finish the formal Use Case Canvas.

Start Section 1
  • Duration90 min
  • LevelApplied
  • PrerequisiteAI 101L recommended first
0 / 3 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor

Info

After you join, your completed lesson numbers are sent to your instructor together with your learner code, so they can see your progress. Use the codes your instructor gave you, and do not enter your name, company secrets or other personal data. Progress is kept with the rest of that session's data.

Learning objectives

  • Use patterns such as summarize/extract/classify/draft/analyze/retrieve/workflow-assist to identify AI's role while keeping human responsibility.
  • Look at value, data availability, verification, and basic risk separately, never a single opaque total score.
  • Complete the formal Canvas, then peer-review it for a vague problem, a smuggled-in solution, or a missing owner/review step.

AI Pattern Matching

30 min

AI Pattern Matching

Learning objective: Use patterns such as summarize/extract/classify/draft/analyze/retrieve/workflow-assist to identify AI's role while keeping human responsibility.

Only after the problem is confirmed do you ask what role AI might play. Common assistance patterns include: summarize, extract (pull out structured information), classify, draft, analyze (find patterns in data), retrieve (find answers in existing material), and workflow assist (help connect steps in a multi-step process). Every pattern is only assistance, the final judgment, approval, and accountability stay with a person, which is this program's core Human-AI-Human principle.

Seven common assistance patterns
FieldMeaning
SummarizeCondense long content into key points
ExtractPull structured information from documents
ClassifySort or tag items
DraftProduce a first draft for a person to edit
AnalyzeFind patterns or anomalies in data
RetrieveFind answers in existing material
Workflow assistHelp connect steps in a multi-step process
Workplace example

For the Lesson 1 meeting-action-item gap, the closest match turns out to be extract (pull action items, owners, and due dates from the minutes) plus classify (sort items into urgent/routine). The extracted result still needs the meeting chair to confirm accuracy and check for missed items before it goes out.

  • Summarize scenario: condense a 30-page evaluation report into a one-page summary for a supervisor, who still reviews the full report.
  • Extract scenario: pull name, unit, and contact details from registration forms into a roster, which staff then spot-check.
  • Classify scenario: sort incoming mail into "leave request", "reimbursement", and "complaint" for different handlers, with the sort still adjustable by hand.
  • Draft scenario: produce a first-draft notice email that a staff member edits for tone and detail before sending.
  • Analyze scenario: find unusually high-priced items in three years of procurement data for the purchasing team to investigate further.
  • Retrieve scenario: find the regulation clauses relevant to a case from internal policy, which a staff member then judges how to apply.
  • Workflow-assist scenario: flag which attachments are still missing at each step of a multi-step approval chain, while the approval decision stays with a person.
Workplace example (contrast case: the wrong pattern)

A unit wants to fix "complaints keep getting the wrong priority" and jumps straight to draft (write a reply draft) as the assistance pattern. But the real gap is in classification and ordering, not drafting; it should pick classify (sort by type and urgency) first, and only add draft later if needed, rather than skipping classification and writing a reply.

Info

Human, AI, Human: a person confirms the problem and judgment first, AI assists only after that, and a person verifies and owns the final result.

Warning

Data safety reminder: every exercise in this course uses course-provided synthetic scenarios only, never real institutional or personal data.
Check yourself

Which of these best fits "AI assists, a person keeps responsibility"?

Which of these best fits "AI assists, a person keeps responsibility"?

Key takeaway: Problem first, pattern second: only after the problem is confirmed do you choose the closest-fitting assistance pattern, and assistance never replaces judgment.

Value / Data / Risk Quick Check

30 min

Value / Data / Risk Quick Check

Learning objective: Look at value, data availability, verification, and basic risk separately, never a single opaque total score.

Before writing the problem up as a formal Use Case, run three separate quick checks, deliberately never combined into one score: Value (is the payoff an activity or an observable result?), Data (what data is needed, where does it come from, is there an authorization concern?), and Risk (if the AI output is wrong, what happens, and what level of human review is needed?). Keeping them separate prevents a single score from hiding a real concern, such as a high "total" masking a serious data-authorization problem.

Three separate checks, never merged
FieldMeaning
ValueIs the payoff an observable result, or just an activity?
DataWhat data is needed? Is the source and authorization clear?
RiskWhat happens if the output is wrong? What review level is needed?
Workplace example

For the meeting-action-item case: Value is "action-item turnaround drops from 3 days to same-day" (observable); Data is "minutes are provided by the meeting chair, no external personal-data concern"; Risk is "if items are missed or miscategorized, some work gets overlooked, so the chair must check every item before it goes out". The three are stated separately, never merged into a single score like "85/100".

  1. Write Value first
    What observable result does solving this actually improve? Time, quality, or capacity?
  2. Then write Data
    What data is needed? Where does it currently live? Who would need to authorize using it?
  3. Then write Risk
    If the AI output is wrong, what is the worst case? What level and whose review would catch it?
  4. Present the three separately
    Keep each paragraph independent and let the reviewer weigh them; never convert them into one combined score.
Workplace example (contrast case: a data concern hidden by a total score)

A unit evaluates "use AI to analyze course-enrollment data and recommend courses". Value looks high (it might improve satisfaction with course selection), so the team gives it a "90/100, recommend proceeding" score. But the Data dimension actually has a serious concern: enrollment records are sensitive academic data, and no clear authorization to use them exists yet. Merged into one score, this data concern is easily buried under the high Value number; stated separately, a reviewer sees immediately that "Data needs an authorization question resolved first".

Info

Human, AI, Human: a person confirms the problem and judgment first, AI assists only after that, and a person verifies and owns the final result.
Check yourself

Which approach best follows this lesson's "no opaque total score" principle?

Which approach best follows this lesson's "no opaque total score" principle?

Key takeaway: Always write Value, Data, and Risk out separately; any automated suggestion must stay explainable and never make the final call for a person.

Use Case Sprint

30 min

Use Case Sprint

Learning objective: Complete the formal Canvas, then peer-review it for a vague problem, a smuggled-in solution, or a missing owner/review step.

This is the course's capstone activity: integrate everything from the first five lessons (problem framing, friction, owner/outcome, AI pattern, value/data/risk) into the formal AI Use Case Discovery Canvas. After completing it, peer-review each other's Canvas for three things: is the problem statement concrete and observable, has a tool crept back into the problem statement (solution bias relapse), and are the owner and review step clearly named.

Three peer-review checks
FieldMeaning
Problem specificityIs the problem field observable and free of tool names?
Solution-bias relapseDid a solution sneak back into any field?
Owner / review completenessAre the owner and human_review fields explicit?
Workplace example

A peer-review checklist: (1) does the problem field mention a tool name? (2) are owner and user filled in as two different people without explaining the relationship? (3) is the next_question field just "keep going" instead of naming the real open question that needs clarifying?

  • Problem: is it free of tool names and observable enough for a third person to verify?
  • Friction signal: does it name the specific type, frequency, waiting, rework, errors, searching, duplicate entry, or judgment load?
  • User / Process owner / Decision owner: are all three filled in, with the relationship between them clear?
  • Desired outcome: is it a measurable result, or did an activity or tool creep back in?
  • AI pattern: does the chosen pattern actually match the problem, rather than picking a favorite tool first and fitting the problem to it afterward?
  • Value / Data / Risk: are the three stated separately, with no merged total score?
  • Owner / human review: are both who drives the next step and who gives final review named?
  • Next question: does it point to a real open question, instead of a vague "keep going"?
Workplace example (contrast case: a Canvas that passed review)

Another team's Canvas: the problem field says "leave requests take an average of 4 business days from submission to approval, often delayed because the supervisor cannot be reached"; owner reads "process owner: HR office; decision owner: unit supervisor"; next_question asks "need to confirm whether there is a delegated approver while the supervisor is traveling". The reviewers agreed this Canvas was concrete, free of solution bias, and had a clear owner and next step, so it could go forward as is.

Info

Human, AI, Human: a person confirms the problem and judgment first, AI assists only after that, and a person verifies and owns the final result.
Check yourself

During peer review you see a classmate's Canvas problem field says "we should use AI to classify complaints". What feedback is most appropriate?

During peer review you see a classmate's Canvas problem field says "we should use AI to classify complaints". What feedback is most appropriate?

Key takeaway: Peer review right before formal submission is the last, and most effective, chance to catch solution-bias relapse and missing owner/review fields.

Save as PDF

What next?