Module 2 | From One-Off Task to Repeatable AI Use Case | AI 101L

Human-AI-Human Workflow

35 min

Human-AI-Human Workflow

Objective: design a Human-AI-Human workflow with every step tagged Human / AI / Tool, and define verification and escalation for each step.

Module 1 practiced context and execution for a single task. Module 2 grows from "one task" to "a workflow": chaining several steps together, each one clearly tagged with who executes it, what it does, and how it is verified.

Five fields of a workflow step
Field	Meaning
Input	This step's input
Action	What this step actually does
Output	What this step produces
Verification	How to confirm this step was done correctly
Escalation	When to escalate or pause instead of continuing

Just "AI handles it"

"AI processes the meeting notes": no idea who is supposed to check it or when to stop.

A complete workflow

Step 1 (Human): Upload the approved meeting notes
Step 2 (AI): Extract action items
Step 3 (AI): Format into a tracker
Step 4 (Human): Confirm owner and due date
Step 5 (Tool): Update the tracker
Step 6 (Human): Approve distribution

Info

AI 101L only teaches the basic Human / AI / Tool tagging of roles; decomposition, sequential / parallel / routing patterns, decision points, and the agent loop are covered in depth in AI 104.

Check yourself

In the flow "AI formats the tracker -> a human confirms owner and due date -> a tool updates the tracker," which step catches errors best?

The AI formatting step, because the formatting is usually tidy.The human confirmation step, because it checks the data before it enters the production system.The tool update step, because tools rarely make mistakes.

Key takeaway: A trustworthy workflow tags every step with its executor and verification method, instead of treating "AI handles it" as one big black box.

Verification and Failure Modes

35 min

Verification and Failure Modes

Objective: verify AI output using five questions (Accuracy, Completeness, Source, Sensitivity, Consequence), and define who checks, when to stop, and when to escalate.

Five verification questions
Question	What it checks
Accuracy	Is the content factually correct, with no hallucinated details?
Completeness	Is anything important missing?
Source	Can the conclusions be traced back to the original data?
Sensitivity	Does the content involve confidential or personal data?
Consequence	What happens if this part is wrong?

Define what to check first: which fields or claims are most error-prone?
Then define how to check it: line-by-line comparison against the source, or spot checks?
Define who checks it: the task owner, or does it need a more senior approver?
Define when to stop: what level of checking counts as done?
Define when to escalate: what should trigger a pause and a report up, instead of deciding to proceed alone?

Warning

The same process steps can carry very different consequences, so verification intensity should differ too. A summary that only goes to a department chat and one that affects external commitments or financial decisions cannot use the same loose standard.

Check yourself

Two "summarize internal feedback" tasks look alike; one is only shared in a department chat, the other will be used to adjust compensation policy. Which statement is correct?

Both can use the same loose verification approach.The compensation-policy one has a more severe consequence and should get stronger verification and more human oversight.As long as both are formatted neatly, verification does not need to differ.

Key takeaway: Verification isn't a yes/no question: it's a question of how much; the more severe the consequence, the more rigorous verification and human oversight need to be.

Data Boundaries and Responsible Use

30 min

Data Boundaries and Responsible Use

Objective: classify information as Public / Internal / Confidential / Personal / Restricted, and judge what should not be given to AI.

Five information classifications
Classification	Example
Public	Already published information, e.g. a website announcement
Internal	General information for internal use only, not yet public
Confidential	Strategy, contract terms, unpublished financial figures
Personal	Names, contact details, national ID numbers, or other identifying data
Restricted	Highly sensitive information accessible only to specific people

Warning

Core rule: do not paste secrets, credentials, personal data, client confidential information, or other unauthorized sensitive information into an AI tool. This rule is not tied to any specific vendor or tool: it applies to every AI collaboration.

Before giving data to AI, ask: which classification does this fall into?
Confidential, Personal, and Restricted content is not provided unless the task clearly needs it and it is authorized
When the classification is unclear, treat it as the more sensitive level and confirm with the data owner
When sensitive data is genuinely needed, prefer de-identification, placeholders, or a summarized description first

Check yourself

You want AI to help analyze an employee satisfaction survey; the raw data includes employee names and departments. Which approach best follows the data-boundary principle?

Give the whole file to AI as-is: it is more convenient.Remove or pseudonymize identifying data like names first, keeping only the fields the analysis actually needs.Keep only the names and delete the other fields.

Key takeaway: Classify first, then decide whether it can be used; data boundaries are a habit to re-check every time you prepare context, not a one-time rule.

From Task to Use Case

30 min

From Task to Use Case

Objective: judge whether a one-off task is worth turning into a repeatable, formal Use Case, using frequency, benefit, and risk.

Completing a task once does not automatically make it a good Use Case. Moving from "done once" to "worth repeating" needs a fuller judgment.

Questions for moving from task to Use Case
Question	What it confirms
How often does it happen?	Whether the frequency justifies the investment
How much time does it take now?	The current cost baseline
What is the current pain point?	The problem actually worth solving
What is the expected benefit?	What roughly gets saved or improved
What data is needed?	Whether the context is stable and available
How is it verified?	Whether the result can be confirmed as correct
What is the consequence if wrong?	The risk level and the oversight it needs
Who is the owner?	Who is accountable for this Use Case's quality
What is the next experiment?	How to test feasibility with a small trial

Info

This is the individual-work perspective: where in my own work could AI help? The department- or organization-level question of "which problems deserve to become formal AI Use Cases" is developed further in AD 101.

Check yourself

Which piece of work is the best candidate for "worth developing into a Use Case"?

A task done only once, that cannot be repeated, and whose result cannot be verified.A task done every week that currently takes a lot of time, where AI-assisted quality can be concretely verified and the risk is manageable.A task involving unauthorized confidential data with no verification mechanism at all today.

Key takeaway: One smooth run does not mean it is worth repeated investment; judge it with frequency, benefit, verifiability, and risk together.

Independent Use Case Sprint

50 min

Independent Use Case Sprint

Objective: independently complete an AI Workforce Use Case Canvas, integrating everything from Module 1 and Module 2 into one finished output.

Info

This is an independent work session with instructor support on hand; there is no automatic business-value judgment. Your job is to write up one real task from your own work as a complete Use Case.

Pick a real work task
Choose one task from your own work that you genuinely want AI to help with.
Fill in the work task and baseline
Write out the department, task, trigger, current process, frequency, time spent, and pain points.
Fill in context and AI assistance
List the data needed, its sensitivity, and what you want AI to help with plus the expected output.
Design the workflow
Write out at least three workflow steps, each tagged with its executor.
Fill in verification and risk
Write out the reviewer, review method, consequence if wrong, and risk considerations.
Write your in-class experiment and key reflection
Record what you actually tried today, and which step is most and least suitable for AI.
Fill in the follow-up plan
Write down a suitable department-interview topic and your recommended next course.

Go to the AI 101L Apply page

Done

After completing the Use Case Canvas, remember to fill in the Exit Ticket to record today's most important learning and what you're still unsure about. This Use Case becomes the starting point for later department interviews and AI-adoption discussions.

Key takeaway: A complete Use Case Canvas is the most concrete proof of turning "knowing how to use AI" into "being able to put AI to work in real work."

From One-Off Task to Repeatable AI Use Case

Learning objectives

Human-AI-Human Workflow

Human-AI-Human Workflow

Just "AI handles it"

A complete workflow

Verification and Failure Modes

Verification and Failure Modes

Data Boundaries and Responsible Use

Data Boundaries and Responsible Use

From Task to Use Case

From Task to Use Case

Independent Use Case Sprint

Independent Use Case Sprint

What next?