Your manager does not ask a clean “AI question”
A manager says: “Can you use AI to help with these three things today — draft a customer FAQ, decide a refund, and write a performance note?”
The trap is that these look like three writing tasks. They are not. One is safe for AI drafting, one needs human approval, and one should not be automated.
Bad example first: the answer looks useful, but the boundary is wrong
Weak prompt
“Use AI to finish the FAQ, refund decision, and performance note.”
Wrong focus: it treats writing, money decisions, and people decisions as the same kind of task.
What goes wrong
AI may produce fluent text, but it can also approve money, invent policy, or write people-impacting judgment without the right owner.
Work-ready version: split the request before prompting
AI may assist
Customer FAQ draft
AI can draft wording from approved notes. A person checks facts before use.
Human review required
Refund approval
AI can summarize case details, but the support manager owns the money decision.
Do not automate
Performance note
AI may help outline neutral questions, but a manager/HR owns judgment and wording.
The 3-label rule
Before you write a prompt, label the task as one of three types. This is the small work habit you are practicing today.
AI may assist
AI drafts, organizes, summarizes, or proposes questions. A person still checks the result.
Human review required
AI can prepare the case, but a named human owns approval, policy, money, or customer impact.
Do not automate
The task affects hiring, discipline, legal, medical, finance, trust, or people judgment.
Boundary judgment trace
This is the small visible work change from today. Keep it shallow, real, and reusable tomorrow.
Before / after: what changed in your work?
A small example is enough. The point is not to design a full workflow today. The point is to see the work step become safer and clearer.
Before learning
After learning
What AI-in-workflow means today
Build today's artifact rows
Complete three short fill-ins. Keep one real work item per row, and change the sample wording so it fits your work.
| No. | Work item | AI role / prompt case | Risk or evidence | Human owner / reader | Next action |
|---|---|---|---|---|---|
| 1 | |||||
| 2 | |||||
| 3 |
Repair one weak row
Use this pass to find the row that could confuse someone, create unsafe AI use, or fail as workplace evidence before you export.
Repair example
Weak row: Work item: Customer issue / AI role: AI handles it / Owner: Me / Next action: Send.
Repair move: Split the task into AI draft versus human decision before using AI.
Better row: Work item: Refund request reply / AI role: draft neutral reply only / Owner: support manager approves refund / Next action: no money decision by AI.
Choose the row you would be comfortable showing to a manager. It should have a clear task, boundary, owner, and next action.
Choose the row most likely to fail, confuse someone, expose sensitive input, or need review.
Write the exact change. Sentence starter: The weak row is weak because ___. I will repair it by adding ___ before using AI.
Pass / fix / stop standard
Do not let AI make person-impacting, money, policy, or trust decisions.
Export artifact
This file is today’s work receipt. It should show what changed in your work habit, not just that you filled a table.
Before you export
- Does the artifact show what AI may draft?
- Does it name the human decision owner?
- Does it clearly stop AI before money, policy, people, or trust decisions?
What this artifact proves: This file proves you can split a messy work request into AI assist, human review, and do-not-automate boundaries.
Weak export: vague rows, no owner, no repair note, or no next-use line.
Good export: one clear work case, one boundary/rule, one repair note, and one tomorrow-use or manager-use line.
Click Generate Markdown to create ai_task_boundary_note.md.