Video version: about 15 min
Opening and Concepts
About 5 minutesLet's first look at Lin Yue's situationFictional example; numbers are illustrative
Weekly department meeting minutes, issued once a week to 12 colleagues. Previous 7 sessions averaged about 90 minutes each.
Common issues include missing action items, incorrect assignees, and missing deadlines, with minutes often sent out the next day.
This time, she pasted the entire meeting transcript to AI, saying, 'Help me organize the minutes.' The minutes looked good, but two action items lacked assignees, and one assignee was guessed by the AI.
A task is a process
First, declare the task category so the review focus aligns correctly.
Organizing
Transforming a set of inputs into a fixed-format output, such as meeting minutes or weekly reports. The key is to check for omissions or fabricated content.
Q&A
Answering questions on behalf of colleagues or clients. This lesson mentions this only briefly; for details, see Appendix 2 in C07 and read it after class.
Action
Tasks involving system changes or external communications. This lesson covers only drafting; C08 will explain the reasoning behind execution.
The default workflow is semi-automated.
Mapping the Process
About 14 minutesExample: Answer these three questions first, and write them down.
Write out a text-based flowchart: each step must have a deliverable and a completion criterion.
Change 'topics' to 'steps'
Topics
- Organize action items
- Review
- Send as soon as possible
- Apply formatting
Steps
- Each to-do item must have an owner and a deadline; mark missing ones as [To Be Supplied].
- Spot-check three conclusions against the original records; all three must be traceable to the original text.
- Send it before end of business on the day of the meeting; record the send time in the email.
- Ensure all four sections are present, in a fixed order, and within one page.
Exercise A: Map your own process.
Have the AI turn your method into a process, and have it ask you questions in return.
Suggested timeI want to break a task into a process. Please act as a process consultant with ten years of experience and help me with three things: 1. Organize the method I describe below into three sections: Input, Processing, and Output. Number the steps in the Processing section. 2. For each step, write the 'deliverable' and the 'completion marker.' The completion marker must be a sentence that allows others to judge whether the step is complete. If a step cannot have a completion marker, mark it as 'Topic, needs breaking down.' 3. Mark checkpoints where human review is needed, and write a sentence for each checkpoint specifying 'what to check.' If my description lacks key information (e.g., where the input comes from, who receives the output), ask me questions one by one; do not fill in the gaps yourself. Base your response only on my description; do not add steps I haven't mentioned. My description: [Write a few sentences describing how this task is currently done]
Human-Machine Division of Labor Color Chart
About 11 minutesHuman-Machine Division of Labor: Three-Color Rules
Green: AI does
Mechanical tasks with clear rules and easily detectable errors, such as organizing, categorizing, extracting, applying formats, or drafting.
Yellow: Human does
Tasks requiring judgment or interaction with systems and individuals, such as confirming with the individual, making decisions, or replacing codes.
Red: Human must review
The AI can draft, but the final review must be done by a human. Anything sent externally, involving money, contracts, or clients, as well as dates and numbers, is always Red.
What to write in each row of the assignment table
Lin Yue's Assignment Table (Part 1)
Lin Yue's Assignment Table (Part 2)
Are too few items marked red, or too many?
Marked too few
Red is only one step away from being 'sent'
- Responsibilities and deadlines are intermediate steps that also need review, but they weren't highlighted in red
Marked too many
Everything requires human review
- The time-saving benefit of AI is lost
- Reserve red only for steps with real consequences
Let AI act as the auditor; you decide the colors
Suggested timeHere is my process and initial division of labor between humans and AI. The rules for division are: Green means AI handles it; Yellow means human handles it; Red means humans must review (AI can draft, but humans must confirm). Any content sent externally, involving money, contracts, or clients, as well as dates and numbers, is always marked red. Please act as a picky internal auditor and check step by step: 1. Are there any steps that should be red but are marked green? Explain why. 2. Are there any green steps where errors are hard to detect? 3. Are there any steps requiring human review where the check content I provided is too vague, or the completion criteria are unjudgeable? Do not change my colors; only raise questions. [Paste your process and division of labor table]
Step-by-Step Execution and Acceptance Checklist
About 12 minutesStep by step: Declare the category first, then proceed step by step
Suggested timeThe task category you are handling is: Meeting Minutes Organization (Organizing type). Please strictly follow the four steps below, proceeding step by step. After completing each step, output the result for that step and indicate the step number. Step 1: List the actual decisions made in the meeting. Only write what was explicitly stated; do not write what was not explicitly stated. Step 2: List all action items. Each action item must be written in three columns: 'Item, Responsible Person, Deadline.' For responsible persons or deadlines not mentioned in the record, fill in '[To Be Confirmed]'. Do not speculate. Step 3: Output the draft minutes using the format below. Step 4: List the 'Checklist for Lin Yue to verify,' separately listing all '[To Be Confirmed]' items, content involving amounts, and content involving clients. Stop at the current step and ask me directly if any of the following occur; do not proceed further: 1. The record contains two contradictory statements. 2. You cannot determine if a statement is a decision. 3. The record contains amounts, contract terms, or specific client information, which I did not ask you to handle. When exceeding your processing scope (e.g., budget approval, contract terms, personnel evaluations), explain the reason and specify who it should be transferred to. Minutes Format: [Meeting Minutes] Date: [Fill in] Host: [Fill in] I. Conclusions II. Action Items (Item | Responsible Person | Deadline) III. Items to Confirm Here are the meeting notes: [Paste the desensitized meeting notes here]
On-site acceptance: Discover omissions, do not fill them in yet, ask questions firstInterface Illustration: Fictional Dialogue
Re-ran the same input after changing the rulesFictional example; numbers are illustrative
Before changing the rules
- Three pending items
- "Get the basics" tasks don't count as pending items
After changing the rules
- Four pending items, with Ding Na's item added
- Ding Na's item has a deadline marked as [To Be Supplemented], not guessed as a specific day
Write an acceptance checklist: each item can be checked or unchecked
- Each item is a sentence that can be checked or unchecked
- Each item specifies the verification method
- Do not write uncheckable terms like 'high quality' or 'clear'
- Includes a gatekeeping check from C04, such as verifying numbers, dates, and names
See a sample (a fictional Lin Yue; the numbers are illustrative)
Each pending item has an owner and a deadline; missing ones are marked [To Be Supplemented]. Verification method: review each item in the three columns The conclusion contains no statements not mentioned in the meeting. Verification method: spot-check three items against the original notes Dates and days of the week match the calendar. Verification method: check the calendar
| Task | Minutes each time | Times per week | Weekly total |
|---|---|---|---|
Self-check: how confident are you right now, on a real task, that it can help?
Exception Handling and Handover Checklist
About 9 minutesException and Handover Checklist: A working process does not guarantee real-world execution
For each exception, clearly document four key points.
How to identify them
Specify the exact signals, such as 'To be supplemented' appearing in the checklist.
Who handles it, and who receives the handoff
For tasks beyond the assistant's capabilities, specify exactly who handles them; do not simply write 'handled by a person'.
What to do
Try a few times; if it still fails, what is the fallback plan?
How to record it afterward
Record one line; otherwise, the process won't be visible when updates are needed.
Q&A-style assistants need two additional components
Use AI to identify potential pitfalls you might have overlooked
Suggested timeI am designing a 'Meeting Minutes Summary' workflow: the input is anonymized meeting notes, the AI extracts conclusions and action items, and a human verifies the responsible person and deadline before sending. Please act as an operations manager with ten years of experience and list the five most likely pitfalls in this workflow. For each, specify: how it’s detected, who handles it, who it’s escalated to, and what to do. Only list common, real-world scenarios; do not invent rare edge cases. [Paste your workflow summary here]
Criteria for reviewing the exception list
Quiz
About 4 minutesThree quick questions to check yourself
Question 1: What does 'Red' represent in the Three-Color Role Assignment Table?
Question 2: What is the issue with a process step labeled 'Check'?
Question 3: Which of the following acceptance criteria is written correctly?
Assignment and Wrap-up (Including Buffer Time)
About 5 minutesKey takeaways from this lesson: five essential phrases.
- A single task equals one workflow: input, processing, output, with checkpoints in between.
- Steps without a completion marker are topics; break them down into steps.
- Three-color division of labor: Green for AI, Yellow for humans, Red for mandatory human review. Red tasks are never handed off entirely.
- The acceptance checklist allows each item to be checked or unchecked; it also serves as the yardstick for polishing C07.
- Exceptions must specify who they are transferred to; the default route is semi-automated.
Next lesson: Create your first dedicated assistant.
Dragon Warrior plus script plus dossier.
- Bring the workflow diagram, three-color chart, and acceptance checklist.
- Prepare 7 past real inputs (desensitized) and the corresponding manual outputs from that time.
Show the full transcript (video narration)
Breaking Down Processes: Input, Processing, Output, and Human-AI Division of Labor
Welcome to Lesson 6 of the AI Implementation Course. In the previous lesson, we selected a task worth delegating to AI. This lesson addresses the question: What steps make up this task, which steps should be assigned to AI, and which must be done by humans? All case studies in this lesson are fictional, and the numbers are illustrative.
Let's first look at Lin Yue's situation
Let's first look at Lin Yue's situation. This is fictional, and the numbers are illustrative. She is an administrative specialist at Yuanfan Technology. The department holds weekly meeting minutes once a week, which need to be sent to 12 colleagues. Previous minutes averaged about 90 minutes per session, with common issues including missing action items, incorrect assignees, and missing deadlines. This time, she pasted the entire meeting transcript to AI, saying, 'Help me organize the minutes.' The minutes looked good, but two action items lacked assignees, and one assignee was guessed by the AI. The problem isn't with the AI; it's that the entire task was handed over without being broken down.
A task is a process
Any task can be broken down into three stages: Input (what you receive), Processing (what you do), and Output (what you deliver). Between these stages, place a checkpoint—a point where someone must review the work. Once broken down, you can identify which steps are mechanical and can be delegated to AI; which step is most prone to error and requires a checkpoint; and how much time each step takes, so you can calculate costs in Lesson 9.
First, declare the task category so the review focus aligns correctly.
Different categories require different review focuses, so you must declare the category first. For Organizing tasks, transforming inputs into fixed-format outputs like minutes or reports, the key is checking for omissions or fabrication. For Q&A tasks, answering questions on behalf of colleagues or clients, you need two additional elements (see the appendix for after-class reading). For Action tasks, involving system changes or external communications, this lesson covers only drafting, not autonomous execution.
The default workflow is semi-automated.
Automation is divided into three levels. Level 1 is manual, with fixed locations; Level 2 is semi-automated, which is our default workflow; Level 3 is fully automated, requiring company approval (discussed in C08). Achieving semi-automation in this course is considered successful. When mapping the workflow today, label each step as either manual or semi-automated.
Example: Answer these three questions first, and write them down.
Step 1: Answer three questions and write them down. What is the input? What do I have in hand before starting? What is the output? What is the final deliverable, and who receives it? What happens in between? From input to output, what manual steps are currently performed? Lin Yue's answer: The input is meeting notes, which are first anonymized; the output is a summary sent to 12 colleagues; the manual steps involve summarizing conclusions, identifying action items, assigning owners and deadlines, applying formatting, reviewing, and sending.
Write out a text-based flowchart: each step must have a deliverable and a completion criterion.
Next, write these steps as a text-based flowchart: one step per line, connected by arrows, without using diagramming software. For each step, specify two things: the deliverable (what is produced) and the completion criterion (what indicates the step is finished). After anonymization, the notes should contain no phone numbers, amounts, or client names. Summarizing conclusions, extracting action items, and applying formatting are drafted by the AI; action items must have all three columns complete, with missing items marked as [To be supplemented]. Then comes manual review, where Lin Yue verifies owners and deadlines, determines which are actual decisions, and ensures amounts and client-related content are reviewed by a human. Finally, she sends it herself.
Change 'topics' to 'steps'
The difference between a topic and a step is the presence of a completion marker. Review your to-dos: if you can't say when it's done, it's a topic. To make it a step, ensure each to-do has an owner and a deadline, marking missing ones as [To Be Supplied]. Verify by spot-checking three conclusions against original records, ensuring all three are traceable to the original text. Change 'send ASAP' to 'send before end of business on the day of the meeting.' Change 'fix the format' to 'ensure all four sections are present, in a fixed order, and within one page.'
Exercise A: Map your own process.
Now, do Exercise A. Map your chosen task into a process using the steps above. If you don't have your own example, use Zhou Kai's sales weekly report. He compiles customer follow-up records into a weekly report every Friday and sends it to the department manager. Check against the four criteria: it has Input, Processing, and Output sections, with numbered steps in the Processing part; each step specifies a deliverable and a completion marker; it includes at least two checkpoints specifying what to look for; and it identifies where the bottleneck in the baseline falls within the flowchart.
Have the AI turn your method into a process, and have it ask you questions in return.
You can have the AI organize your verbal description into a process. The key is three requirements: base it only on your description without adding unmentioned steps; mark steps without a completion marker as topics that need breaking down; and ask for missing information one by one without filling in the gaps yourself.
Human-Machine Division of Labor: Three-Color Rules
The rules for the three-color table are simple. Green: AI does mechanical tasks with clear rules and easily detectable errors, such as organizing, categorizing, extracting, applying formats, or drafting. Yellow: Human does tasks requiring judgment or interaction with systems and individuals. Red: Human must review; the AI can draft, but the final review must be done by a human. Anything sent externally, involving money, contracts, or clients, as well as dates and numbers, is always Red. Red steps are never handed over entirely.
What to write in each row of the assignment table
Each row of the assignment table must include five elements: step number and name; who does it; color; why (written as 'what happens if it goes wrong,' not 'because it's important'); and what I check, along with the completion criteria. When determining what to mark red, first consult the company's AI usage policy. Any step explicitly prohibited by the company from being handled by AI must be marked red, with the basis stated. Company policy prevails.
Lin Yue's Assignment Table (Part 1)
Looking at Lin Yue's assignment table, this is fictional. De-identification is yellow, requiring her to judge what counts as sensitive. Organizing conclusions is green, handled by AI. Extracting action items is red, as incorrect responsible persons or deadlines cause errors; AI drafts, and she must review. Applying the format is green.
Lin Yue's Assignment Table (Part 2)
Determining actual decisions is yellow, requiring understanding of the meeting. Content involving amounts and clients is red, requiring sentence-by-sentence verification against the original transcript. Restoring real names and sending is also red, sent on the same day with no mix-ups. Step 7, 'Send,' is always handled by a human.
Are too few items marked red, or too many?
There are two common biases when marking red incorrectly. Marking too few means red is only the final step; intermediate steps like responsibilities and deadlines, which also need review, are missed. Marking too many means everything requires human review, and the time-saving benefit of AI is lost. Reserve red only for steps with real consequences. The reason must be written as 'what happens if it goes wrong,' not just 'it's important.'
Let AI act as the auditor; you decide the colors
After writing the division of labor table, you can have AI act as a picky internal auditor to check step by step for steps that should be red but are marked green, or green steps where errors are hard to detect. Note that it only raises questions and does not change your colors. You decide the colors.
Step by step: Declare the category first, then proceed step by step
This prompt has three parts. First, proceed in four steps, outputting the result for each step and indicating the step number. Second, stop and ask humans when encountering contradictions, uncertainty about whether something is a decision, or the appearance of amounts or client information. Third, for matters out of scope, explain the reason and specify who it should be transferred to. Its expectations are specific: if an action item lacks a deadline, write 'To Be Confirmed'; do not guess.
On-site acceptance: Discover omissions, do not fill them in yet, ask questions first
When accepting on-site, if you discover omissions, do not rush to fill them in. Count: there are four things in the record that require human action, but only three in the action items. Ask it first: why wasn't Colleague D's item included? What assumptions in your rules would cause an item not to be counted as an action item? It replied that it treated 'get to know' as something not yet formed. It didn't make a mistake; it had an undiscovered assumption. Find it, write it into the rules. You are changing the rules, not just fixing this single result.
Re-ran the same input after changing the rules
Re-ran the same input after changing the rules. This time, the pending items included Ding Na's item, with the deadline marked as [To Be Supplemented] instead of a specific day. After success, don't rush to celebrate. Changed another condition, such as adding a note to 'provide feedback by Thursday' and re-running only that input, to check if the new item is correct and if other entries were affected. If the change didn't take effect, the most common reason is that it wasn't saved; confirm before re-running.
Write an acceptance checklist: each item can be checked or unchecked
Each item in the acceptance checklist must be checkable or uncheckable, with a specified verification method. Do not use uncheckable terms like 'high quality' or 'clear'. This list will be very useful later: when polishing the script in C07, re-run the same input for each change, then check off items one by one. Only when the number of checked items increases is the change considered complete. Therefore, do not modify it casually after writing it.
Exception and Handover Checklist: A working process does not guarantee real-world execution
A process that works on paper doesn't necessarily work in reality, so you must plan for potential issues in advance. Here, we display the first five rows of Lin Yue's checklist. If documents are missing, the checklist shows 'To be supplemented'; confirm in person before proceeding. If documents contradict each other, the AI pauses to ask questions, deferring to the host's judgment. If unsure whether a decision is required, Lin Yue makes the call. If the format is incorrect, send it back for reprocessing; if it still fails, handle it manually. For budget-related amounts, forward to the finance colleague, copying the original amount exactly, as the AI does not perform conversions. Review the remaining four rows after class.
For each exception, clearly document four key points.
For each exception, clearly document four things: how it’s detected (e.g., an item appears in the checklist as pending completion), who handles it (escalate to someone else if it’s beyond your capability), what to do (try a few times; if it still fails, what’s the fallback plan?), and how to record it afterward. An exception with no assigned owner is effectively unhandled.
Q&A-style assistants need two additional components
If your role involves Q&A support—meaning the assistant answers questions on your behalf for colleagues or clients—the exception list must include two additional items: an FAQ library listing the 20–30 most common questions with standard answers and sources, and a handback protocol specifying which questions are not to be answered, who they should be escalated to, and what script to use. Details are in Appendix 2 of C07; read this after the course.
Use AI to identify potential pitfalls you might have overlooked
Finally, you can ask the AI to act as an operations manager and list the five most likely pitfalls in this workflow, specifying for each how it’s detected, who handles it, who it’s escalated to, and what to do. Whether to adopt them is your decision; only list real-world scenarios.
Criteria for reviewing the exception list
Review the exception list for four key points. There should be at least five exceptions, each with four components. One must be a fallback for manual processing. One must address what to do and who to notify if inappropriate content is found. Each scenario beyond your authority, such as budget, contracts, or staffing, must specify the escalation contact. A common bias is focusing only on AI errors while missing data issues or human oversight.
Quiz Question 1
Question 1. What does 'Red' represent in the Three-Color Role Assignment Table? Options: A) An error occurred at this step; B) The AI can draft, but a human must review and approve before proceeding; C) This step must be performed entirely by a human; AI cannot be used. Answer: B. Red means human review is mandatory, not that AI is banned, but that it serves as the final checkpoint.
Quiz Question 2
Question 2. What is the issue with a process step labeled 'Check'? Options: A) It lacks deliverables and completion criteria; it is a topic, not a step; B) The word count is too low; C) It should be assigned to the AI. Answer: A. If you cannot define what 'done' looks like, you cannot verify completion or estimate the time required.
Quiz Question 3
Question 3. Which of the following acceptance criteria is written correctly? Options: A) The meeting minutes are clear; B) Each action item has an owner and a deadline, with missing items flagged; C) The quality is high. Answer: B. Acceptance criteria must be binary (pass/fail); 'clear' and 'high quality' are not verifiable.
Key takeaways from this lesson: five essential phrases.
Finally, take away five sentences from this lesson. First, a single task equals one workflow: input, processing, output, with checkpoints in between. Second, steps without a completion marker are topics; break them down into steps. Third, the three-color division of labor: Green for AI, Yellow for humans, Red for mandatory human review, and red tasks are never handed off entirely. Fourth, each item on the acceptance checklist can be checked or unchecked. Fifth, exceptions must specify who they are transferred to, and the default route is semi-automated. Post-class tasks: workflow diagram, three-color chart, acceptance checklist, exception list, and use the acceptance checklist to review 7 past manual outputs to fill in any gaps in the baseline.
Next lesson: Create your first dedicated assistant.
In the next lesson, we will turn today's workflow into a truly usable dedicated assistant, which is Dragon Warrior plus script plus dossier. Please prepare 7 past real inputs in advance, desensitized, and bring them along with the corresponding manual outputs.