TL;DR: the job itself needs to be clearly defined. That means knowing what the job is, giving the AI the right business context, defining what a good job looks like, and deciding when the AI can proceed on its own and when it should ask for human judgment. This article walks through five reusable assets that help teams build workflows that AI can consistently and confidently support.
In my last article, Redesigning the Workflow Before Adding More AI Agents, I argued that companies should start by redesigning the workflow before rolling out more AI tools and agents.
You may be wondering, if creating more agents is not the first step, where should we start? If a team wants to move beyond individual products, one-off pilots, and scattered demos, and make AI part of the product lifecycle and daily operations, what needs to happen first?
The more I work on projects to enable AI workflows, the more convinced I am that the missing piece is optimizing the workflow itself.
Before an AI can faithfully perform any emerging task, someone must define what the task is, why it exists, what information is important, what a successful outcome looks like, and where the AI should stop and ask for human judgment.
The reality today is that this arrangement is rarely written down. Many groups jump straight into introducing agents. Many believe that giving teams access to advanced models will produce better results. Here's the flip side: the better the model works, the more valuable those missing explanations can be.
The difference is how you ask. The conversation begins with the request: “Analyze this.” “Make an introduction.” “Compress these files.”
The program starts with a job well done. What effect is required? What sources are authoritative? What decisions can AI make on its own? What does the transient output look like? When should AI stop and ask for help?
In this article, I want to show what you can do before handing over work to an AI assistant or agent. For your team, work can begin with these five reusable supplies. Instead of continuing to collect the best information, create these assets, improve them using them, and move them to all the AI models and tools you adopt over time.
1. Duplicate Work Assets
Identify tasks that occur frequently, take reasonable time, follow repeatable steps, use the same types of inputs, or manage enough value or risk to ensure a reusable AI workflow.
It could be a weekly report, a monthly business update, a customer proposal, a contract review, a product launch package, or a quarterly planning process.
You need a simple inventory of how you and your team spend your time, what tasks are repetitive, and where a reusable process can help. A duplicate task list helps you select tasks based on frequency, effort, risk, and value.
You are my workflow organization assistant.
Based on the work description below, identify the recurring tasks that are most suitable for AI.
Requirements:
1. Include tasks that:
- Repeat regularly
- Follow a consistent process
- Consume meaningful time
- Carry a high risk of avoidable errors
- Or depend on repeated analysis, drafting, review, or coordination
2. For each task, state:
- The specific action
- How often it happens
- Required input
- Expected output
- Evaluation standard
- Time currently required
- Main source of difficulty or error
3. Classify each task as:
- Better suited for a one-time AI conversation
- Better suited for a reusable workflow or agent
- Better kept primarily human-led
4. Explain why you selected each classification.
5. Avoid general advice. Use specific actions.
My work description:
[Paste your description here]2. Legacy of Work
Most people provide the title and leave it all to AI. They ask AI to “analyze data,” “prepare a presentation,” or “summarize documents.” The model then must consider audience, decision, format, source value, level of detail, and quality threshold.
AI often fills in missing information on its own. If that assumption is wrong, your output may sound confident while still pointing in the wrong direction.
Functional inheritance helps to remove these hidden concepts. It turns a vague request into a task that AI can perform. Define the purpose, the audience, the materials, the obstacles, the steps, what it looks like, and when the program should stop and ask. It also gives you something concrete to review before giving the AI more responsibility.
Convert the vague request below into a task package that AI can execute.
Output:
- Objective
- Business purpose
- Audience
- Decision or action this work should support
- Materials to use
- Authoritative sources
- Reference-only sources
- Constraints
- Execution steps
- Required output format
- What a good result looks like
- Acceptance criteria
- Risks I need to confirm
- Information that is still missing
- When you must stop and ask me
My request:
[Paste your request here]3. Inheritance of Content
Context helps AI see what's important. Teams can reorganize, priorities can change, results are updated, policies are rewritten, and information you relied on six months ago may be out of date. A useful context tells the AI what it knows and what it still needs to confirm.
Stop redefining your job in every interview. One short document can cover who you are, what you do, what sources you can trust, how you make decisions, and what kind of product you won't want.
A useful context should not be the kitchen sink at all. Keep it short and to the point. Don't dump months of old chat history into context, because outdated information and unrelated conversations can bury valuable information.
Create a concise project context document for AI.
I will reuse it in ongoing AI tasks so I do not need to explain the same context each time.
Include:
1. Who I am
2. What I am currently working on
3. My current objective
4. My target audience
5. The decisions I am trying to support
6. How I usually work
7. The tools and materials I commonly use
8. My preferred output style
9. The types of output I dislike
10. Things I cannot say, publish, share, or do
11. Which sources are trustworthy
12. Which sources are for reference only
13. Important definitions or business rules
14. Facts that may expire or change
15. Information that must be confirmed before use
16. The date this document was last updated
Keep it short, precise, and relevant.
Separate stable information from temporary information.
Flag anything that may need to be updated within the next 30, 60, or 90 days.
My background:
[Paste your information here]4. Material Acceptance Test
You need to know what failure looks like before the AI output reaches the customer, the production system, or the public. Test the AI agent against existing prototypes before using it for ongoing work. Use examples from your own work, including those you have adopted and given birth to.
Acceptance testing turns your expectations into something you can test. They show both you and the AI what a good result looks like. Accept and reject examples make it easy to distinguish between output that sounds confident but is incorrect and output that you can use.
I want to assign this task to AI on a recurring basis.
Create an acceptance-test set for it.
Task:
[Describe the task]
Accepted examples:
[Paste one or more outputs I approved and explain why]
Rejected examples:
[Paste one or more outputs I rejected and explain why]
Use these examples to identify the quality standard.
Do not invent a quality standard that is unsupported by the examples.
Identify any standard that I still need to define.
Provide:
1. Five test examples
2. The passing criteria for each example
3. The evidence required to confirm that each case passed
4. Common errors
5. How to detect fabrication or unsupported conclusions
6. How to detect use of outdated or unauthorized sources
7. Situations that must be given to a person for judgment
8. Any unresolved quality standard that requires my decision
Include:
- A normal case
- A missing-information case
- A conflicting-information case
- A difficult edge case
- A case requiring human judgment5. Consent Material
A human agent system works best when everyone knows where the lines are: what AI can do on its own, what it can prepare for you to approve, and what it shouldn't do on its own. AI can handle repetitive manual work, approve important recommendations, and there is a record of how the final result was produced.
It is also important for irreversible actions. Deleting a file, modifying a production system, authorizing a purchase, or publishing something publicly can cause consequences that are difficult to reverse.
This property can be your personal agent system. It explains what the agent can handle, what still needs your permission, what data it can use, and where you live. It also keeps a record of the guesses it made, what it changed, and who approved the final result.
Create a permission policy for this AI workflow.
My recurring tasks:
[Describe the tasks]
Divide all activities into three categories:
1. AI may do this directly
2. AI may prepare a draft, but I must approve it
3. AI may never do this directly
For each activity:
- Give one specific example
- State when AI must stop and ask me
- Identify any irreversible action
- State which data or systems AI may access
- State which data or systems AI may never access
- State whether the action must be logged
- State what evidence must be retained for review
- State who is accountable for the final result
Pay particular attention to:
- Sending emails
- Deleting files
- Modifying production systems
- Purchasing anything
- Publishing publicly
- Contacting other people
- Making final decisions for me
- Accessing confidential information
- Using employee, customer, financial, or legal data
- Changing source data
- Approving transactions
- Creating external commitmentsUsing the Five Assets
Once you have these five assets, one key piece of information is to assemble them into a reusable workflow.
I want to use you as an AI assistant that can complete complex work.
Do not execute the task yet.
Based on the materials I provide, create a reusable workflow.
Define:
1. The standard input
2. The standard output
3. The steps between input and output
4. Which steps AI can perform directly
5. Which steps require my approval
6. Which steps must remain human-led
7. The acceptance standard
8. The permission limits
9. The sources AI may use
10. The evidence that must be retained
11. A minimum working version I can test today
12. The risks I should resolve before increasing access or automation
My task:
[Describe the task]
My materials:
[List or attach the materials]A Final Thought
Reorganize Work Before Adding More AI Agents lays out five leadership decisions that must be made before scaling AI agents. Packaging what your team already knows into reusable assets turns that strategy into action. These five assets give you an effective way to create value with AI.
Before assigning multiple tasks to AI agents, write down what they need to do. Start with one defined function, one reliable input, one standard output, and one authorization point. Use real examples for it. Compare the output with your accepted and rejected conditions. Correct gaps before adding more access, more steps, or more autonomy.
The person collecting the data only asks the model to make predictions. Models, licenses, and platforms will always change. The value comes from turning what your team already knows into work that AI can repeat, people can review, and the business can trust. It requires clearly defined workflows, quality standards, correct sources and context, and clear points where AI should stop.
It only offers “help me with this,” and can only guess. Give the AI a scene, materials, and levels, and it can do it. This is where AI transitions from experimentation to practical business value.
* Author's Note: All images in this article are created by the author using AI tools.
