What This Article Is About
I have long appreciated Professor Jian Lifeng's perspective on AI. He can examine Taiwan's broader industry and technology trends while also returning to the practical issues enterprises face every day, including work, talent, and adoption. His macro view and detailed practical observations coexist.
I especially appreciate his recent explanation of AI agents. AI used to stop at answering questions; now it is moving toward planning, using tools, and completing tasks. Professor Jian's phrase, 'From Answer to Action,' captures this shift clearly. I reviewed several of his public talks and reports and compiled this article around a practical enterprise question: when AI starts doing work, how should we integrate it into a workflow?
- The company has allowed staff to use generative AI, but the results are still limited to summaries, Q&A, and personal efficiency.
- Managers want to adopt AI Agent but don't know where to start with the first workflow.
- Teams are worried about data, permissions, errors, and responsibility, and want to start with a controlled pilot.
- A plain-language framework for understanding Answer to Action.
- A screening table for choosing the first Agent scenario.
- A six-step process and 90-day roadmap for enterprise small-scale pilot.
Answer to Action changes more than just a slogan.
In the past, when we used generative AI, common actions included asking questions, looking up information, summarizing, and generating suggestions. AI provided the answers, and the subsequent judgment and execution were still handled by humans.
As AI moves toward Action, it begins to participate in a task: understanding the goal, reading the context, planning steps, using tools, producing results, and then handing over the parts that require judgment back to humans. This is where AI Agent becomes truly important for enterprises.
In a public talk at the 2025 Future Commerce Expo, Professor Jian Lifeng summarized the evolution of large language models as a move from Answer to Action. He noted that agents can understand, reason, collaborate, and execute tasks. He also placed AI within workplace relationships by discussing human-machine collaboration and AI colleagues. This advances the enterprise adoption question: we are designing both a good answer and a process that can be delegated, checked, and taken over.
Agent workflow, which can be considered as six actions
I have organized the understanding, reasoning, tool usage, and human-machine collaboration mentioned in public sources into a workflow that is easy for enterprises to discuss.
What needs to be completed and who will receive the results.
What data, rules, templates, and historical records are needed.
Break the task into checkable stages.
Query data, update systems, or produce documents within authorized scope.
Clearly explain what has been completed, what is still missing, and any exceptions.
Have people confirm, correct, or stop at key nodes.
Companies can start with four types of scenarios
Organizing data, comparing documents, producing drafts, preparing meeting summaries, and answering internal questions with clear sources.
Assisting with programming, testing, data cleaning, and repetitive processing, while retaining code review and formal deployment permissions.
Preparing draft replies, organizing customer needs, comparing product conditions, so that people can focus their time on relationships, negotiation, and exception judgment.
Tracking task status, checking for missing items, comparing rules, and preparing procurement or inventory information.
The first scenario should be small enough and true enough
Many companies start by looking for the most powerful Agent. I will first check whether this scenario can be safely learned.
| Filtering Issues | Signals Suitable for Starting |
|---|---|
| Does it frequently occur? | Occurs every week or every day |
| Is the input accessible? | Files, fields, templates, and rules can be found |
| Can the output be checked? | There is a clear format, standard, or responsible person |
| Can errors be taken over? | Can the process be paused and handed back to human when errors occur? |
| Can the benefits be measured? | Can time, quality, cost, or exceptions be compared? |
| Can permissions be reduced? | The first version only gives access to what is needed to complete the task |
If four to five of the six questions can be clearly answered, the scenario is worth entering a pilot. If even the input, output, and responsible person are unclear, introducing tools usually only amplifies the originally vague process.
First complete a small pilot
Write down what you want to see after completion, for example, a summary, missing item list, or pending confirmation replies.
List files, templates, fields, judgment rules, and update responsibilities.
Each step clearly states what was read, what was done, and what was left.
Decide who confirms, when to stop, and which exceptions must be handed over.
Review quality, time, cost, exceptions, and manual review volume together.
Feedback errors and exceptions to data, rules, templates, and role assignments.
Data, talent, and guardrails determine whether an Agent can actually work.
Data and systems
For an AI to complete a task, it must be able to read the company's data and use tools under clear permissions. Companies can start by inventorying readable content, structured data, necessary tools, update mechanisms, and minimum access scope.
As future searches and business processes involve more AI, companies should also make product catalogs, specifications, inventory, and transaction conditions easier to be consistently understood. This is a practical preparation direction I derived from public discussions on Agentic Web, conversational commerce, and AI procurement.
Human Role
One of Professor Jian Lifeng's public observations strongly resonates with me: when AI starts participating in work, people should focus on asking questions, exercising judgment, integrating information, and assuring results. Public reports describe this shift as moving from Work Faster to Work Smarter.
Six basic guardrails
- Quality standards: what results are acceptable, which fields must be checked.
- Data permissions: which data can be read, which data should not enter the model or process.
- Tool permissions: Can query, draft, modify, or can formally submit.
- Human nodes: Which steps must be confirmed by a person.
- Operation records: Retain necessary records of tasks, data, tools, and results.
- Stopping conditions: What situations require pausing or transferring to human intervention.
Make a first round of learning that can be judged within 90 days
Inventory high-frequency processes, available data, roles, inputs and outputs, and human checkpoints.
Let the Agent work with limited data and permissions, retain human review and record exceptions.
Compare quality, time, cost, and human workload to decide to expand, adjust, or stop.
References and boundary of compilation
The following data are all from public speeches or articles. This article only provides summaries, translations, and practical enterprise organization:
- INSIDE: Jian Lifeng, large language models from Answer to Action, are changing all Agent services
- INSIDE: Jian Lifeng, generative AI entering the 2.0 era
- Central News Agency: Jian Lifeng, conversational commerce in e-commerce is a trend
- CommonWealth Magazine: Jian Lifeng on what not to ask AI to do when it becomes an employee
- Manager Person: Is your enterprise on the AI procurement list?
- IMC Elite: Facing Taiwan's Golden Decade, AI Agent trends under enterprise organization and talent transformation
I am Coach Jiang
I am Jiang Yude, also known as Coach Jiang. I focus on how AI enters real work processes and help people organize their professional knowledge, judgment, and experience into AI-understandable and usable knowledge.
If you are also thinking about how to introduce AI Agent into your enterprise, you can start with a small process: write input, output, rules, permissions, exceptions, and evidence of completion on the same piece of paper. This paper is often closer to the real starting point than buying a set of tools.
I will share new articles, public lectures, and practical summaries in the LINE community.
Join LINE Community