Over the past two years, most enterprises have continued to think of AI primarily in terms of “question and answer”: a person asks a question, receives an answer, and decides whether to use it. Agentic AI turns this logic on its head. Rather than simply waiting for prompts, AI can plan its own steps, invoke tools, and advance a task over hours or even days until it is complete. This shift may sound like a technological upgrade, but the recent “Human-AI Collaboration and AI Governance Seminar,” jointly organized by Deloitte Taiwan and BIG DATA CO., LTD., offered an answer grounded more firmly in real-world practice: what truly determines whether an enterprise can put Agentic AI to work is not the capability of the model, but the completeness of its governance framework.
From “Answering Questions” to “Getting Work Done”: An Agent Is Not Just a Smarter Assistant
Shu-Kai Fan, Dean of the College of Management at National Taipei University of Technology, noted at the seminar that as frontier models gradually acquire capabilities such as long-horizon planning, tool invocation, memory updates, and autonomous research, the focus of enterprise AI competition is shifting from “being able to answer questions” to “being able to complete tasks and continuously improve.” This observation deserves closer consideration from decision-makers in traditional industries. In the past, the acceptance criterion for an AI implementation was whether its answers were correct. With Agentic AI, the criteria become whether a sequence of actions follows the rules and whether errors can be traced and reversed.
Fan also cautioned that the development of AGI should not focus solely on expanding capabilities. It must also establish “governable autonomy.” Mechanisms such as Agentic Workflow, MCP/A2A, Evaluation Harness, and Human-in-the-Loop can make decisions traceable, verifiable, interruptible, and accountable. At the same time, the human role shifts from executing processes to serving as an intelligent control tower for goal setting, risk governance, and value judgments. For traditional industries accustomed to managing production lines and quality control through SOPs, this analogy is not unfamiliar. The only difference is that the entity being managed is now software capable of acting on its own.
The Six Pieces of a Governance Framework: Every Missing Piece Leaves a Gap
Drawing on his consulting experience, Deloitte Taiwan Management Consulting Executive Vice President Jung-Lun Chiang proposed that enterprises should develop a comprehensive governance framework alongside their AI applications. This framework should cover six areas: strategic objectives, roles and responsibilities, risk assessment, effectiveness monitoring, decision traceability, and legal and regulatory compliance. He emphasized that future competitive advantage will depend not only on AI capabilities, but also on whether an enterprise can reshape its management mechanisms to balance innovation, governance, and trust. By embedding AI governance throughout the entire AI lifecycle, enterprises can develop AI from a tool for improving efficiency into a partner that drives transformation and sustainable growth.
Of these six areas, most enterprises are already familiar with the first two—strategic objectives and roles and responsibilities. The four that follow are more easily overlooked: Who is responsible for risk assessment? How frequently should effectiveness be monitored? Can decision records be retained? Can a compliance audit trace a problem back to the exact step where it occurred? In scenarios such as export controls, trade sanctions compliance, and supply chain risk monitoring, once an AI Agent enters the decision chain, these four issues are often the first questions an enterprise must answer—and the hardest to address after the fact.
Human Oversight Is Not a Delay; It Is the Enterprise’s Reserved Power of Veto
Chiang stated at the seminar that human-AI collaboration mechanisms are essential to ensure that critical decisions remain subject to appropriate human oversight and professional judgment. This can easily be misinterpreted as simply “slowing AI down so someone can take another look.” In practice, however, it involves much more specific design questions: Which types of decisions must always receive human approval? Which can an AI Agent complete autonomously and report afterward? Can an AI Agent be stopped immediately when something goes wrong?
This design does not reflect a lack of trust in AI. Rather, it treats “human oversight” as a reserved power of veto—not as an optional checkpoint in the process. For decision-makers in traditional industries, this trade-off closely resembles the questions they faced years ago when introducing automated production lines. Whether an automated machine should be able to decide when to stop has never simply been a question of how intelligent the machine is. It is a question of whether the enterprise is willing to delegate the authority to halt operations.
Once a Persistent AI Agent Is Running, Who Is Watching It?
The technical perspective shared by NVIDIA at the seminar illustrated what a governance framework looks like in practice. Persistent AI Agents represent AI’s evolution from assistants activated by prompts into software systems that can continuously retain context, coordinate tools, recover from interruptions, and advance complex workflows over hours or even days. Once this type of AI Agent starts running, it no longer operates in a simple “one question, one answer” interaction model. Instead, it becomes an ongoing process within enterprise systems.
From a governance perspective, this means enterprises need visibility into what an AI Agent is doing, the ability to approve or terminate its actions at critical checkpoints, mechanisms for retrying failed operations, and policy enforcement that extends across the entire workflow rather than being applied only once at the outset. These requirements may sound highly technical, but they follow the same logic as the quality inspections, abnormal production-line shutdowns, and rework processes familiar to traditional industries. The only difference is that the object being monitored has shifted from a production line to an AI Agent’s execution trail.
Before Governance, Identify Where Knowledge Is Scattered
Yu-Ting Yuan, Director of Innovation Marketing at BIG DATA CO., LTD., shared an observation with particular relevance to traditional industries: enterprises have accumulated a wealth of information and practical experience, but that knowledge is often scattered across different documents and systems—or even stored only in the minds of senior employees. This not only makes the knowledge difficult to organize, but also complicates retrieval and transfer. When helping enterprises adopt AI, the process should begin with an assessment of business needs and actual work scenarios, making existing data and experience easier to find, understand, and apply.
This also echoes a point made at the seminar by Chih-Wei Chiang, Chief Operating Officer of BIG DATA CO., LTD.: the end goal of AI adoption is not to add another AI tool, but to genuinely transform AI into an organizational capability. For traditional enterprises that rely heavily on veteran employees’ experience, with SOPs scattered across paper documents and verbal handovers, the order of these priorities is especially important. A governance framework determines how AI performs its work, but a knowledge inventory determines whether AI has anything to learn from or use in the first place. If this step is skipped and governance is addressed directly, even the most comprehensive framework will be an empty shell.
FAQ
Will a governance framework slow down AI adoption?<br>In the short term, establishing roles and responsibilities and implementing risk assessment mechanisms does require additional time. However, investing that time early helps prevent enterprises from discovering only after scaling that decisions cannot be traced or that accountability cannot be assigned when something goes wrong. The experience of most enterprises shows that time saved by skipping governance ultimately turns into remediation costs.
What should traditional industries do first when adopting Agentic AI?<br>According to the insights shared at the seminar, the first step is not choosing a tool or model. It is identifying where knowledge and data are scattered across the enterprise and who holds them, then determining which decisions must retain human approval. Only on this foundation can a governance framework be built.
What does “human oversight” look like in practice? Does every decision require human approval?<br>No. In practice, decisions are usually classified by risk level. High-risk or irreversible decisions retain human approval, while low-risk and reversible decisions can be completed autonomously by an AI Agent, with records preserved for subsequent review. The key is for the enterprise to establish this classification in advance.
How can a small or medium-sized enterprise without a large consulting team begin building these mechanisms?<br>It can start with two of the six areas proposed by Chiang—risk assessment and decision traceability—and extend its existing quality control or audit processes rather than creating an entirely new governance department. The scale of a governance framework should be proportionate to the size of the enterprise. What matters is that someone is accountable for each of the six areas, not that the most comprehensive possible framework be implemented all at once.
References
- Building the Dual Engines of Human-AI Collaboration in the Age of Agentic AI—Deloitte Taiwan: Enterprises Must Balance Innovative Applications with Governance Resilience — Official press release from Deloitte Taiwan
