Why is 'using' AI not enough anymore? The business revolution brought by Agentic AI
Until now, 'conversational' and 'generative' AI have been the mainstream. However, recently, Agentic AI, which 'thinks and acts more autonomously,' has been attracting attention.Andrew Ng is a figure at the very forefront of this trend, and at the 2025 Masters of Scale Summit, he spoke candidly about the potential of Agentic AI, the gap between that potential and reality, and the challenges that society and businesses face. In this article, starting from Andrew's lecture, we will organize what Agentic AI is, what is possible now, where the limitations and risks lie, and how we should approach it.
1. What is Agentic AI — Differences from conventional AI
1-1. Definition and Essence
Agentic AI refers not to 'mere response-based AI (prompt -> output),' but to 'autonomous agents' that receive a goal and plan, act, and reflect on their own to achieve it.
Conventional generative AI (so-called chat-based or one-off text generation) had the job of providing 'one-off responses.' On the other hand, Agentic AI attempts to complete tasks through 'multi-step workflows.'
1-2. Typical Architecture and Features
Agentic AI is not a single 'model,' but is built based on design patterns such as the following.
Planning: First, break down tasks into smaller subtasks
Tool Use: Use external APIs, databases, web clients, etc., to acquire information and execute actions
Multi-agent / Orchestration: Multiple agents share roles to distribute and coordinate processing
Reflection: Look back at generated results, discover errors and points for improvement, and loop them back
With this structure, Agentic AI goes beyond mere 'text generation' or 'image generation' to enable actual decision-making, action, and task execution.
2. Andrew Ng talks about the 'current' state of AI — Potential and cautions of Agentic
2-1. Distance from misconceptions surrounding Agentic
Andrew Ng is wary of the fact that Agentic AI is 'easily abused as a marketing term.' For example, he points out that there are many cases where it is labeled 'Agentic' but in reality, it is just simple prompt-generating AI. Furthermore, he emphasizes that 'it only demonstrates its true value when the Agentic structure is designed correctly.'
2-2. Realistic introduction value and application areas
Andrew Ng argues that Agentic AI is the 'smartest bet' for companies. This is because it has the following advantages:
Automation of complex and long workflows — For example, multi-step processes such as research -> analysis -> report creation -> distribution can be automated end-to-end.
Scalability — Many tasks can be processed in parallel without significantly increasing manpower.
Flexible application range — Applications are envisioned in various fields such as customer support, supply chain management, financial auditing, research and development, and compliance.
In other words, Agentic AI can function not as a 'chat partner' or a 'writer that generates text,' but as a 'semi-autonomous business operator.'
3. Limitations and Risks — Why it is not 'all-powerful'
3-1. The structure and constraints behind autonomy
Agentic AI operates strictly based on the 'objectives and design structure' initially defined by humans. Unless goal setting, tool design, workflow segmentation, and evaluation criteria are properly designed, you will not get the intended results.
In short, it is not a case of 'just hand it over to Agentic AI and it's done'; the skills and responsibility of the designer become even more critical.
3-2. Concerns regarding data quality, security, and unexpected behavior
Especially when introducing it into a company, there are risks such as the following:
The possibility of making incorrect judgments or meaningless processing due to inappropriate or incomplete data
Security and compliance concerns when access rights to internal systems or confidential data are required
Unintended consequences due to goal misalignment (e.g., focusing too much on cost reduction and causing quality to drop)
Also, it is necessary to correctly understand that Agentic AI is not an 'all-purpose automation device' but has 'areas of strength and areas of weakness.'
4. Why Agentic AI is important now — Implications for society and business
4-1. The 'redefinition of labor' in the digital age
Agentic AI does not reduce the 'work that humans do not need to do,' but rather 'redefines the role of humans.' For example, Agentic AI will handle monotonous data processing and routine analysis, while humans will focus on tasks that require more creativity and judgment — such a transformation in work styles will progress.
Furthermore, for many professionals (marketers, finance, HR, etc.), the ability to collaborate with AI by having a little programming knowledge will become a 'new literacy.' This has the potential to spread across many job categories, not just technical roles.
4-2. Democratization of innovation and acceleration of speed
Conventionally, AI utilization and automation were led by large corporations and well-funded startups. However, Agentic AI makes it possible to launch prototypes and business systems 'with a small number of people,' 'on a small scale,' and 'rapidly.' Ng himself also calls out at the end, 'Now is the time to build (build build build).'
This change overturns the conventional 'technology → large corporation → system' structure and opens the door to innovation and entry by more diverse players.
5. How should we face this — Toward practice in the Agentic era
Start by trying with small workflows
Rather than handing everything over, it is realistic to first conduct a PoC (proof of concept) for Agentic AI in areas such as 'routine but time-consuming tasks,' 'tasks with multiple steps,' and 'regular reporting or data organization.'Do not neglect data governance and monitoring design
It is important to incorporate human intervention and check mechanisms, such as data quality, access rights, log management, and error handling.Redefining education and skills
This is an era where value lies not in being a mere 'tool user,' but in being 'talent who can design and instruct tools.' Possessing basic knowledge of programming and workflow design will become a highly versatile skill.
Conclusion
As of 2025, Agentic AI is no longer just a trend or a marketing buzzword; it has reached a stage where it can serve as a starting point for corporate and organizational DX, business transformation, and even the creation of new businesses as a technology with real value in practicality, versatility, and scalability.
However, if you misunderstand it as 'convenient magic,' you may fall into pitfalls such as design errors, data leaks, quality degradation, and malfunctions. That is precisely why the approach of 'setting up purpose design, data systems, and evaluation/monitoring mechanisms,' starting small, and gradually scaling will be the key to surviving the Agentic era.