"Hand a Whole Job Over to AI and You Won't Build New Skills": Professor Lee Joong-hak's Advice

SK Telecom’s newsroom has published a column by Lee Joong-hak, a professor of business administration at Dongguk University, titled “What Will We Build With AI?” In it, Lee stresses that dividing work into stages comes before ramping up how much AI you use. Only after that breakdown, he argues, should you reassign which parts belong to people and which belong to AI. Even when AI does the work, he writes, the judgment and the responsibility for the outcome still rest with humans.

Writing a Report: How Much Should You Delegate?

Lee explains that the unit by which work is divided is shifting. In the past, people split tasks among themselves and each used AI as a tool. Now, within a single task, humans and AI take on different stages. Writing a report, he suggests, needs to be broken down into defining the problem, gathering material, analyzing it, drafting, and reviewing and deciding, so that you can pick out which parts to hand to AI.

As an example from South Korea, he points to SK Telecom’s “AX Innovation 2.0,” an AI transformation initiative. When the company announced it in June, it gave AI agents that handle work on their own an employee ID number, an organizational assignment, a job function and a set of permissions. It also introduced an “AX Sandbox” for redesigning existing work from scratch on an AI foundation. Lee cites this as a case of redistributing human and AI roles stage by stage.

What Happens If You Submit AI’s Answer As Is?

A Harvard Business School study cited in the column sorts ways of working with AI into three types. The research analyzed roughly 5,000 exchanges between AI and 244 Boston Consulting Group consultants. The “cyborg” type, who traded outputs back and forth with AI across the entire process, preserved their existing domain expertise while also gaining skill at using AI. The “centaur” type, who used AI only at certain stages, produced the most accurate conclusions of the three groups.

The “self-automator” type, by contrast, handed the whole task over to AI and built neither domain expertise nor AI skills. 44 percent of this group submitted the answers AI produced without editing them. For people already comfortable with their own way of working, Lee sees starting with AI at a few selected stages as the realistic entry point. Over the longer run, he recommends the approach of thinking and revising alongside AI throughout the entire process.

He identifies three capabilities people need to hold onto: deciding first what problem is to be solved, correcting course while AI is working, and checking the final answer. The column also cites research from the Wharton School. In the first experiment, incorrect answers were followed as given in roughly 80 percent of the cases in which AI was consulted, which the column points to as evidence that verification is necessary.

When Productivity Drops Early in Adoption

Lee argues that the trial and error of early AI adoption should be understood as a learning process. To support this argument, he cites the “J-curve of AX,” a concept introduced by SK Telecom CEO Jung Jae-hun in a September 1 guest lecture for the joint AI program run by SK Telecom and Seoul National University. The idea is that productivity dips right after AI is introduced, then climbs steeply once a certain point is passed. Lee explains that this period is when people learn the criteria for what to delegate to AI and when to step in.

The column’s recommendations move from individual practice to organizational performance management. Lee notes that pressing for results from the outset makes it easy for employees to dump entire tasks on AI. Organizations, he says, need to provide time and tools for experimentation and an environment where failure is tolerated. What to watch for in the workplace, he suggests, is not the volume of AI usage but the division of roles, verification procedures and the conditions that support trial and error.