One use is not one comparable task
Korea Investment Corporation recorded 3755 uses of its internal AI service in May and 798 in September. User numbers fell from 339 to 147 over the same period. Those figures raise a workplace question: which tasks do employees complete with this tool, and which do they take elsewhere? I would first examine the point where a task meets the application, rather than the model’s name. Falling use indicates that employees are using the service less; it does not establish why. Work may have changed, the application may not have proved useful, or employees may have found another way to obtain the same result. Reading a usage count directly as productivity erases those distinctions.[1], [2]
Parliamentary data reported by Daily NTN on October 19 put employees’ September ChatGPT use at 36881. The difference is substantial. Maeil Business’s report on the same day also carries the institution’s explanation: its internal service is a limited application supporting 9 types of work. That explanation does not remove the decline; it changes the comparison that needs to be made. Comparing total use of a broader service with total use of an application reserved for selected tasks cannot establish where the same work is done better. Shared tasks must first be separated out. Summarizing a document and asking a general question do not become equivalent outputs merely because each counts as one use. This is a way of specifying the measurement, rather than defending low adoption.[1], [2]
The work remaining after the response
The internal division of tasks is concrete. Qwen3 handles language work, BGE-M3 retrieves documents, Llama3 filters content and Gemma3 processes images. Document summaries, translation and searches of internal material are among the applications described. An employee’s work may continue after a model responds. Checking that the retrieved document is the right one, comparing a summary with its original and making the result usable also require labor. I am not claiming a measured duration for those activities at KIC. I am proposing that they be made visible. If an apparently successful response requires extensive correction afterward, counting only the time saved in producing the first answer understates the employee’s contribution to the completed task.[1], [2]
Asking employees why they stopped using the service should therefore complement the usage table. Is a summary difficult to verify? Can the required document be found? Is the result usable in daily work? These are questions I propose for an assessment, rather than answers collected from staff. Fewer requests could mean fewer retries; they could also mean avoidance of an unhelpful tool. The available counts do not distinguish those possibilities. Employees’ say over tool selection matters here. Explaining their experience should not become another assignment undertaken merely to meet management’s adoption target. An arrangement that makes work easier needs to be distinguished from one that simply produces more clicks.[1], [2]
What capacity delivers at work
The hardware plan is a separate decision. The reports describe an expansion from 4 graphics processors to 12, which the institution links to upgrading its internal AI platform. More hardware does not establish a lower employee verification burden. Falling use does not automatically invalidate every future capacity requirement either. Workloads and planned tasks need to be considered together. I would judge the workplace value of the investment through comparable tasks employees can complete, the total time they spend and the correction burden they carry. Those are proposed measures to place alongside adoption counts, not published outcomes for KIC. A useful tool helps employees finish their work without requiring additional use simply to demonstrate the tool’s own importance.[1], [2]