KIC’s internal AI use falls as hardware expansion continues
Korea Investment Corporation disclosed a sharp fall in use of its internal AI service between May and September. Its parliamentary information also sets out the reasons for choosing different models and expanding graphics-processor capacity. The institution says the internal application supports nine kinds of work, limiting a direct comparison with employees’ much higher ChatGPT use.
Artificial Intelligence··Midday
Use declined over four months
Korea Investment Corporation’s internal AI service received 798 requests in September, compared with 3755 in May. User numbers fell from 339 to 147. The South Korean public investment institution’s parliamentary disclosures, covered by Daily NTN and Maeil Business on October 19, put these adoption figures alongside plans to expand computing capacity. Employees recorded 36881 ChatGPT requests in September. KIC says its own application supports only 9 kinds of work. The two services cover different task ranges. The published figures count requests and users, without providing a measurement of completed work or time saved for each employee.[1], [2]
Four models divide the work
The internal system divides work among four models. Alibaba’s Qwen3 supplies written responses, summaries and translation. BGE-M3, developed by the Beijing Academy of Artificial Intelligence, retrieves internal documents. Meta’s Llama3 filters harmful content, while Google’s Gemma3 processes images. KIC and its implementation contractor selected the models together. The institution identified accuracy, reasoning, suitability for its tasks and benchmark results among the considerations. Its initial capacity of 4 graphics processors also shaped the choice. KIC explained its preference for Qwen3 through cost, available graphics memory and multilingual suitability. These are the institution’s stated selection grounds, rather than an independently measured ranking of the models.[1], [2]
Expansion is planned for November
KIC signed the capacity-expansion contract on May 7, the same day it opened the service to employees. The project will increase graphics processors from 4 to 12, with completion planned for November. A request document prepared in March had cited increasing work demand. The institution subsequently linked the equipment to development of its internal AI platform and said procurement took account of hardware lead times and rising graphics-processor and memory costs. Use of the internal service declined after the earlier capacity planning. They describe an investment still scheduled for completion, and do not yet measure the effect of additional hardware on employee use.[1], [2]
Related columns
For more information on this topic, you can read the related columns.
