KT released the final results of its joint AI research with Seoul National University and KAIST on Aug. 30. The focus of the research was not on building a bigger model. It was on how accurately an autonomous agent reads a user's intent and situation, and how much the processing load consumed in that process can be reduced. It is a case showing that the concerns of AI development in Korea are moving from performance competition to operating costs.
The research with Seoul National University dealt with agents' understanding of situation and context and with reasoning technology. Added to this was work on setting evaluation standards for RAI, or responsible AI, and improving the reliability of answers. Collecting datasets suited to Korean-language environments, a model that predicts user behavior, and a reinforcement learning framework that feeds user feedback back into training were also within the scope of the research.
The research with KAIST dug into a different area. Prompt compression is a technique that shortens input length by stripping out parts of the text fed to a model that say the same thing twice or that have little effect on the result. When the input gets shorter, the number of tokens the model has to read falls, and when tokens fall, the cost and time for a single response come down together. In situations where an agent calls a model dozens of times while handling one task, this difference accumulates.
Other news released in the same week touches on this. SK covered material related to the token economy, to the effect that token governance is needed in the era of AI agents, in internal documents. Abroad, as AI contracts at Salesforce and Workday increase, billing is appearing in a form that adds usage-based charges on top of a flat subscription fee. Under a system where you are billed for what you use, token consumption becomes an operating profit line item.
For companies, this means what has to be managed changes. More important than the decision to adopt a single model is the work of looking at how many tokens are used a day, where and how, after adoption. If usage by each person is not tallied when agents are turned loose across departments, the usage pattern only becomes known after the bill arrives.
That is also why reliability research is tied to the cost question. If an agent gives an off-target answer, a person has to ask again, and tokens go out again for each re-ask. Raising the rate of getting it right the first time is both a quality improvement and a cost reduction. That is why work on setting RAI evaluation standards is read as an operating metric beyond an ethics item.
How far these results go into products is the next gate. KT said it would internalize the research results as core technology and apply them in the field, but it did not say which services would reflect them or when. Measurements such as the scale of token savings or improvements in response latency did not come out along with them either. The actual usefulness of the technology will become clear at the point those figures are disclosed.
An employee who has AI draft a report in the office does not know how many tokens they used. There is no need to know. Instead, the company settles that total every month.
If compression and reliability research work out, more work can be done for the same fee, and that many more people will call on AI without hesitation. The change in a day starts at that point.
