Korea puts KT, SK Telecom and Kakao to work turning AI chat into filed paperwork

What the three consortiums are actually bidding for is permission to write into systems they do not own. SK Telecom's headline feature, issuing 83 kinds of certificate including resident registration abstracts, works by linking to Government24 and PASS ^[5]. KT's rental scenario runs on Zigbang listings and ends with the assistant filing the move-in report, changing the postal address and moving the utility accounts ^[9]; its homework example leans on EBS and Qanda content ^[8]. Answering questions is a demo; acting on someone's behalf requires credentials, and the portal side is what grants that trust.

Which is the awkward part for anyone modelling this as a platform contest, because if all three plug into the same national rails, the certificate menu becomes table stakes quickly and the only durable layer is whatever the assistant remembers about you. KT is the one selling memory as the product: with consent it retains interests, schedules and tasks in progress, then raises the next step unprompted ^[7]. That is the closest thing to a switching cost in any of the three descriptions, since leaving means teaching a new agent everything again.

Then the cash, or rather its absence. Nothing in the reported program carries a contract value or a consumer price, and no funding split has been disclosed either ^[15], so what an investor gets is a calendar across housing, healthcare, education, employment and public services ^[2]. On that calendar SK Telecom's beta sits one month behind Kakao's ^[13], and only two of the three named leads have a dated beta at all ^[14]. Until a figure appears, the near-term effect on any of these income statements is spend: integration work, and inference on errands users have not been asked to pay for.

The paths diverge more than the launch schedule suggests. It could settle into procurement work funded by the ministry, with thin margins but a sticky client. It could become a retention feature folded into mobile plans, in which case the benefit never shows up as a line item and you end up arguing about churn instead. Or the agent layer stays shallow, the state portal does the real work, and three assistants converge on the same product with different logos.

This is probably wrong, but I would rather own the memory layer than the certificate count, because the certificate count is a published integration and memory is a habit. What would break that view: if the consented profile turns out to be portable or trivially reconstructed, the moat is a rental, and if Government24 ever fronts a first-party agent of its own, the three consortium leads are contractors with a September deadline and a support ticket queue.