A Japanese Telco Launched an AI App Whose Selling Point Is That It Cannot Use the Open Web

KDDI launched Buffmee at 01:00 UTC on 28 July, a consumer AI assistant built on a constraint rather than a capability: it answers only from content its publishers have licensed, and cannot draw on the open web at all.

The product

Roughly 150 licensed sources — books, magazines and web media — across approximately 30 partner news organisations and publishers. Named partners include Orange Page, Takken-shi Goukaku Navi and Railway Fan. The free tier allows 100 chats and 10 image generations a day; premium runs ¥980 a month and is free for one year to early registrants. It ships on iOS and Android, built on Google Cloud under a partnership dating to October 2025.

The distinction the framing hides

"Licensed publisher content" implies something like the OpenAI–News Corp class of agreement. It is not the same bargain. Buffmee licenses retrieval with citations, not training data — the model is not being trained on these publications, it is being pointed at them at query time. That is a dramatically cheaper rights deal for the platform and a dramatically weaker one for the publisher, because it pays for lookups rather than for the corpus. Roughly 150 titles is also a curated shelf, not a market-wide licence.

What the publishers are actually getting

Right now, a share of nothing. Premium is free for the first year, so whatever revenue-sharing arrangement exists is currently dividing zero subscription income. The publisher case rests on the second year and on referral traffic from citations — the same bet European and American publishers have been making, with results that have not yet justified it anywhere.

Why Japan, and why a telco

Japanese publishers have been among the most resistant in the world to AI content licensing, and Japanese consumers have been slower to adopt Western assistants. A carrier with a billing relationship and a distribution channel is well placed to sell a product whose pitch is trustworthiness rather than capability — and a walled garden is exactly what a rights-holder that has refused to license training data will agree to.