Korea and Japan: Shopping Agents Shipped, Agent Payment Rails Did Not

South Korea and Japan did the visible half of agent commerce and skipped the invisible half. Both countries shipped consumer shopping agents inside dominant platforms in the first half of 2026. Neither shipped a payment protocol for agents. The agents live inside wallets that already existed, spending money the way a human would, and no wallet in either country has published a single agent-initiated transaction number.

That is the finding. The rest of this page is the evidence, and the reason the absence matters more than the launches.

What actually shipped in Korea

Naver. CEO Choi Soo-yeon said on 6 February 2026 that the shopping AI agent had reached closed-beta readiness, with public release targeted for late February. It landed on 26 February 2026 as "shopping AI agent version 1.0", a *beta* inside the Naver Plus Store app, covering digital, living and lifestyle categories, with beauty and food promised within H1 2026. Naver describes a sub-agent architecture over "Shopping Intelligence", a commerce-specialised LLM trained on its own shopping-ecosystem data.

On 1 June 2026 Naver announced the agent now initiates conversations, prompted by clicks, saved items and cart additions. That press release contains no user count, no transaction count, and no payment integration — checked directly against the release.

Worth naming, because it is the exact failure mode this wiki exists to catch: the launch-day trade headline reads "Naver Launches AI Shopping Agent for Voice-Activated Purchases", and the body of the same article describes an agent that summarises, compares, analyses reviews and guides navigation. No purchase execution appears in it. Naver's own Q1 2026 commentary is consistent with the body, not the headline: the agent is a shopping guide gaining Membership and delivery integration, AI monetisation is tested in Q2 2026 and goes full-scale in Q3 2026, and the stated aim is raising "the likelihood of leading to actual purchases and reservations" — a conversion goal, not a shipped checkout. Anyone who cites the headline has cited a claim its own source does not make.

*Reported, not verified:* Naver Plus Store at 7.77 million MAU in March 2026, up from 7.1 million in February (attributed to Mobile Index); agent users +20%, conversations +40%, clicks on recommended products +140% versus initial introduction (attributed to Naver). Percentage deltas with no absolute baseline, reaching English third-hand. They give direction and nothing about size.

Kakao. "Kanana in KakaoTalk" entered beta in October 2025, iOS-only, initially requiring an iPhone 15 Pro or newer, with Android promised at full launch. Kakao scheduled an update for 21 February 2026, iOS first, to let it recommend restaurants and complete reservations inside the chat window — reservations via a booking button, not payments. That it shipped is attested after the fact rather than on the day: on 8 July 2026 Kakao announced it "has upgraded the place recommendation and booking functions" of Kanana, extending them "well beyond its original restaurant focus" into lodging, exhibitions, performances and cinemas. You cannot broaden a booking function that does not exist. Still no payment step appears in the July announcement either. On 9 March 2026 Kakao folded its existing AI shopping feature into Kanana. Kakao says more than 80% of invited testers completed the required on-device model download and roughly 70% still use the service regularly — self-reported, on an invited cohort, not a market figure.

On commerce, the verifiable version of Kakao's roadmap is narrower than the one in circulation. CEO Shina Chung told the February 2026 earnings call that "commerce was the scenario with the highest proportion among agents running on Kanana in KakaoTalk" and that "in the first half of the year, our goal is to more actively explore contexts where transactions can occur." Separately, Kakao's 2026 plan is to fully unveil agent-based Kanana in H1 2026 and then integrate it into services including payments, maps, music and reservations. No date for in-chat payment has been published, and no payment step had shipped as of 3 August 2026.

KakaoTalk's nearly 50 million domestic users are the whole argument for this architecture: the agent is not acquiring users, it is being placed in front of a population that already opens the app daily.

The duopoly, and what it is not

Naver and Kakao are a genuine duopoly of attention, and both own a wallet. Naver Pay processed KRW 24.2 trillion in Q1 2026 (+23% YoY), of which KRW 13.5 trillion (56%) was external payments (+33%), against KRW 86 trillion for full-year 2025. Kakao Pay has moved the other way, toward financial services: KRW 145.9 billion of financial-services operating revenue in Q1 2026 (+82.0%), now 49% of operating revenue.

Neither company has announced an agent payment protocol, an agent identity scheme, a mandate format, or an agent transaction disclosure. Their agents are UI on top of consumer checkout.

Japan: enormous rails, zero agent traffic disclosed

PayPay is the rail. From its SEC Form F-1 — the highest-quality disclosure in this region, because it carries securities liability:

| Metric | Value | Period | |---|---|---| | Payment Segment GMV | ¥15.39 trillion | FY ended 31 Mar 2025 | | Transactions via PayPay app | 7.8 billion | FY ended 31 Mar 2025 | | Registered users | ~72 million | as of 31 Dec 2025 | | Monthly transacting users | ~40 million | Dec 2025 | | Share of Japan code-based payment GMV | 64% | as of 14 Mar 2025 | | Share of all Japanese cashless transactions | 20% (of ~38.8bn) | 2024 |

The F-1 does not contain the words "AI agent", "agentic" or "agent commerce" anywhere. Its only AI reference is to "generative AI, particularly in coding workflows". A company registering securities in the United States in early 2026, in the middle of the loudest agentic-payments news cycle on record, did not consider agent commerce a business line, a strategy, or a risk factor worth naming. That silence is the most reliable signal on this page, because it is the one statement here made under legal liability.

A separate PayPay release of 18 March 2026 reports 40 million of 73 million registered users identity-verified, those users accounting for 87.8% of payment volume and 85.8% of transaction count as of February 2026, across ~1,000 connected financial institutions, with a ¥1 million per-transaction cap on balance payments. (The 72m/73m step between December 2025 and March 2026 is growth, not a contradiction.) Note what that cap is: a per-transaction ceiling on a human, not a delegated spending mandate with scope (A Spending Limit Is Not a Conduct Limit).

Rakuten Pay is second. Rakuten does not publish a Rakuten Pay registered-user figure; the ~44 million number in circulation comes from April 2026 press coverage of a points-to-XRP feature and should be read as reported, not disclosed.

The Japanese agent product is LY Corporation's "Agent i", launched 20 April 2026, merging Yahoo! JAPAN's AI Assistant with LINE AI. Its own release schedules complex task-execution for "around June 2026", an AI Shopping Memo linked to Yahoo! JAPAN Shopping, and Agent i Biz in August 2026. The release does not mention PayPay — the wallet owned by the same corporate family — and describes no payment mechanics. Reported via coverage of LY's Q1 FY2026 investor slides, and not audited: 12 million DAU as of June 2026 across roughly 25 domain agents, with the DAU composition undefined.

The protocol question, answered in the negative

There is no Korean agentic payment protocol. There is no Japanese one either. We looked for a national standard, an industry consortium, a bank-led scheme, a wallet-issued specification. As of 3 August 2026 we found none.

What exists instead:

1. Two regional names on someone else's protocol. Google's AP2, announced 16 September 2025 with "more than 60" collaborators, names 45 of them publicly — 42 with an attributed quote, plus Etsy, Revolut and UnionPay International named in passing. Exactly two are from these countries: JCB (Japanese card network) and KCP / NHN KCP (Korean payment gateway). Not Naver Pay, not Kakao Pay, not PayPay, not Toss. A card network and a PG — the plumbing, not the wallets that own the users. (The negative covers the named roster; at least fifteen unnamed collaborators are unverifiable.) 2. One internal demo. Exchangers, a Japanese foreign-exchange firm, verified an autonomous agent payment loop using x402 (x402: HTTP 402 Finally Gets a Job, at 32 Cents a Transaction) with a yen-denominated e-money token (XJPY) on its own private blockchain, gasless via its own miner, from a KYC-completed wallet — reported 9 July 2026. It ran "in a closed internal environment"; real-environment testing is a plan. Its priority claim is narrow and self-declared: no prior Japanese x402 implementation combining yen, KYC wallet and gasless payments. This is a demo, and we count it as one. 3. A Chinese protocol arriving through the side door. Ant International launched the open-sourced Agentic Mobile Protocol (AMP) on 27 April 2026, explicitly targeting digital wallets and super-apps, with a Know Your Agent identity framework and an Agent Trust Rating. Alipay+ connects 40+ wallet partners, 1.8 billion user accounts and 150 million merchants. Ant names no individual wallet partner, so no adoption by any Korean or Japanese wallet is proven — but the structural fact is that both Kakao Pay and PayPay are already Alipay+ counterparties.

That last point is the one worth sitting with. The only cross-border agent-payment plumbing plausibly reaching Korean and Japanese wallets is Chinese — and that is a claim about opportunity, not about a signed deal.

The cross-border rail that already exists is Chinese

The Korea–Japan payment corridor is real and it is intermediated by Alipay+:

- April 2023 — PayPay and Kakao Pay sign a business agreement to expand cross-border acceptance over the Alipay+ network. PayPay was then at 55 million registered users. - November 2024 — Korean tourists gain the ability to use Kakao Pay at Japanese merchants via PayPay's local network. - 25 September 2025 — Japanese PayPay users can pay at more than 2 million Korean merchants. Reported as PayPay's first overseas expansion, executed as a three-way arrangement between Kakao Pay, PayPay and Alipay+. Context: 1.92 million Japanese tourists visited Korea January–July 2025.

Two national champions could not connect their wallets to each other without a Chinese gateway sitting between them. Whatever agentic layer eventually rides this corridor inherits that dependency.

The market-size number does not survive the central bank

The figure circulating for Korean mobile payments is USD 44.38 billion (2025) → USD 48.25 billion (2026), reaching USD 73.33 billion by 2031 at 8.72% CAGR. We traced it to Mordor Intelligence, a market-research vendor. Mordor states the unit explicitly — transaction value in USD — and states that the number comes from a "proprietary estimation framework" whose mechanics are paywalled.

Because the unit is stated, the comparison with the Bank of Korea is like-for-like, and it fails badly:

| Series | Value | Annualised | Ratio to vendor 2025 figure | |---|---|---|---| | Mordor "mobile payment market", 2025 | USD 44.38bn | — | 1x | | BOK simple payment services, H1 2025 | KRW 1.046tn/day (USD 745.77m/day) | ≈ USD 272bn | ~6x | | BOK mobile-device card payments, 2025 | KRW 1.7tn/day | ≈ USD 442bn | ~10x |

BOK's H1 2025 simple-payment figures are 1.046 trillion won and 33.78 million transactions per day, up 11.4% and 13.7% year on year. Its 2025 full-year figures, reported 30 March 2026, put total daily card spending at 3.1 trillion won, of which mobile-device payments are 1.7 trillion won per day (+7.3%, an all-time high) and 54.3% of card spending (from 52.4% in 2024), against 1.4 trillion won per day on physical cards (-0.4%). Annualisation uses the FX rate implied by the Korea Times conversion, ~1,403 KRW/USD.

Salesmart S.r.l., trading as Sinapsi, which publishes this wiki, traced that figure to its vendor and ran the comparison above. What follows is our judgement, offered as judgement and not as a reported fact: the vendor figure is irreconcilable with the central bank's own transaction-value series, by roughly six to ten times depending on which BOK series you pick, and Mordor publishes nothing that would let anyone diagnose why. We do not claim to know what it measures. We claim it should not be cited as "the Korean mobile payments market". Use the Bank of Korea. This is the same discipline the rest of this wiki applies to self-reported protocol volumes (The x402 Market, Counted: 14,766 Listings, 12,741 Calls a Day, One Winner).

Against China: same architecture, no protocol layer

China Shipped Agent Payments First. The 1,000x Volume Gap Does Not Survive the Numbers. describes a closed, platform-internal, identity-first stack that reached agent-payment scale precisely because payer, payee, agent and identity all sit inside one super-app. Korea and Japan have the identical architecture — Naver, KakaoTalk, PayPay, LINE are all closed loops with verified identity and a captive wallet — and have produced no equivalent output.

The difference is not architectural, so it is not explained by architecture. Three candidate explanations. These are Sinapsi hypotheses, not reported facts, and each is written to be falsifiable:

- No protocol layer was attempted. Chinese players shipped governance, technical and wallet-side specifications (see China Shipped Agent Payments First. The 1,000x Volume Gap Does Not Survive the Numbers.); Korea and Japan shipped product features. You cannot report agent transaction volume if nothing in your stack distinguishes an agent transaction from a human one. *Falsified by:* any Korean or Japanese agent-payment specification published under its author's own name. - The regulatory clock runs later. Korea's AI Basic Act — a horizontal AI regime, not a financial one — took effect 22 January 2026, with MSIT granting a one-year grace period before administrative fines are imposed. Its high-impact category is described in the law-firm briefing we fetched as "healthcare, energy, transportation, hiring and biometric analysis"; finance is not named in it, and we make no claim that it is. Sector rules arrived later still and softer: the FSC's revised AI guidelines for the financial sector are self-regulatory, were introduced on 18 June 2026 and take effect 22 June 2026, and the FSC states that rules on the responsibility and authority of AI are still to be prepared "as AI agents are expected to player a greater role in the areas of making recommendations for financial products, subscribing to financial products, and making payments" — with test operations on AI agents only "from the second half of this year". So every binding or quasi-binding instrument landed after or alongside the product launches, and none of them defines an agent payment. Dates read against the Cooley briefing and the FSC's own English release, not against the primary statute. *Falsified by:* evidence that a wallet had a protocol ready and held it back for reasons other than regulation, or by an equivalent Japanese timeline that does not fit. - Scale changes the payoff. Ant is building for 1.8 billion Alipay+ accounts. PayPay's 40 million monthly transacting users may not repay a bespoke protocol when an open-sourced one is free. *Falsified by:* PayPay or Naver Pay publishing its own specification anyway, or by adopting AMP and saying so.

The throughput gap is real, but it points the other way from the agent story: PayPay alone runs 7.8 billion transactions a year, roughly 21 million a day, against the x402 open market's measured 12,741 calls a day (The x402 Market, Counted: 14,766 Listings, 12,741 Calls a Day, One Winner) — about 1,700x, more than three orders of magnitude. And none of PayPay's are disclosed as agent-initiated. Enormous rails, no reported machine traffic. That is the precise inverse of the The Earn-Spend Loop: Why Machine Payment Is Half-Built problem in the West, where the machine traffic exists and the rails do not.

What would change this picture, and how you would know

Four checkable events, none of which had occurred as of 3 August 2026:

1. Naver, Kakao Pay or PayPay publishes an agent-initiated transaction count. Any number at all converts this page from an absence to a measurement. 2. A Korean or Japanese entity publishes an agent payment specification — mandate format, agent identity, delegated authorisation — under its own name rather than implementing someone else's. 3. Kakao connects Kanana to Kakao Pay for in-chat payment. Kakao has said payments are among the services Kanana will integrate, without a published date. If the payment step arrives with no agent-specific authorisation layer, that confirms the thesis here: agents inside wallets, not wallets for agents. 4. PayPay's next annual filing mentions agentic commerce. Its F-1 does not. A company that must disclose material risks adding the category would be a stronger signal than any press release either country has issued.

Until then the honest sentence is: Korea and Japan have consumer shopping agents with undisclosed usage, wallets with enormous human volume, no native agent payment rail, and a cross-border corridor that runs through a Chinese gateway.