[Edaily Reporter Kim Hyun-ah ] Users can earn Naver Pay points by writing reviews after purchasing products on Naver Shopping. Naver (NAVER(035420)) awards 50 won for text reviews and 150 won for reviews that include photos or videos. After a certain period following purchase confirmation, users can also write an additional monthly review. Additional points may be awarded depending on the product and seller settings.

Although the amount isn’t large, writing reviews and selecting photos can be a hassle. Cases have emerged where users are entrusting even this task to artificial intelligence (AI) agents. When a user makes a request, the AI logs into Naver and writes the purchase review. It goes beyond simply searching for products or providing answers; it even takes over tasks that users previously had to perform themselves to earn points.

AI agents are evolving to the point where they can log into actual online services and handle tasks. As AI takes over not only product searches, orders, and writing purchase reviews but also restaurant reservations, KTX ticket bookings, and even phone calls, changes to the business models of existing platforms are expected.

At the Future Technology Strategy Forum held on the 5th (co-chaired by Song Sang-hoon, President of the TTA, and Choi Kyung-jin, President of the Korean Association for Artificial Intelligence Law), Kim Ji-hoon, Head of the AI Business Division at SKTelecom(017670), explained the market shift toward action-oriented agents, citing the Silicon Valley startup Pickle as an example.

Pickle is an Agent AI service created in Silicon Valley by Korean founders. It accesses external websites through its own cloud browser and performs tasks using accounts linked by the user. A key feature is its ability to integrate with domestic services such as Naver, Kakao, and Government 24.

It supports product searches, ordering, and payment for shopping; reservations at restaurants and hospitals; and KTX ticket bookings. It also checks for canceled tickets on sold-out trains, sends notifications, and makes reservations upon user request. It can handle tasks that were previously processed through customer service centers, such as changing mobile carrier plans or requesting refunds. It also supports product searches and orders on shopping services like Coupang and Daiso Mall.

Since its launch on October 1, the service has become the biggest talking point in the IT industry. An industry insider stated, “Pickle excels at naturally handling tasks such as restaurant reservations and orders,” adding, “The scope of AI’s activities in real-world service settings is expanding to the point where there are reports of self-employed business owners being surprised by calls from AI agents.”

Kim Deok-jin, director of the IT Communication Research Institute, also drew attention by sharing his personal experience using Pickle. In a Facebook post, Director Kim recommended Pickle, stating, “I’m sharing this because I was genuinely impressed after trying it myself—without any sponsorship or special compensation.”

He commented, “It’s easy to set up, and I’m currently using it to handle tasks by connecting various services—it works really well,” adding, “From a Korean user’s perspective, it’s far superior to Muse in terms of usability.” While noting that he is currently using it to handle actual work by connecting multiple services, he added, “I just hope there aren’t any security issues.”

Meta’s Agent AI Muse, which Director Kim compared Pickle to, has not been officially launched in Korea. It appears he highly rated Pickle for its integration with domestic platforms and ease of use.

Search- and Ad-Centric Platforms Are Being Transformed by AI Agents

Traditional digital platforms were structured to connect buyers and sellers through search and generate revenue through advertising. Users compared search results and directly selected the products or services they wanted to purchase.

In the era of action-oriented agents, this process changes. When a user states their desired conditions, the Agent AI searches for options, narrows down the candidates, and then carries out follow-up procedures such as making a reservation or completing a purchase.

Kim cited the example of a request to “book a quiet dining spot for four people on Friday evening.” The agent first verifies the requester’s identity and identifies the number of people, time, location, and preferred conditions. It then searches for restaurants, narrows down the options, and confirms the reservation details with the user. Upon approval, it finalizes the reservation and handles schedule registration and notifications.

“The agent acts on my behalf,” explained Director Kim, adding, “It’s important that it not only conducts searches for me but also carries out actual tasks based on those search results.”

This shift could pose a threat to existing platform operators. If users entrust tasks to Agent AI agents instead of accessing Naver or other platforms directly, control over customer touchpoints and the transaction process could shift to the agent providers.

A prime example is the practice of writing Naver Shopping purchase reviews on behalf of users. Reviews serve the dual purpose of providing users with points while helping the platform accumulate product-related content. If an external agent writes reviews on behalf of users, the platform must manage not only the customer touchpoints but also the authorship and quality of the reviews, as well as the criteria for earning points.

Citing the example of Pickle writing purchase reviews on Naver Shopping, Director Kim noted that platform operators are inevitably faced with the dilemma of whether to allow or block access by external agents.

Beyond API Openness: A Framework for Permissions, Security, and Accountability Is Needed

For executable agents to become widespread, service providers must open their functions to external AI. This involves enabling agents to call service functions—such as reservations and purchases—through APIs (Application Programming Interfaces) or MCP (Model Context Protocol). For services without official APIs, computer manipulation techniques—where agents interact with screens via a browser in the same way a human would—are also utilized.

However, platform operators’ responses have been mixed. This is because if external agents take over customer touchpoints, it could weaken the existing platform’s customer relationships and control over transactions. The development costs and liability issues associated with opening services to agents also pose a burden.

Director Kim cited the example of Korail blocking access to certain Agent AI services. In the U.S. as well, responses vary by operator: while Amazon has blocked Meta’s agents from making purchases, Shopify has been proactive in utilizing agents.

Security and access control also pose challenges. When agents perform tasks involving financial liability—such as stock trading or payments—user authentication, prior approval, and transaction log management become critical. This is because significant damage could result if the permissions granted to agents are misused or if authentication credentials are compromised.

Pickle explains that it encrypts and manages passwords and card information, and for critical tasks such as payments or reservations, it displays the details and costs to the user before obtaining approval. This is the security policy disclosed by the service provider; it does not imply that it has undergone independent security verification.

Director Kim emphasized that observability and traceability—the ability to monitor and track the agent’s execution process—as well as a secure operational framework between the service and the model will become increasingly important.

“As competition in Agent AI accelerates, security and trust are what matter most,” he said, adding, “Service providers and partners must work together to build the ecosystem.”