Claude Duet, Kanana, Muse: The Era of AI Assistants That Proactively Speak Up Begins

AI assistants are moving away from question-and-answer chat windows. They are shifting toward checking linked email and calendars, proactively alerting users to issues, suggesting needed information during conversations, and continuing assigned tasks while the user is away. As a result, the scope of competition among AI assistants is also expanding beyond answer accuracy. That expanded scope covers what to notify users about, when to do so, and how far the assistant should go in acting on their behalf.

Examples of this trend include Anthropic's 'Claude Duet,' Kakao's 'Kanana,' and Meta's 'Muse.' The three services have different points of contact across work, messaging, and daily-life management. What they have in common, however, is their ability to understand the user's situation.

There are also differences. Those differences appear in the scope of proactive information suggestions, the scope of actual task execution, and the stage at which the service is offered.

On the 21st, Anthropic unveiled a research preview of Claude Duet for professionals and small business owners. It will be rolled out sequentially to U.S. Claude Pro and Max subscribers after they join a waitlist.

Claude Duet is an active agent that summarizes key points from connected apps into a morning briefing and sends proactive messages when additional confirmation is needed. Its defining feature is that information checking and task execution are linked.

Users can identify priority tasks based on email and calendar information, and delegate work such as meeting preparation and document drafting. Claude Duet runs on a cloud computer with its own browser, and it can continue working even after the app is closed.

Structurally, however, it does not continuously monitor connected accounts in real time. Anthropic said it checks accounts every few hours to manage usage. That creates a limitation in that important tasks may come in between check intervals.

Kakao has adopted a model in which AI is embedded in the user's existing conversation space, developing the service so that AI can understand context and offer help within KakaoTalk conversations users are already having. The focus is on reducing the burden of having to restate the situation in a separate AI chat window.

'Kanana in KakaoTalk,' a service in this direction, includes functions that proactively suggest schedule management, information searches, and product and place recommendations based on understanding conversation context, and it was officially launched in March this year.

At its developer event 'if Kakao 26' on the 13th, Kakao outlined a direction for expanding task execution and unveiled 'Kanana Space.' 'Kanana Space' is a call-based AI assistant for KakaoTalk with functions for sending messages, managing schedules, and carrying out web search requests.

Through 'Kanana Space,' Kakao said it aims to provide an experience in which users can talk with Ryan anytime and anywhere and feel naturally connected with AI. 'Kanana Space' is scheduled to undergo closed beta testing from December for some pre-registered users, and the company plans to later unveil additional characters that leverage various intellectual property (IP).

Kanana is planning to combine its existing proactive suggestion features with execution capabilities in the future. The idea is to attach an actual execution step to advance suggestions so the service can extend through ordering and payment. In that process, it will reference past order history and available coupons, and use a method that preps the ordering and payment process in advance. Future expansion plans include ordering, gifts, reservations, and remittances.

Meta announced Muse on the 8th of last month. Muse is positioned as a personal AI agent that handles life goals and to-dos across daily life. Users can talk with Muse through a dedicated app or WhatsApp. Muse performs tasks after connecting the apps and services the user has allowed.

Meta said Muse remembers preferences and conditions the user has previously mentioned. It also suggests unsolicited ideas. As a result, a cooking video saved on Instagram can be turned into a grocery list. When suggesting a menu for a dinner gathering, it can also take a friend's dietary restrictions into account.

Muse runs on its own cloud virtual machine and browser. Because of that, it can continue working even after the app is closed. Both Kanana and Muse show a trend in which AI goes beyond making suggestions and extends to preparing or carrying out real actions.

Each service shares a common direction: users should be able to receive the help they need during their existing activities without having to seek out AI separately. That common direction involves reducing the number of times users need to search for AI and providing help within the context of existing activities. There are, however, differences in task priority for the connected services. Muse places greater emphasis on continuously reflecting life goals and preferences, while Kanana prioritizes spotting needed help within messenger conversations.

For such proactive suggestions to be useful, AI must fully understand the user's situation. The usefulness of proactive suggestions depends on AI sufficiently understanding the user's situation. At the same time, as access to email, calendars, conversations, and purchase history expands, the impact of mistaken judgments also grows. It is also pointed out that expanding access to email, calendars, conversations, and purchase history increases the impact of erroneous judgments.

Accordingly, a need has been raised to separate permission to view information from permission to transmit information externally or make payments. Anthropic said that, by design, Duet requires user approval before sending messages, making payments, signing documents, or deleting files. Meta said Muse applies confirmation steps to sensitive actions such as sending emails and making purchases. Meta also said users can choose which apps to connect and the scope of access granted.

In this context, Kakao announced plans to further strengthen its agent safety framework in the future. Kakao plans to design the system so that only the information users need is delivered within the permitted scope. It also plans to design the system so that actions are taken only within the scope of consent.

An industry source said the competitiveness of AI assistants lies in timely alerts about tasks users may have missed and in reducing the burden of handling them. The source added that personalization requires the use of necessary information, but users need the ability to control external information transfer and payment actions, and that such user control is a condition for building trust.

Source: IT DAILY · Lee Jae-young
Original: https://www.itdaily.kr/news/articleView.html?idxno=242441

References

• https://www.itdaily.kr/news/articleView.html?idxno=242441