Meta Repeats History: AI Agent Muse Turns Out to Be Assisted by Human Staff

Meta is back in the spotlight after it emerged that the expansion of its Muse artificial intelligence (AI) agent service is not entirely run by machines. Some outbound calls made by Muse for business purposes—such as asking about prices or making reservations—are reportedly still handled by trained human staff. This fact surfaced in a report by 404 Media citing internal leaks from Meta employees, and immediately sparked discussion about how autonomous the AI agent touted as a breakthrough really is.

The announcement of the expansion of Muse's outbound call capacity was made by Meta Chief AI Officer Alexandr Wang last week. According to the report, Meta is testing a mechanism in which some AI calls and certain requests are routed to human agents. The goal makes sense: in the early launch phase, Meta wants to ensure the service runs smoothly while refining the system. However, this move reminds the public of a similar strategy that failed more than a decade ago.

In 2015, Meta launched M, a virtual assistant inside Messenger that was claimed to be able to answer questions and carry out user tasks, including calling restaurants for reservations. Behind the scenes, M was supported by AI and human trainers ready to help when the bot ran into difficulty. Users were not told which parts were handled by humans and which by the machine. Although it worked, this hybrid model was difficult to scale because operating costs ballooned, and user interest turned out to be low. The case was revealed by Casey Newton of The Verge in 2018, who said he did not find much benefit in M. When Newton told Meta CEO Mark Zuckerberg about his experience, the CEO acknowledged that the experience represented many other users.

The failure of M did not immediately make Meta abandon the idea of AI agents. The company remains convinced that intelligent assistants will improve users' lives, whether people like it or not. However, a crucial question arises: if the hybrid model failed before, why repeat it now? Some analysts believe advances in AI technology—such as more sophisticated large language models—may make this approach more economically viable. Nevertheless, the presence of humans in the service chain still raises questions about scalability and consistency.

Equally important, Meta reportedly once tested what was called a "human concierge" to complement Meta AI responses with real human staff. However, the concept was later cancelled due to concerns that users' personal data could be exposed to human contractors. This shows that privacy is a serious consideration limiting how far humans can be involved in AI operations.

For Indonesia, this development offers valuable lessons. The country's large digital market and rapid AI adoption make this kind of hybrid model potentially testable here. However, labor costs and increasingly strict personal data protection regulations—through the PDP Law—are determining factors. If Meta or other players want to deploy AI agents with human support in Indonesia, they must ensure compliance with local rules and transparency toward users. In addition, an ethical question arises: to what extent do users have the right to know whether they are interacting with a machine or a human?

Going forward, the direction of Meta's AI agent development will be a barometer for the industry. If Muse successfully passes the hybrid phase and moves to full autonomy, it could become a blueprint for similar services. Conversely, if history repeats itself, Meta may need to review its strategy. What is clear is that the public and regulators will continue to scrutinize how tech giants balance innovation, efficiency, and responsibility.