Meta’s new Muse personal AI agent is supposed to handle real-world tasks for users, from making bookings to contacting businesses. But behind some of that automation, there may still be a human on the other end.
Meta has been testing the use of human contractors to handle certain phone calls initiated through Muse, according to reporting from Reuters and Social Media Today. This arrangement is part of Meta’s effort to expand what its AI agent can accomplish. The technology is still developing.
It also brings back an old question for AI assistants: how much of an “AI” experience is actually automated?
Meta Muse Can Call Businesses on a User’s Behalf
Muse is Meta’s latest push beyond the familiar chatbot format. Rather than simply generating answers, the personal agent is designed to take action. Meta says Muse can send emails, book travel, fill out online forms, make purchases and work across connected services. Additionally, phone calls are becoming part of that strategy. This gives users a way to ask the agent to contact businesses without making the call themselves.
Meta Chief AI Officer Alexandr Wang has highlighted outbound calling as one of Muse’s expanding capabilities. In practical terms, the feature could allow the assistant to contact a restaurant, hotel or other business to check availability, confirm information or complete a reservation. Moreover, it fits the broader shift toward agentic AI, where assistants are expected to perform tasks rather than simply explain how to do them.
Some Muse Calls Have Been Handed to Human Contractors
Reuters reported that Meta has tested what it internally called a “human concierge” system. Under that setup, human contractors could step in and handle certain phone calls initiated through Muse. This happens when the AI was unable to complete the conversation on its own.
That approach makes sense from a reliability standpoint. Real phone calls are messy. For example, businesses put callers on hold, employees ask follow-up questions, automated menus change and unexpected details can quickly derail a scripted AI interaction. As a result, a human operator can help bridge those gaps while Meta continues developing the underlying technology.
The presence of human contractors, however, makes the product less straightforward than a fully automated assistant. In addition, users may believe they are interacting with an AI system while part of the task is quietly being completed by a person behind the scenes.
Privacy Quickly Becomes Part of the Conversation
Human involvement also creates privacy questions. Reuters reported that some Meta employees raised concerns that contractors handling Muse calls could potentially hear or receive sensitive user information while carrying out requests.
Meta said the testing process was intended to gather feedback and help the company determine the right privacy and safety protections before any broader rollout. The company has also indicated that appropriate disclosures would be needed if human assistance becomes part of the experience.
That detail matters because Muse is designed to work with personal context and connected services. Meta says users can control which services Muse can access and how much permission it receives. Certain actions can also require direct user approval before the agent moves forward.
Introducing human operators into that workflow changes the privacy calculation. Even limited human access could become an important trust issue if users are not clearly told when a person may be involved.
Meta Has Tried the Human-Assisted AI Model Before
Meta has experimented with this type of hybrid model before. In 2015, Facebook introduced M, a Messenger-based digital assistant that relied on artificial intelligence. Human trainers could step in when the system could not complete a task.
The concept was ambitious for its time, but it proved difficult to scale. Human support increased operating costs, while the product failed to attract enough sustained demand. Facebook eventually shut down the broader M assistant experiment in 2018.
Muse arrives in a much more advanced AI environment. Current models can navigate websites, understand complex requests, use tools and coordinate multi-step workflows far more effectively than consumer AI systems could a decade ago. Therefore, human help may play a smaller role this time, acting more as backup than as a core part of the service.
Muse Is Designed to Do More Than Answer Questions
Meta introduced Muse as a personal AI agent built to perform tasks rather than stop at recommendations. The system can reportedly browse websites, fill out forms, manage projects, send emails, arrange travel and carry out longer-running tasks across multiple services.
Muse can also continue working after a user leaves the app and return when it needs approval or has an update. Furthermore, that background capability is central to Meta’s pitch. It moves the product closer to an autonomous digital assistant rather than a traditional chatbot.
Meta has also created security controls around the service. Users can limit permissions, approve sensitive actions and restrict what the agent is allowed to access. The company has described a separate safety system designed to interrupt certain actions when additional confirmation is needed.
Human-assisted calling adds another layer to that structure. In other words, it raises a simple but important question: when an AI agent completes a task, users may eventually want to know exactly how much of that task was actually handled by AI.
Human Help Could Solve an AI Reliability Problem
Using people as backup could help Meta improve Muse during its early development. For example, when an AI system fails to understand a caller, encounters an unusual request or gets stuck in a conversation, a human can step in and keep the task moving.
That can create a smoother user experience, especially while the technology is still being tested. Instead of ending with an error message, the service has another way to finish the request.
The bigger challenge is scale. Muse is being positioned as a consumer-facing personal AI product. If large numbers of users begin making calls through the service, depending too heavily on human contractors could become expensive and difficult to manage.
That suggests human intervention may be temporary scaffolding rather than a permanent feature. Meta could use the system to study difficult conversations, improve the AI and gradually reduce the number of situations where people need to step in.
Meta’s Bigger AI Agent Strategy Is Taking Shape
Muse reflects Meta’s wider push into AI agents that can actually perform actions across the internet. The goal is no longer just to provide answers. Companies increasingly want AI systems that can book, buy, schedule, contact and complete tasks on behalf of users.
Phone calls expose how difficult that transition can be. Websites offer predictable buttons and forms. Human conversations do not. For instance, a restaurant employee may suggest a different time, a hotel may ask for additional details, or a business may change the terms of a booking halfway through the call.
Those situations are exactly where autonomous AI still faces challenges. They require context, judgment and the ability to respond to unexpected information in real time.
For Meta, the next test will be reducing the need for human assistance while making sure Muse remains useful, transparent and trustworthy. If agentic AI is going to become a normal part of everyday life, users will likely expect clarity about whether they are dealing with software, a person or some combination of both.
Sources
Social Media Today — Meta Muse uses human staff for some functions
https://www.socialmediatoday.com/news/meta-muse-uses-human-staff-for-some-functions/831097/
Meta — Introducing Muse
https://about.fb.com/
