Meta presents Muse as an artificial intelligence agent capable of sending emails, preparing shopping lists, or calling a company to complete a task on our behalf. But Reuters reveals that the company also tested a system called a "human concierge": subcontractors could handle certain calls that users thought they were entrusting to their digital agent. The trial was reportedly suspended following internal privacy concerns. This case raises a simple question: when a service presents itself as AI, must it clearly state that a human can access our request?
Key points to remember
- Reuters reports that Meta has tested human operators to carry out some calls handled by Muse.
- These subcontractors could be exposed to sensitive information transmitted during the preparation of the call.
- The test was reportedly suspended due to the risk of data leaks and internal concerns.
- Meta also claims to have designed Muse as a private and secure agent, running in a dedicated IT environment.
- Occasional human intervention is not necessarily misleading if the user is informed and can choose. The problem stems primarily from a lack of transparency.
- The case illustrates a current limitation of AI agents: speaking naturally is easy, reliably performing a real task remains much more difficult.
What is Muse?
Launched by Meta on September 8, 2026 in the United States, Muse is presented as a personal agent, not a simple chatbot. The difference lies in the action. A chatbot answers a question; an agent receives an objective, prepares several steps, and uses digital services to achieve it.
According to Meta, Muse can help users organize their calendars, send messages, plan trips, search for products, or turn long-term goals into action plans. The service is accessible through a dedicated app and can also interact with users within WhatsApp.
To accomplish these tasks, the agent needs access to more information than a typical search engine: contacts, calendar, emails, preferences, purchase history, or project-related data. The more useful the assistant becomes, the more crucial the issue of confidentiality becomes.
Why did Meta resort to human operators?
Making a phone call seems simple for a human, but it's a complex task for a machine. The agent must understand the request, dial the correct number, get through a switchboard, explain the reason for the call, interpret an unexpected response, and sometimes negotiate or confirm an expense.
According to internal documents seen by Reuters, Meta experimented with a "human concierge" to study the reactions of people who did not want to speak directly to an AI. Call center employees could then discreetly make certain communications to book a service, check stock levels, or request a price.
Human intervention can also help prevent the failure of a still-imperfect product. When an agent doesn't understand a user or encounters an unusual procedure, an operator can complete the task. This method is sometimes used in emerging digital services to test a feature before fully automating it.
Where does the problem lie?
A hybrid service is not automatically problematic. Many companies combine software and human intervention. An anti-fraud system can forward a case to an analyst; a medical assistant can have an alert reviewed by a professional.
The difference lies in consent. If a person asks an AI to call a clinic, an insurance company, or a bank, they may be transmitting information they wouldn't have shared with an unknown operator. Therefore, they need to know whether their request will be processed by a computer system or if a subcontractor will be able to listen to, read, or rephrase it.
Reuters reports that Meta employees were concerned about the potential transfer of sensitive data to call centers. The test was subsequently suspended. This suspension does not mean that all Muse calls were handled by humans, nor that the system is still in use today. It does, however, highlight the need to clarify the line between automation and human assistance.
How does an agent normally operate when making a phone call?
A voice agent combines several technologies. A model transforms the user's request into an action plan. A speech synthesis system pronounces the sentences. Speech recognition transcribes the user's responses. Finally, the model chooses the next action: asking a question, accepting a proposal, requesting clarification, or soliciting the user.
Each step can fail. A product name might be misunderstood, the price might change, an important condition might be overlooked, or the agent might approve an expense without fully understanding the answer. To mitigate risks, the user must be able to define rules: maximum amount, prohibited information, mandatory validation before payment, and limited data retention.
What precautions should be taken before using an AI agent?
- Check if humans or subcontractors can access the requests.
- Limit permissions to services that are truly necessary.
- Avoid transmitting confidential medical, banking or professional data.
- Require confirmation before a purchase, reservation, or sending of an important message.
- Monitor the history of actions performed and remove unused access.
Our reading
The real issue isn't whether humans have definitively disappeared from customer service. It's who is acting, with what data, and under what responsibility. An agent can be very convenient while remaining a hybrid. But the user must be able to distinguish between automated actions, human intervention, and decisions that involve their money or identity.
Meta's promise—a personal AI capable of acting in the real world—will only be sustainable if trust accompanies automation. Otherwise, every spectacular demonstration will raise the same question: is it really the AI doing the work, or someone behind the scenes?
Frequently Asked Questions
Were all of Muse's calls made by humans?
No. Reuters describes a test involving a human reinforcement system. The available information does not allow us to confirm that all calls were affected.
Is the "human concierge" still working?
According to Reuters, the test has been suspended. This experiment should be distinguished from Muse's other automated features.
Is Muse available in France?
The initial announcement was for the United States. Availability, features, and privacy policies may vary by country.
