Meta implemented a test using a type of 'human concierge' for its new personal artificial intelligence agent, Muse. According to internal company documents seen by Reuters, this initiative involved employees discreetly handling some of the phone calls that were supposed to be managed by the digital agent.
This test was notified to employees the week before the public launch of Muse's phone calling feature. However, the implementation caused apprehension among staff, as some expressed concern that human involvement in these conversations could lead to the unintentional sharing of confidential data with the attendants.
A Vice President of Meta's SuperIntelligence Labs admitted in an internal communication that it was a mistake to start tests of calls made by contractors without providing adequate information to users. For this reason, the company decided to temporarily suspend this feature. The executive also stressed that Meta continues to improve the functionalities of automated AI calls.
Some trials demonstrated that human intervention could increase the success rate of calls to between 95% and 98%, surpassing the percentage achieved when the process is entirely executed by artificial intelligence.
More information about the test
While developing new capabilities, Muse has shown high acceptance in the United States. Data provided by market intelligence company Sensor Tower indicates that the agent led app download rankings in the country during the two weeks following its launch, accumulating over 2.5 million installations.
Muse constitutes a fundamental element in CEO Mark Zuckerberg's project to provide 'personal superintelligence' to the billions of daily users of Meta's services. This AI agent has the ability to perform tasks autonomously, such as scheduling trips, making purchases, or sending correspondence on behalf of the user.
At the time of the launch, Meta strongly emphasized the assistant's security measures. The company guaranteed that each agent would have its own secure 'virtual machine'—a dedicated cloud computer designed to keep user sessions isolated—and that sensitive data, such as passwords, would be stored in a protected and separate environment.
With the calling functionality, the user can ask Muse to contact companies in the United States and interact with people on the other end of the line to fulfill various obligations. Examples include scheduling a haircut, confirming item availability in a store, or obtaining quotes from different service providers. After the call is completed, the agent sends the user a summary and a transcript of the conversation within the Muse app.
Meta began internal testing of this calling feature in August under the method known as 'dogfooding,' and later started releasing it progressively to Muse users after the official application launch. Last week, the company activated the 'human concierge' feature, also called 'calls made by human agents,' for half of its own employees. As detailed in an internal publication, Muse gained the option to direct certain requests to a trained agent, who would make the call and manage the interaction. Employees who chose not to participate in the test had the option to join an exclusion group.
