Meta Platforms told employees that human callers handling requests for its Muse AI agent achieved success rates of 95%-98% in some tests, outperforming AI-only calls, though the company did not disclose the AI-only rate. Muse, an AI agent that can call businesses to book appointments, check product availability and obtain quotes, began phone-call testing with employees in August. Meta later gave half of its employees access to a human concierge, allowing trained contractors to complete tasks Muse could not handle. The experiment was rolled back after an employee said a contractor made a racist reference while negotiating a cable bill; Meta apologized and said the contractor would no longer work on its projects. The test also exposed a broader limitation: businesses or individuals may refuse to interact with an AI agent even when the technology can complete the task. Meta said employee feedback was overwhelmingly positive and that it would publicly launch human assistance only when ready and with proper disclosures. The approach recalls Facebook’s M assistant, which reportedly relied on people for about 70% of requests before being shut down in 2018. Muse has become important to Meta’s effort to turn heavy AI investment into a consumer business. Meta shares have risen more than 20% since Muse launched on Sept. 8, adding over $200 billion in market value, while the agent recorded 2.8 million downloads in its first 12 days. Polymarket traders assign a 74% chance of Meta reaching $760 before the end of September, with about $70,000 wagered. JPMorgan said Muse could become the most widely used consumer AI app since ChatGPT, and Jefferies estimated $10.8 billion in annualized revenue if it reaches 1 billion users by the end of 2027 and converts 3% to paid plans.