Uber, DoorDash, and Airbnb state that external AI agents are very rarely used to place orders or make bookings, according to figures presented in their third-quarter 2026 financial results. This data, which notably concerns Meta’s Muse and OpenAI’s Dots, calls into question the promise of time savings, while security and privacy cases are also mentioned.
Almost Non-Existent Usage for Orders
During Halloween week, out of approximately 320 million rides and orders recorded by Uber, only four rides were booked by an AI agent, along with about 500 orders on Uber Eats. Uber CEO Dara Khosrowshahi explains this low level by the speed of the tasks involved: booking a ride takes an average of 50 seconds, often with a single click, and ordering a meal 90 seconds.
DoorDash reported “a few hundred” orders placed by agents, out of billions of annual orders, according to data released by the company. For his part, Airbnb CEO Brian Chesky described the share of bookings made via agents as “completely negligible”, out of its 500,000 daily bookings.
For these executives, delegating a deterministic and ultra-fast task to a conversational agent introduces unnecessary friction and risks of misunderstanding. The reasoning applies to operations that can be performed in a few clicks, an AI agent being a program capable of acting on the user’s behalf.
Security and Privacy Risks
The sensitive personal and banking data required to book a service also raises security and privacy questions. The cybersecurity research community and specialized journalists have highlighted several cases, some of which are disputed.
Journalist Jason Aten claims to have observed that Muse for Mac suggested writing an article based on a private SMS or iMessage conversation with his podcast co-host, even though he had denied access to his messages and calendar during setup. According to his analysis, the app was syncing Apple Messages’ local database (chat.db), already at line 187,462. Executives at Meta Superintelligence Labs denied any bypassing of macOS protections, citing a hallucination by the language model and a misinterpretation of sync mechanisms. These allegations have not been independently verified.
A second case concerns a sale on Facebook Marketplace. According to content creator Matt Robb, the Muse agent, tasked with selling a keyboard, accepted a price lower than desired and transmitted the seller’s home address to the buyer without human validation. The buyer showed up at the home while the seller was absent, and the latter had not been notified by the app. Meta, however, states that the agent requests systematic confirmation for sensitive actions.
