Meta’s new Muse app promises to handle the tedious online tasks people usually do themselves, from booking flights and restaurant tables to shopping and updating calendars. But early glitches — and resistance from major companies — show how difficult it may be to turn AI agents into reliable digital assistants.
An AI that does the clicking for you
Muse represents a different approach to consumer AI. Instead of simply answering questions or generating text inside an app, the agent is designed to navigate the internet and complete tasks on a user’s behalf.
The idea is straightforward. Tell Muse that your child needs textbooks for college, for example, and the app can check course requirements and attempt to order the books. Users can store login details in a secure vault and use Link to make payments.
The concept has obvious appeal. Much of online life consists of repetitive tasks that consume time without offering much satisfaction — searching for products, making reservations, updating calendars or filling out forms.
Muse has also received significant attention, becoming the top-ranked app on Apple’s App Store. But its biggest challenge may not be the technology itself. It is the companies whose websites the agents need to access.
Amazon, for example, has blocked Muse from shopping on its platform. The conflict reflects a fundamental difference between an AI agent and the traditional online shopping model.
When a person shops on Amazon, the company has opportunities to show advertisements, promote Prime and encourage additional purchases. An AI agent, by contrast, has little reason to browse, explore recommendations or respond to marketing. It simply wants to complete the transaction.
Amazon has also been involved in a dispute with AI company Perplexity over similar shopping-agent technology.
Meta has had more success elsewhere. Expedia announced a partnership with Muse, suggesting that some companies may see AI agents as a new route to customers rather than a threat to their existing business models.
When robots compete for restaurant tables
Restaurant reservations could become another major battleground.
On paper, booking a table is exactly the kind of task an AI agent should handle. A customer knows where they want to eat, while the agent can search for availability and make the reservation without forcing the user to repeatedly refresh a website.
But widespread adoption could create a new problem: agents competing against one another for limited tables.
If hundreds of customers send AI agents to search for reservations at the same popular restaurants, the process could become an automated race. There are also concerns about people using agents to reserve multiple tables or failing to realize that an agent has successfully made a booking, potentially leaving restaurants with empty seats.
Reservation platform Resy has already suspended a user’s account after detecting AI-agent activity, although it remains unclear how such activity fits within the platform’s official rules.
The more companies restrict AI agents, the less useful those agents become.
A 30-minute flight-booking failure
The limits of current technology became apparent in one test of Muse.
The writer attempted to use the app to book a flight for himself and his wife. After finding the desired flight through Google Flights, he gave Muse access to his airline credentials and asked it to complete the booking.
The process quickly ran into a CAPTCHA. Muse asked whether it should attempt to complete the human-verification test, but failed. The user then took control, logged in and handed the process back to the agent.
A few minutes later, Muse was kicked out of the browser.
After roughly 30 minutes, the booking had gone nowhere. The writer eventually completed it himself in about five minutes, paying slightly more than $500 for the flights.
A second test was even more basic. At 12:40 p.m. Eastern time, Muse was asked to add an event to the calendar for 10 a.m. the following day. Instead, it created the event for a time that had already passed by two hours and 40 minutes.
For an app whose central promise is taking everyday digital chores off users’ hands, reliability is crucial.
The bigger fight may be unavoidable
Despite the problems, the underlying technology is advancing quickly.
AI systems are increasingly capable of navigating websites, clicking through interfaces and working with complex computer applications. More powerful systems can already perform tasks such as reconciling financial spreadsheets and updating accounting software.
The question is whether those capabilities can be made reliable enough for ordinary consumers — and whether the companies whose websites agents need to access will allow them to operate freely.
Some companies may welcome the technology. Microsoft, for instance, has been developing systems designed to work with AI-driven shopping and brand agents.
Others have strong reasons to resist. AI agents could reduce advertising exposure, disrupt business models, create security problems or generate huge numbers of automated requests. Banks have also warned about potential fraud, scams and privacy risks associated with AI shopping bots.
That could slow adoption even if the technology itself continues to improve.
The future of AI agents may therefore depend on more than whether they can learn to click the right buttons. They also need permission to enter the digital spaces where those buttons exist.
For now, Muse is showing both sides of that equation: AI can increasingly perform internet chores, but it is still far from doing them reliably — and many of the companies controlling the internet have little incentive to make the process easy.
Be First to Comment