AI agents brought Uber just 4 rides out of 320 million weekly orders

Uber ran a rare and therefore especially valuable experiment: instead of showcasing an impressive AI-agent demo, the company revealed how users actually behave. With roughly 320 million rides and food orders per week, agents handled about 500 deliveries and just four rides.

The numbers look almost comical, but the conclusion is serious. Users do not hire a digital intermediary merely because it is called an agent. If a familiar button already completes a task quickly and predictably, an additional conversational layer is more likely to get in the way than help.

Why the agent lost to the conventional interface

Booking an Uber ride takes about 50 seconds. Repeating a familiar Uber Eats order takes around 90 seconds. The user opens the app, sees the price, checks the address, and taps a button. The task has a clear flow, few steps, and immediate feedback.

Now imagine an agent. The user must explain where to go or what to order, then verify that it correctly understood the address, restaurant, order contents, and price. Even if the agent is technically capable of completing the action, the user still bears the risk of an error. Saving a few taps does not offset the need to supervise an intermediary.

This is where a common product hypothesis breaks down: a conversational interface does not automatically become more convenient than a graphical one. For short, frequent operations, a button is often more efficient than a dialogue. It displays the available options, constraints, and price before confirmation. An agent must either repeat all of that or conceal it behind its own decision.

Where an agent can genuinely be useful

The problem is not agents as a technology, but the scale of the chosen task. Booking one familiar ride is too simple. Organizing a complex itinerary is more compelling: selecting a departure time around a flight, booking a car for several passengers, arranging a child seat, comparing the price with public transport, and changing the booking if the flight is delayed.

The same applies to food delivery. Repeating yesterday's order takes one button. Building a weekly menu for a family while accounting for allergies, budget, delivery windows, and several people's preferences is a chain of decisions. Here, an agent can save not seconds but tens of minutes of mental work.

For developers, there is another important boundary: any action involving money requires transparent confirmation. The agent must show exactly what it selected, how much it costs, and what data it used. Critical parameters cannot be hidden behind the phrase “I arranged everything.” The higher the cost of an error, the more important the activity log, cancellation option, and clear handoff to manual control become.

What product teams should test

Do not begin with the question, “Where can we add an agent?” It is more useful to measure the existing process: how long it takes, how many decisions it requires, and where users make mistakes or abandon the task. Then compare the agent-based flow with the conventional interface by completion time, success rate, number of corrections, and repeat usage.

Uber's experiment is a reminder that technical capability does not equal product demand. If an agent replaces three taps with a long dialogue, it makes the product worse. If it gathers data from several sources, coordinates participants, and independently completes a multistep process, its value becomes tangible.

The practical conclusion is simple: look not for operations that AI can perform, but for tasks that people genuinely struggle to complete on their own. An agent is justified when it removes an entire chain of routine work, uncertainty, and supervision. For everything else, a good button remains the stronger interface.