Uber: AI agents booked just 4 rides and 500 meals out of 320 million during Halloween week

Over the past year, we have repeatedly been told that personal AI agents will take routine tasks off our hands: booking rides, ordering food and buying things. Startups are receiving valuations in the tens of billions, while apps such as Meta Muse are rising to No. 1 in the App Store. Then a public company's financial results come out, and the picture becomes far more modest.

What Uber revealed

On the third-quarter earnings call, CFO Balaji Krishnamurti said that during Halloween week, October's busiest, Uber and Uber Eats processed about 320 million rides and orders. AI agents accounted for just four rides and 500 food orders. That is hundredths of a percent of the total volume—essentially a statistical rounding error.

CEO Dara Khosrowshahi described this as “very early and very limited traction,” adding that the company expects growth. A week earlier, DoorDash reported almost the same thing: agents generated only a few hundred orders, while the overwhelming majority of customers still used the regular app.

Why an agent cannot beat a button

Uber's main argument is straightforward: booking a ride in the app takes an average of 50 seconds, while reordering a favorite meal takes about 90 seconds. The apps have been optimized for these scenarios for years, with saved addresses, order history, payment details and one- or two-tap flows.

Now imagine using an agent. You have to formulate a request, wait for a response, verify that it selected the right fare and pickup point, and confirm payment. For a familiar, recurring task, this is no faster and can sometimes be slower. There is also the issue of trust: a mistake in the address or price costs real money.

Agents win when the task is complex and loosely defined: “Find dinner for six with vegetarian options near the hotel and make a reservation.” There is no point handing them a simple one-tap order.

What this means in practice

Uber is nevertheless relaxed about the situation. Khosrowshahi sees agents as another customer-acquisition channel and expects them to provide a “net positive” for demand. According to him, every AI company is talking to Uber; there are currently around eight major assistants, and that number will grow beyond ten. For the platform, this is inexpensive: connecting to another channel is easier than building it itself.

The takeaways for developers and product teams are:

• Do not build an agentic flow on top of a task that the app completes in a minute. First measure the time and number of steps in your own product.
• Look for tasks with high cognitive load: comparing options, planning and multistep bookings.
• Prepare APIs and structured data for agents in advance. It is inexpensive, but do not expect a quick return on investment.
• Start tracking the “orders through agents” metric separately now so that you can spot an inflection point if it arrives.

Limitations

These are figures from one week and two companies. They do not mean that agents will never take off: the early numbers for the web and mobile apps also looked modest. It is also unknown how many users ask an agent for recommendations but then complete the order themselves in the app. Such journeys are not visible in agent-order statistics.

The conclusion is simple: hype around agents is running ahead of actual human behavior. For now, the convenience of a regular app is winning, while the advantage of agents lies in complex tasks—not in replacing an “Order” button.