AI Agent Bookings Remain Rare, Uber and Airbnb Report

Uber, DoorDash and Airbnb have each put numbers on a question the AI agent industry has mostly answered with demos: are people actually letting agents make bookings and orders for them? Based on what these three companies are reporting, the answer is barely.

The numbers are tiny

The companies say AI agents have handled only a small handful of transactions in recent weeks. Uber's example is the starkest. During the week of Halloween, riders booked 320 million car rides through people's own actions in the app. Agents accounted for four.

DoorDash CEO Tony Xu described a similar pattern. Agents were responsible for several hundred orders on the platform. That's a bigger absolute number than Uber's, but DoorDash handles billions of orders a year, so the share is still negligible. Airbnb CEO Brian Chesky said that only a small fraction of bookings on the app come through agents.

The gap matters because these platforms are exactly where agents were supposed to shine. Ride-hailing, food delivery and short-term rentals involve repetitive decisions, clear inputs and a lot of comparison shopping. If agents were going to become a default way to transact, these are the places it would show up first. Instead, the traffic is almost entirely human.

Why the uptake may be lagging

Part of the explanation is that the tasks agents are meant to speed up often aren't slow to begin with. Uber's chief financial officer, Balaji Krishnamurthy, pointed out that a rider can book a trip in about 50 seconds. There isn't much drudgery to automate. An agent that has to parse a request, confirm preferences and then place the order can easily end up slower than tapping a few buttons.

Airbnb's Chesky offered a different reason. He suggested that many travelers enjoy researching trips themselves, and that an agent can't judge what a rental actually feels like or whether the neighborhood suits you. A booking involves taste and context that's hard to delegate, and for a lot of people, the research is part of the trip.

That's a useful distinction. Agents look most compelling for tasks where the user already knows what they want and the outcome is easy to verify. Choosing a place to stay or deciding where to eat tends to fall into the other bucket, where the process carries as much value as the result.

The pushback is more than a marketing problem

The skepticism isn't limited to executives reading numbers off a dashboard. On Reddit, a highly upvoted comment argued that deciding what to eat or do is part of the fun, and that few people want to become "hyper-efficient at having fun." The worry is that agents strip the small, enjoyable choices out of everyday life.

Another user made a more practical complaint. Ordering a pizza through an agent meant describing the craving, then specifying the budget, instead of opening the app and tapping. Whatever the agent saves in the moment, it can add friction at the front end. That's a design problem as much as a demand problem: the agent has to be faster than the interface it replaces, and in many simple cases, it isn't.

The broader productivity picture doesn't help the case either. A recent report found that people using AI tools were sometimes working up to 90 minutes extra each day, and a separate study found workers spending more time correcting AI-generated content than doing the original work. Those findings don't measure agents specifically, but they feed a growing sense that AI's time savings are more conditional than the marketing suggests.

Some organizations are pulling back

A few organizations have moved past skepticism into restriction. The Wall Street Journal reports that United Airlines is banning AI bots from booking flights. Alibaba has stopped its workers from using Claude Code, and Sweden has blocked both Claude and ChatGPT over security concerns.

These cases differ in motive. Restricting bots on a booking channel is about controlling who and what can transact on a company's system. Limiting internal tools is about security and oversight. Still, taken together, they show that even where AI tools are available, organizations are deciding case by case where they belong.

What to watch next

My read is that the early numbers say more about fit than about the technology's ceiling. Agents may find a real role in tasks with many steps, strict constraints and little personal preference involved, such as coordinating multi-leg travel or managing recurring purchases. Those are harder to measure than a single ride, and they may not show up in the same platform metrics.

The useful test for the next year is whether agent-driven transactions grow in categories where the human process was genuinely slow. If they don't, the agentic future that vendors have been promising may look more like a set of narrow tools than a general shift in how people buy things. For now, the strongest evidence suggests that most people still prefer to tap the app themselves.