Skip to content

Four Rides Booked by Third-Party Agents in an Entire Quarter

Uber disclosed that third-party AI agents arranged four rides across an entire quarter, and about 500 Uber Eats orders in Halloween week out of roughly 320 million transactions. The 50-second booking time behind that gap says a lot about which parts of a creative workflow are worth handing over.

AI-assisted draft. Reviewed and edited by the Phosphene team before publication.

On this page
A plain brown paper takeout bag on a wet concrete doorstep at night, lit by a single bare bulb, with an out-of-focus car headlight in the background.
A few hundred orders, four rides, and one clear signal about what people will hand to an agent.

Uber finally put a number on agentic commerce. It is small enough to fit in a screenshot, and the way it was reported tells you something about how loosely these numbers travel.

On the third-quarter earnings call, Uber disclosed for the first time how much business third-party AI agents bring in. Four rides, across the entire quarter. For Uber Eats, the company recorded around 500 orders placed by agents during Halloween week, out of roughly 320 million rides and deliveries that week.

Read that again with the units straight, because they are easy to blur. The four rides are a quarterly figure. The 500 orders are one week. The 320 million is that same week, both services combined. Some coverage ran the four rides and the 320 million together as if they shared a denominator, which makes the ratio look even more absurd than it is. It is absurd enough already.

The disclosure is specifically about agents operating from outside the app, not about Uber's own in-app AI assistant. That distinction matters: the number measures whether people will let a third party transact on their behalf, which is the actual bet behind every "agentic commerce" pitch.

The comparison numbers are the useful part

The quarter was good in the ordinary way. Uber reported 4.1 billion trips and 59.8 billion dollars in gross bookings. That is the business agents are supposedly about to eat.

DoorDash gave the same story from the other side a week earlier. CEO Tony Xu said third-party agents had generated a few hundred orders in total, not a few hundred per week, but a few hundred since such agents existed. Meta's Muse, the personal agent that launched on September 8 and became the most downloaded free app in the US App Store within a week, is past 10 million users by Meta's own count, and shopping is one of its top activities. Instinct, another personal-agent startup, reportedly closed a round at a 10 billion dollar valuation.

So you have millions of agent users, billions of dollars of enthusiasm, and a transaction count you can count on your fingers. Uber is not gloating about this. Dara Khosrowshahi said on the call that he hopes there will be more traction, that Uber wants to work with third parties, and that the company wants to avoid depending on any single player. That is the language of a platform hedging, not a platform that thinks the threat is over.

Fifty seconds is the whole argument

Khosrowshahi also gave the explanation, possibly without meaning to make a general point. The average Uber user takes about 50 seconds to book a ride in the app. Reordering a favorite on Uber Eats takes around 90 seconds.

Sit with that. A delegate that takes a minute to explain the task to, and that then needs to be checked, is not saving anything against a task you finish in 50 seconds with your thumb. It is slower, it is less certain, and if it goes wrong you have to unwind it.

Compare that with the errands nobody wants. Cancelling a subscription, arguing with an airline, comparing the same pair of shoes across six shops, chasing a refund. Those take minutes to hours, they are boring in a specific way, and the result is cheap to verify: the money moved or it didn't. That is where agents have room.

I would like to say the 500 versus 4 gap proves the point, because food ordering has more open decisions in it than tapping the same destination you tap every Tuesday. It doesn't prove it. Uber did not publish separate totals for the two services, the checkout surfaces are not equivalent, and the integration work third parties have done for food is not the integration work they have done for rides. The gap is consistent with the complexity story. It is also consistent with a dozen boring explanations about API access. Treat it as a hypothesis that fits, not a finding.

The shops are also closing their doors

There is a second brake that has nothing to do with user habits. Amazon blocked Muse, citing security. The commercial logic is easy to guess: an agent does not look at sponsored placements, does not get pulled in by a deal badge, and does not impulse-buy the thing next to the thing. A shopping agent is a browser with the funnel removed. Nobody who sells ads is in a hurry to host one.

Uber's public position is friendlier than that, and still guarded. Work with the agents, keep the app good enough that nobody needs one. That is a strategy a lot of software companies are quietly adopting right now.

What this means if you build creative tools

I keep seeing creative AI products pitched as "the agent does your workflow end to end". The Uber numbers point at the wrong half of the work.

Look at what an image session is actually made of. Typing a prompt, pressing generate, and looking at the result takes about as long as booking a ride. Maybe less. Nobody wants an agent in that loop, because the loop is where the taste lives and the feedback is instant. Handing it off buys nothing and costs you the moment where you notice the hands are wrong.

The expensive part is on either side of that loop:

  • Culling 200 variants down to the six that are actually usable.
  • Checking every shortlisted frame at export size for the thing that always breaks, whether that is fingers, a logo on a shirt, or a step in a gradient sky.
  • Renaming, resizing, cropping for four aspect ratios, and writing the alt text.
  • Keeping the record of which model, which prompt, and which retouch produced the file that shipped.

That work is dull, it is measurable, and you can check the output in seconds. It is also where a competent agent earns its keep, because the alternative is you doing it at 1am with your attention gone.

The rule I take from Uber's numbers: delegate when the manual path is long, when the result is cheap to verify, and when a mistake can be undone. Skip the agent when the manual path is short and the judgment is the point. Most creative work is a mix of both, and the useful question is not "can an agent do this" but "which twenty minutes of this morning am I willing to lose".

What would move the number

Two things, and neither is a better model.

The first is agent-native checkout that shops can live with, which means identity, payment limits, and something better than a blocked user agent. That is a payments and trust problem, not a reasoning problem.

The second is a reason to delegate at all. Fifty seconds is a low bar, but it is a real one. Products that make the boring part of a task vanish will get used. Products that ask to take over the interesting part will keep posting launch threads that far outperform their transaction counts.

Uber's four rides are not an embarrassment. They are the most honest adoption metric published this quarter.

Sources

Share this guide
X LinkedIn