AI agents can order your dinner. The numbers show nobody trusts them to

Ordering dinner is the perfect AI agent demo: a clear goal, a familiar app, a satisfying "done" at the end. So why is almost nobody doing it for real?

The numbers are stark. On its third-quarter earnings call, DoorDash CEO Tony Xu said, in the words Ars Technica quotes, that "AI agents are generating several hundred orders [on DoorDash] across the time that they've been up and running." He called that a "very low use case" for a company that handled just shy of 1 billion orders in the quarter. Days later, TechCrunch reported that Uber counted only four rides ordered through an AI agent, plus around 500 Uber Eats orders placed by agents in the week of Halloween, when the company handled roughly 320 million trips and orders.

Those are rounding errors. And they arrive while personal-agent startups raise money at eye-watering valuations and Meta says its Muse app topped 10 million users.

The models aren't the bottleneck

The easy reading is that agents aren't good enough yet. We don't buy it. Ordering food is among the simplest tasks an agent could be handed: pick a restaurant, pick items, pay. If capability were the problem, we'd expect failures on harder jobs, not a near-total absence on the easy one.

The more convincing explanation is convenience. Uber's CEO Dara Khosrowshahi pointed out that booking a ride in the app takes about 50 seconds on average, and re-ordering a favorite meal about 90 seconds. An agent has to beat that, and it also has to be briefed, supervised and double-checked. Saving a minute is hard when you spend one explaining what you want. For a task this routine, the app is already the shortcut.

Trust is a bigger obstacle than it looks

Food is a decision people care about in small, personal ways: no cilantro, not that place again, not at that price. Handing it to a bot means accepting that it will spend your money and be wrong sometimes, with no obvious undo button. That's a control problem, not an intelligence one.

The wider environment doesn't help. Security firm Elastic's new threat report, covered by ITWire, warns that an agent holding credentials and internet access is a new attack surface, and that when an attacker is behind it, its activity looks exactly like the employee who installed it. That's about enterprise tools, not food apps, but the instinct carries over. An agent with your payment details and your account is a lot of authority to delegate for a burrito.

We'd also note a thin-data caveat: one Hacker News commenter argued the headline is misleading because the integration is available to almost no one anyway. A Winzheng write-up on the DoorDash story describes the connector as still limited to a small number of "early beta partners," mainly used for corporate lunch orders. That's a fair point, and it means low usage partly reflects low access. But it cuts against the hype too: if the plumbing isn't open, "agents will replace apps" is a long way off.

The platforms have no reason to hurry

There's a third factor that has nothing to do with users. As The Verge notes, platforms like Uber and DoorDash have little incentive to let agents barge in and replace their own customers. Uber says it will work with third parties but bets its direct services will be the most "delightful and easy," and it's developing its own AI tools, including a Cart Assistant for Uber Eats.

That's rational. If an agent owns the customer, the delivery app becomes a commodity back end. We expect cooperation in words and friction in practice.

What we'd take from this

The lesson isn't that agents are a bust. It's that "can it do the task?" is the wrong test. The right ones are: will people let it spend their money, can they see and reverse what it did, and is it actually faster than the thing they already do in 90 seconds?

Until an agent clears those three, demos will keep outrunning behavior. If you build or buy agents, we'd focus on narrow, reversible, clearly confirmed actions rather than "do everything" promises. And if you're an investor or executive, treat dinner as the benchmark: a few hundred orders against nearly a billion is the honest scoreboard, and for now trust looks like it's winning.