The first consumer AI agent that actually makes sense

I love going to the grocery store so much. There are just so many new and exciting innovations lining the shelves, and the packaging is always so pretty. New flavor of Talenti Ice Cream, sign me up. A sweeter version of a plum, I'll take three. New type of nut that can be milked, I’m confused, but I’m trying it.

Maybe I just like food.

But I dread carrying my groceries back. Those of you in the suburbs have the luxury of driving to the store and filling your trunks, but here in New York, we have to shlep our groceries all the way home. It’s killer on the forearms.

Of course, you could use services like Instacart and Uber Eats, and I’ve definitely tried them, but honestly, I can't give up perusing for new items myself at the grocery store.

This week, I’m talking about Uber Eats and how they’ve rolled out one of the first consumer-agentic experiences that actually makes sense, because it solves the thing most agents can't: getting you to trust it enough to use it.

Grab a snack and read on!

TL;DR

The thirty-second version of what Uber Eats did.

The word “agent” is in every investor report, news article, and billboard ad. Every product wants you to believe that they have an agent that can execute tasks on your behalf, but most of the time, they are just glorified chatbots.

And it's not that the tech doesn’t work. It's that we won't let it. We'll let AI suggest a recipe all day, but we get nervous the second it tries to buy, book, or cancel something on our behalf. That gap, between AI that advises and AI that acts, is the hardest thing for consumer AI to cross right now, and it's why so many "agents" retreat into being a chatbot.

That isn’t necessarily a bad thing. These chatbots can be very useful if they are grounded in real data and can provide recommendations. For example, the Google Maps AI chat provides great insight about the quality of WiFi at a coffee shop, but it’s still just a chatbot.

Earlier this year, Uber Eats launched Cart Assistant, a beta inside the Uber Eats app. You give it a list, a photo of a handwritten one, or even a screenshot of a recipe, and it builds a checkout-ready cart, leaning on your past orders to guess the right brands (your oat milk, your go-to Ben & Jerry’s flavor). Then you review it, swap anything that's off, and make the purchase yourself.

It’s an agent with an asterisk. It does work for you, but at the end of the day, you still make the final purchasing decision. It does the heavy lifting of putting together your cart and saving you time, but it defers to your judgment on what actually makes it through checkout.

Under The Hood

The three strategies that make it work.

1. Pick a bounded, checkable task.

Uber chose the perfect job to be completed by an agent: putting together a grocery cart. This task works well because of two main properties:

• Bounded = the agent isn't inventing options; it's matching your words against a finite shelf. Easy to generate a candidate.

• Checkable = your grocery list is the answer key, so you can glance at the cart and see if it covers what you asked. Easy to verify.

A bounded option set means that you don’t need to worry about all the ways the agent could go wrong because it is restricted. If a task isn’t easily checkable, like the prompt "plan my week”, you can't tell at a glance whether the agent did well, and you're back to doing the work just to check their work.

Another example that would work well is scheduling a meeting. The agent can only pick from open slots in a calendar (in theory, it could put a 9 PM Friday meeting if it hates you), and it’s easy to verify if it picked a good time.

2. Keep humans in the loop.

Unfortunately, we are not yet at the point where we can be the humans in Wall-E and have our every whim catered to by AI. AI is still far from perfect, and it needs a human in the loop (yay humans!!!).

The Cart Assistant fills the basket for a user and then hands it back for a final review. Swap the weird cereal, fix the milk, and hit buy when you're ready. The agent does 80% of the work and leaves you with the 20% of judging the items in the cart.

In a recent essay by Sequoia Capital, they argued that tasks that still require judgment can’t fully be placed on autopilot. You, as the user, have specific tastes. The agent uses all the data it has to determine the type of Oreos you want, but you need to be the one to make sure it is getting double stuffed.

In a Quad and Harris Poll survey, only 39% of Americans trust an AI agent to make everyday purchases for them, and just 34% for bigger-ticket items, describing AI as "a helpful companion rather than a replacement decision-maker". Visa puts a finer point on the fear: about half of shoppers worry about an agent processing "decisions made without me" (Visa). A cart you approve is the only design that, for now, would be adopted by consumers.

My takeaway here is that even in low-stakes situations, consumers don’t trust AI to act without some level of supervision. This makes it hard to build agents because, by definition, they work autonomously against a goal without supervision. Therefore, the journey to pure autonomy looks more like:

chatbot → human in the loop → autonomous agents

3. Own the data.

It feels like every week I end up mentioning something about data, and this week is no different. Uber has the right data to successfully run the agent.

First, it knows you and your order history. When your grocery list says milk, it knows you want Oat milk instead of having to guess. Second, it knows what is on the shelf because it pulls the retailer’s own product feed instead of trying to scrape it.

Compare that to OpenAI, which tried this first. It launched a feature called Instant Checkout that lets you buy a product right inside a ChatGPT conversation, no trip to the store's website required. The issue: OpenAI didn't have a direct product feed from those stores, so it pieced together product details by crawling and scraping their sites. That meant it was often guessing about what was in stock and what it cost, and that information went stale fast. As Forrester's Emily Pfeiffer put it, "crawling and scraping is inadequate to get the full breadth of product data that you need to do a good job of commerce”

And Uber doesn’t always get it right, but groceries have so much variety that they allow for substitution, so a slightly-off guess is a different oat milk, not a broken order. An owned context plus a forgiving category means the agent is usually right and harmless when it's wrong.

The Playbook

The Agentic Task-Fit Checklist

Three questions to ask before you let AI do a task for someone. Three yeses, and it's a real agent. Any no and you've built a fancy chatbot.

1. Can it be done? Does the task have a finite set of right answers that the user can verify at a glance? Grocery shopping: yes. "Write my marketing plan": no.

2. Will they allow it? Does the agent hand back a draft for approval instead of just completing the task? People will let AI do the work long before they'll let it make the final call.

3. Can only you do it? Does your product hold context that a generic chatbot doesn't, like past choices and real-time data, so it isn't guessing? Being wrong should be no big deal (different milk), not a disaster.

What I’d Ship

Google Flights is the obvious one, and it's pushing the idea farther than Uber. Booking a trip is a real agent's job: there are a thousand ways to get you from A to B, but they're all actual flights you can scan in seconds, so it stays bounded and checkable.

Google Flights understands all of your preferences, compares fares, assembles the trip, and takes you straight to the airline's booking page to confirm. The agent does all of the work; you just review and make the purchase.

But the stakes are higher here. A wrong carton of milk is not that bad. A wrong nonrefundable fare is a bad afternoon and a worse credit card bill. That's exactly why the approval step matters more here, not less. Same pattern as the grocery cart, higher stakes, and it still holds: do the work, then hand over the keys.

Extra Credit

Google, OpenAI, and the race to build agentic commerce

Quick read for understanding:

• Why purchasing infrastructure using LLMs is so complicated

• Why Google is better positioned for agentic commerce (surprise its’s thanks to Froogle)

• The protocols that the new commerce infra is being built on

The Invention of the Shopping Cart

Fascinating read on the invention of the shopping cart. The funniest part is that at the beginning, men felt emasculated by them and insisted they were strong enough to carry their own baskets, and women didn’t want to push something that looked like a baby stroller.

The adoption of the cart only grew through advertising and hiring actors to pretend to be shoppers who pushed the carts around the store.

I think that we will slowly start to see more and more agents who actually do work for us. We just need to build the trust that they won’t mess everything up. I don’t know about you guys, but I have way too much on my plate and would love an agent to handle it all.

A question I’d love to hear from you on: What is a task that you would trust an agent to do for you?

— Amaraj (aka the guy that’ll take one of everything)

The Meme