Earlier this year, I wrote an article on agentic commerce (link), arguing that agentic commerce might not be as big a deal as some were saying. In the ensuing few months, agentic commerce and its related discussion waned. Few signs of adoptions were observed.

Since August, Instinct (a personal AI agent app developed by a startup) and Muse (a personal AI agent app developed by Meta) have attracted the attention of both consumers and investors. On the consumer end, as of today, Muse is the No.1 most downloaded app across both Apple and Android. On the investor end, investors sold consumer-related stocks, in fear of the “disintermediation” risk—consumers might ditch existing e-commerce and food delivery apps and order via personal AI agent apps going forward.

I have been using Muse for the past few weeks; I am also an Instinct user. The more I use them, the more I become convinced that the idea of “agentic commerce” is probably overhyped. People will use personal AI agent apps more and more, but probably not for agentic commerce.

Limited Room for Dramatic Improvement

This, I think, is the most fundamental argument. Little room is left for “agentic commerce” to make e-commerce dramatically better than what e-commerce is today.

From traditional offline experiences to today’s e-commerce, improvements were in fact massive. Before e-commerce, to buy something, consumers had to drive some 20 minutes to a nearby store, hoping the items they wanted were in stock at good prices; he or she spent some 10 minutes in the store navigating the shelves, spent another few minutes in the check-out line, then drove some 20 minutes back home. Lots of effort.

With e-commerce, you sit on your sofa, say no words, move your fingers, make a few clicks; you enjoy a wide selection of products at competitive prices. Similar things can be said for food delivery (which is a form of e-commerce) as well.

Consumers moved to e-commerce apps because the experiences are dramatically better: better selection at competitive prices; hours cut down to minutes. Consumers move if the experience is dramatically better. Taxi to Uber. In-store to Amazon. In-person pick-up to DoorDash.

When average consumer time spent on e-commerce apps is cut from hours to a few minutes, from this point onward, how much “time” can you dramatically save for consumers?

When many items consumers buy on Amazon are at similar prices as Walmart in-store prices, and delivery is often free, how much more “money” can you help consumers save in a dramatic way?

So, I reason, we probably have reached a state of e-commerce that there is limited room for dramatic improvement—without a foundational shift in how humans shop.

If there is little room for dramatic improvement, there is little chance for major disruption.

Significant Room to Regress

Last week, I used (or at least, tried to use) Muse to order food delivery.

I typed texts to Muse that I wanted Mexican food: a few sentences, like a mini essay. After a few minutes, Muse came back to me with a cart (1x Chips & Quac, 1x Burrito, 1x Quesadilla).

I then typed texts to Muse, asking what the proteins are. Muse said both Burrito and Quesadilla are chicken.

I typed to Muse, asking what other protein options were. I waited a few minutes. Muse came back to me with a long answer. I read its responses. In short, beef was another option but beef would cost $3 extra. Meanwhile, during the same turn, without asking me, Muse decided to save me money by downgrading my entire cart to vegetarian (veggie Burrito, veggie Quesadilla), so Muse claimed that it successfully saved $4 for me.

I really wanted to see what other meat options there were. I opened my DoorDash app, only to see the menu as follows: Vegetable +$1.00, Grilled Chicken, Honey Chipotle Chicken, Shredded Chicken, Steak, Ground Beef, Beef Barbacoa, all +$3.00 each.

Wait…

• There are more protein options than what Muse typed to me in text format.
• “+$3.00” does not mean beef costs “+$3.00” more than chicken. All protein is “+$3.00,” same price. Moving from chicken to beef is free. Unfortunately, Muse read “+$3.00” in a literal way, interpreting it as adding extra three dollars.
• Muse decided for me that I need to eat vegetarian, so that Muse can successfully build the lowest price cart for me. It is a typical “gaming” behavior that AI often exhibits.

Having double-checked with the original DoorDash app, I went back to Muse, typing a short essay, asking Muse to change it from vegetable to ground beef. A few minutes went by. Muse told me the cart is ready.

Hmmm… I thought, should I personalize the order? The “spicy salsa”, “jalapenos”, “Mexican corn” and 10 to 20 other options that I can choose from?

My stomach was hungry. I looked at my watch—30 minutes had passed since I began this food delivery order with Muse!

If I went on to personalize this order with Muse, another 30 minutes might not even be enough!

No way. I gave up. I typed to Muse, asking it to place the order.

Wait, did the order go through successfully? Looking at Muse’s text screen, I was not sure. I went to my email to check the confirmation email.

Wait, where is my driver? When will my order get delivered? There were no updates on Muse. I opened the original DoorDash app to check for updates.

Having spent a solid 40 minutes ordering this Mexican meal that otherwise would take me just a few minutes at most, I learned a few lessons:

• More Communication Costs: Verbalizing shopping desire into text is mental work. Reading text responses from AI is mental work. If you want more personalization, write longer essays to AI please! If you want more choices, get ready to read many essays from your AI agents! AI agents are adding communication costs into an otherwise easy e-commerce shopping experience. Do you want to write and read essays, or do you want just a few clicks?
• Fewer Choices: Text interface means the choices presented to consumers are limited. By “abstracting” away the e-commerce user interface, consumers lose choices.
• Much Slower: AI inference takes seconds and minutes to come back. Even if inference time is cut to a few seconds each turn, it is probably slower than me clicking through an e-commerce app that I am already familiar with.
• More Uncertainty: There are plenty of room for AI to make mistakes in shopping. That “+$3.00” was totally unexpected, and I would not have caught it if I did not check the original DoorDash app myself. Some mistakes are obvious; some mistakes are hidden. Your imagination is the ceiling. If consumers cannot be sure about a shopping experience, they will double check. I felt uncertain throughout the whole shopping journey. Are there other proteins to choose from? Are there other personalization options? Are there different sizes of my burrito? Are other merchants running interesting promotions? Do I have unused vouchers or credits to apply to the order? Does the order go through?

AI agents take away consumer choices, add extra mental work (write, read), take longer time to complete, open a whole world of possibilities for mistakes, make the ordering experience less certain, and ultimately make the entire shopping experience less fun.

Wait—and after regressing so much, your agentic commerce AI agent may be blocked from visiting your favorite merchants. Amazon already blocked Muse.

In a broader sense of “agentic commerce,” there might be valid use cases for personal AI agents. For example, tracking flight prices: you identify a specific itinerary, and AI agents check prices regularly for you. I already have multiple such tasks set up on Muse. Yet, even those use cases can often be done on sites such as Google Flights. So, I am just using Muse to double check results from Google Flights, which is kind of pointless, I think.

We are reinvesting the wheel. And the new wheel seems to be square.

Conclusion

In Chinese, “多快好省、省心、省力” describes an ideal consumer experience. Translate: “Greater selection, faster delivery, better quality, lower prices—with total peace of mind and zero hassle.”

AI agents, as I discussed above, risk going opposite to what consumers want:

• Adding communication costs (text, write, read);
• Taking away choices (limited options displayed via text);
• Making it slower (back-and-forth turns);
• Removing peace of mind (AI mistakes and human double-checking).

That is regression, not advancement.

Therefore, for agentic commerce, I make the following prediction: Agentic commerce does not have a supply problem. Technology will always get better and more companies will release their personal AI agent products. Agentic commerce will likely have a demand problem. For agentic commerce, most people may not want these AI agents.

Want something funny to take away? There is a rude slang in Chinese: “脱了裤子放屁.” Translate: “Like pulling your pants down to fart.” A pointless extra step, totally unnecessary. I think most “agentic commerce” use cases belong to this unfortunately funny category.

Agentic commerce, in its form today, misses the mark. It is probably overhyped.

(END)