Online shopping is becoming the next major test for AI agents. Instead of asking a person to search retail sites, compare products and complete checkout, tech companies want software to handle more of that work directly.
Perplexity has already moved first among major AI startups with a shopping agent for paying customers in the United States. The product is meant to browse retail websites, find items and press the checkout button for the user. But the early picture is more complicated than a simple handoff from shopper to bot.
Why AI shopping agents matter now
Millions of Americans open their laptops during the holiday season to buy gifts. That familiar flow is what AI companies are trying to compress: search, compare, select, pay and ship.
Perplexity is not alone in that ambition. OpenAI and Google are reportedly developing AI agents that can make purchases, including booking flights and hotels. Amazon already has Rufus, an AI chatbot in a place where millions of people search for products, making checkout assistance a logical next step for that kind of product experience.
The pitch is straightforward. If an AI shopping agent works well, it could help a user find a product or deal they might not locate through ordinary browsing. It could also reduce the time spent on repetitive tasks, such as looking for a cheap flight or choosing a birthday present.
That promise is still theoretical in many cases. The current tools are early, and the challenge is not just answering a shopping question. The agent has to operate across websites, handle payments, respond to stock changes and avoid buying the wrong item.
How Perplexity’s checkout works
Perplexity’s shopping feature includes an option called Buy with Pro. In TechCrunch’s test, a request to buy toothpaste produced choices from Walmart, Amazon and smaller websites. Some results sent the user to the retailer, while others offered the Perplexity-powered buying flow.
In one test, TechCrunch selected a tube of Crest from Walmart and checked out inside the Perplexity app. The bank statement showed payment to Perplexity’s agent rather than Walmart. Three hours later, Perplexity emailed that the agent could not complete the purchase because the item was sold out at Walmart.
A second attempt to buy another tube of Crest worked, but only after eight hours. That points to a basic weakness in the experience: the information surfaced by the agent may not match the retailer’s live inventory by the time the purchase is attempted.
The process is not the same as shopping through a marketplace where merchants upload and manage storefronts. Perplexity’s agent appears to gather product information from retail sites and later tries to execute the user’s instruction. Perplexity declined to comment on whether retailers like Walmart knew their products were appearing inside its app.
Payments create both safety and trust questions
Perplexity is working with Stripe on the payment side. Stripe is using single-use debit cards for Perplexity’s AI agent, based on a repurposed version of Stripe Issuing. The point is to let the agent spend only what it needs for a specific purchase, rather than giving it access to a full bank account.
That setup matters because AI agents can make mistakes. If an agent buys the wrong pair of socks, the damage is smaller than if it had open-ended access to a user’s money. In this model, the user gives Perplexity the exact amount for the item and instructs the agent to buy it with the user’s name and shipping address.
Google’s AI agent reportedly needs access to credit card information. That could make some consumers hesitate, although several companies already hold billing details and autofill forms during online shopping, including Google, Amazon, Apple and Shopify.
Privacy is another issue. Perplexity says human checkers are involved to help ensure the agent works accurately. Human oversight is common in AI, and companies such as Scale AI and Turing have built large businesses around that kind of work. But shopping adds sensitive context, including what a person buys and where it will be shipped.
Perplexity spokesperson Sara Platnick told TechCrunch that human oversight provides occasional support and helps transactions finish in a timely manner while avoiding issues such as buying the wrong product. Perplexity declined to answer questions about how often human oversight is needed, how involved humans are and whether checkers watch purchases happen in real time.
Retailers and advertisers may push back
If AI shopping agents become common, fewer people may visit online storefronts directly. That could affect retailers that rely on their own websites to promote additional products, encourage impulse purchases or shape the buying journey.
Advertisers could also lose useful shopper data if agents stand between customers and retail websites. That gives retailers and advertisers a reason to resist tools that reroute attention away from their own pages.
Some AI companies are trying to avoid direct dependence on retailer permission. Rabbit’s LAM Playground lets an agent navigate websites for a user through a computer in a data center. Anthropic’s computer use agent does something similar, but from the user’s personal computer.
These systems interact with websites through ordinary user interfaces: clicking, typing and moving through pages in a browser. That approach can make the agent behave more like a person using a site, rather than an outside service requesting access through a back end.
Rabbit CEO Jesse Lyu has said that AI agents are getting better than humans at solving CAPTCHA, the tests used to separate people from bots. If that continues, website owners will need stronger ways to prove personhood online.
The next phase of online shopping is still unsettled
AI shopping agents could eventually improve online buying, especially when users need to compare options across many sites. But Perplexity’s current shopping agent shows how much still has to work before the experience feels easier than buying directly from a retailer.
The open problems are practical and commercial. Agents need accurate product availability, reliable checkout, clear payment controls, transparent human oversight and a way to operate in a retail ecosystem that may not welcome them.
More AI shopping agents are expected in 2025 from companies including Perplexity, OpenAI and Google. The race is not only about who can build the smartest agent. It is also about who can make automated buying trustworthy enough for users and acceptable enough for the online retail system around them.