Amazon is moving generative AI deeper into online shopping. Its current step is practical: AI-generated shopping guides for hundreds of product categories. Its longer-term goal is more ambitious: AI agents that can help customers choose, cart, and potentially purchase products with less manual effort.
The company is not presenting this as a finished future. Executives describe it as work in progress, with prototypes and unanswered questions still ahead. But the direction is clear: Amazon wants AI to become more than a search box or chatbot. It wants AI to act inside the shopping journey.
From Rufus to autonomous shopping help
Amazon added a chatbot called Rufus to its platform in February 2024. Rufus can answer questions about products across Amazon, and it is powered by a bespoke large language model also called Rufus.
That model was trained on large amounts of internet text, including publicly available websites, and then fine-tuned with Amazon’s proprietary data for commerce. Trishul Chilimbi, a VP and distinguished scientist at Amazon, says Amazon’s LLM has “hundreds of billions of parameters.” He also confirmed that Amazon is training a larger model, while declining to say how large it is or what new capabilities it may enable.
For now, Rufus is mainly a conversational assistant. The next stage could be more proactive. Chilimbi says the first move toward AI agents will likely involve chatbots that suggest products based on a customer’s habits, interests, and broader trends.
“If it’s no good and annoying, then you’ll tune it out,” he says. “But if it comes up with surprising things that are interesting, you’ll use it more.”
That line captures the central challenge. A shopping agent has to be useful without becoming intrusive. It must know enough to make a relevant suggestion, but not so much that the experience feels pushy or out of the customer’s control.
What Amazon wants agents to do
AI agents are different from ordinary chatbots because they are designed to carry out tasks. In the broader technology world, companies are exploring agents that can write code on the fly, enter text, move a computer cursor, navigate websites, or complete routine digital work for users.
Shopping is a natural test case for that idea because the task has a clear commercial end point. A customer wants to find something, compare options, decide whether it fits their needs, and buy it. Amazon’s vision is to reduce friction across that path.
Rajiv Mehta, a vice president at Amazon who works on conversational AI shopping, describes a possible Rufus agent that notices when the next book in a series someone is reading becomes available. It could recommend the book, add it to the cart, or even buy it.
“It could say, ‘We have one bought for you. We can ship it today, and it will arrive tomorrow morning at your door. Would you like that?’” Mehta says.
Amazon executives also describe a broader scenario. A customer might say, “I’m going on a camping trip, buy me everything I need.” In that case, an agent could assemble a set of relevant products. Chilimbi also raises a more extreme possibility: agents that determine when a customer needs something and then buy and ship it. “You could maybe give it a budget,” he says with a grin.
Mehta says Amazon is also thinking about how advertising can be incorporated into the model’s recommendation. That detail matters because a shopping agent is not just a convenience tool. It could become a powerful new layer between customers and products.
AI guides are the first visible step
Amazon’s AI-generated shopping guides were announced at its Reinvent conference in Nashville and are initially available on the company’s US mobile website and app. They cover hundreds of product categories and use the Rufus LLM to generate information that could otherwise take shoppers hours to collect online.
Brett Canfield, a senior product manager on the personalization team at Amazon, says unfamiliar categories can be time-consuming because shoppers must understand the market, available features, and product choices. The guides are meant to compress that research into a more direct buying experience.
Canfield showed WIRED shopping guides for televisions and earbuds. Those examples included technical features, explanations of key terms, and product recommendations. The underlying model can draw on product information, customer questions, reviews, feedback, and users’ buying habits.
“This is really only possible with generative AI,” Canfield says.
The guides also show why generative AI is attractive for ecommerce. A human editorial team might not produce full guides for every narrow category. Amazon’s system can generate guides even for niche areas, including “The definitive hedge trimmers.”
The hard part is reliability
The agent vision is compelling because it promises to automate repeated, everyday tasks. Ruslan Salakhutdinov, a computer scientist at Carnegie Mellon University who is working on AI agents, says, “Every major company is now doing [AI] agents.” He also says, “On the ecommerce side, if agents can find the best possible outcome for me, that's amazing.”
But making agents dependable is difficult. Tasks that seem simple to a person often require an AI system to understand visual information, evaluate many possible options, and choose the correct next action. Salakhutdinov and colleagues at CMU developed a dummy ecommerce website as part of Visual Web Arena, a platform for testing AI agents.
According to the source, key challenges include helping agents interpret visual information and training them to explore large sets of possible choices while narrowing in on the right one. That may require more advanced reasoning abilities.
Data is another major factor. Salakhutdinov says that information about how users handle common tasks such as shopping may help agents stay focused. “Data is going to be very important,” he says.
Search, publishers, and ecommerce economics
Amazon’s push also points to a bigger shift in search and shopping. AI-generated search results increasingly provide product comparisons and opinions directly. That can reduce traffic to outlets that make money by producing shopping guides, reviews, and related articles.
The tension is that AI systems may generate those answers using information drawn from the web, including material from publishers. Canfield declines to say what additional training data was used to build Amazon’s new AI shopping guide feature. The source also notes that WIRED’s parent company, Condé Nast, entered into a partnership with OpenAI in August of this year.
Still, interest in AI for ecommerce appears unlikely to slow. Machine learning is already widely used in ecommerce for analytics, search, and product recommendation. With large language models creating new use cases, one analyst report suggests the market for AI in ecommerce will grow from $6.6 billion in value in 2023 to $22.6 billion by 2032.
Mark Chrystal, CEO of Profitmind, says, “LLM agents are a customer service game changer.” He also says large companies such as Amazon may benefit most because they have extensive data to feed into their models.
The immediate product is a guide. The larger ambition is an assistant that shops with increasing independence. For Amazon, the prize is a smoother path from product discovery to purchase. For shoppers, the question is how much control they want to hand to an AI system built inside the store itself.