Why Amazon Put Down Rufus, Its Loyal AI Assistant

Photo-Illustration: Intelligencer

After two short years, Rufus, Amazon’s “generative AI-powered conversational shopping experience,” has been sent to a server farm upstate. It will be remembered, almost, for appearing in the corner of hundreds of millions of Amazon users’ screens and for responding to questions like “What are good gifts for Valentine’s day” with answers like “flowers” and “chocolates” and links to buy some. It is survived by its replacement, Alexa for Shopping, a “personalized, agentic AI assistant on Amazon,” which is kind of the same thing, except chattier and with more “personalized knowledge and [the] context of Alexa+.” Rufus was preceded in death by countless other AI assistants who once asked, repeatedly, if we needed any help.

Of all the strange and insistent features that tech companies have come up with to get regular people to use AI that isn’t ChatGPT, AI assistants are by far the most cursed. The best-case scenario is that they’re easy to ignore. At their worst, they show up out of nowhere, little chatbots dangling off the side of familiar interfaces, pestering users until they get our attention, which they then waste. In contrast with stand-alone chatbots and AI-agent tools like Claude Code and Codex, assistants are forced into contexts — websites, apps, operating systems, software suites — that seem to confound them and us. Whether they arrive in 2011, like Siri, or in 2024, like Amazon’s Rufus, they’re better at showing us the limits of their underlying technology than they are at making the case for their potential.

Rufus was something in between, a premature attempt to match the fuzzy logic of a generative chatbot with the sprawling infrastructure of an enormous e-commerce system. It was, if nothing else, a cleaner way to interact with Amazon’s product listings, effectively a search page with fewer ads and less noise, not unlike Google’s AI Mode; mostly, though, it was a more awkward way to get where you needed to go on the platform, sacrificing the already diminished enjoyment of tapping around the shop without gaining much in the way of efficiency. At least you could still mess with it:

In the meantime, AI shopping really did start to grow but largely through much better chatbots that people were already using and just occasionally asking about products: ChatGPT, Gemini, and Claude. It’s a pattern that extends beyond shopping — nobody wants an AI widget. Users seem to prefer general-purpose chatbots, like ChatGPT and Claude, that are unaffiliated or at least feel independent from existing products and interfaces (contrast this with, say, Meta AI, which spent a couple years jostling for placement across the company’s platforms to unclear ends). Programmers have moved away from coding assistants, which were inserted awkwardly into developers’ workflows, toward applications that handle more if not most of the process, like Claude Code or Codex. Shortly before Amazon moved on from Rufus, Microsoft — the company with perhaps the most extensive history of unasked-for software assistants, starting with Bob and Clippy and running through Cortana — announced, after correctly assessing that its billions of customers found its numerous attempts to incorporate AI into Windows annoying at best, that it would now be trimming some AI overgrowth from its LLM-era general assistant, Copilot:

You will see us be more intentional about how and where Copilot integrates across Windows, focusing on experiences that are genuinely useful and well‑crafted. As part of this, we are reducing unnecessary Copilot entry points, starting with apps like Snipping Tool, Photos, Widgets and Notepad.

This is partly a story of capability. AI agents, which can perform multistep processes that involve outside tools, are more flexible and powerful than they used to be. At the low end, for example, chatbots can search the web to help formulate a response to a user’s question. This, it turns out, is a better answer to the question of “AI search” than bolting an LLM onto the side of a search engine. In 2026, the most successful applications of AI treat it as an interface rather than an add-on.

But Rufus’s short, unhappy life is also a story of being sort of trapped. For a company like Amazon — or Microsoft, Apple, or Meta, or even Google — the overwhelming temptation and institutional incentives will be to make a new technology subordinate to an old and dominant product. (This is the story of Siri but also Google Gemini on Android, a far more comprehensive product made by a company with its own AI models that is still reaching for usefulness.) Rufus was a fairly straightforward example of a why-not chatbot. Rolling it into Alexa, which has struggled for years, against both its limited capabilities and Amazon’s awkward ecosystem incentives, to become more than a glorified kitchen timer — even after a big LLM refresh — is a sign that this strategy doesn’t really work.

Alexa for Shopping, Rufus’s replacement, feels somewhat less like a chat widget. You can ask it for recommendations or product research but also to change delivery addresses, set price alerts, or pull together product comparisons (all things you could do or approximate before, of course, in various interfaces on Amazon’s apps and websites). It’s full of suggested prompts that look and function an awful lot like buttons. It’s far from a general-purpose chatbot and perhaps best understood as a way for Amazon to test out a broader redesign — a new, slightly chatty, lightly “agentic” interface for the same old product. It’s also a sign that Amazon feels a bit of pressure. To the extent people now want to shop with AI, it’s as part of an experience — throwing questions and prompts at a trusted chatbot — that treats Amazon like it treats everything else: as a peripheral source of data to be consumed, summarized, and made invisible.