Can Amazon AI voice replace customer reviews? It’s starting to try

Anyone who has ever gone in search of a product review on Amazon knows how valuable the experience of other shoppers can be, and how easy it is to fall down the rabbit hole of customer comments, from five-star raves to one-star takedowns — sometimes hundreds of words to get to the point. As Amazon continues to roll out AI features it says will make shopping easier, AI-generated audio descriptions of products are in the mix and scraping that customer commentary — maybe ultimately settling into a position to replace it as a go-to source of buying information. 

Called “Hear the Highlights,” the AI-voiced product descriptions rely on a large language model to script the summary based on a variety of sources, pulling from Amazon’s product catalog, customer reviews, and information from across the web, and then translating the content into short-form audio clips. The summaries began rolling out during the summer on select products to a subset of U.S. customers and have now reached all U.S. customers as a button in the mobile shopping app, scaling up to cover over one million products.

A unique appeal of Amazon, and other e-commerce options, has always been the ability to get information from actual users, not just product descriptions. Of course, bogus reviews have long been a problem for Amazon, even though it bans use of paid promotion and other inauthentic forms of review writing. You still have to sort through various efforts to game the system, and in recent years, signs that the newest review writers are chatbots like ChatGPT. But actual people sharing their actual, idiosyncratic experience with a product — customers as a source of information, and informed decision making — have been a key part of the learning process from clothing and shoe sizes to safety and more narrow questions: When purchasing a new toaster, does it do bagels, is the timer setting accurate, how easy is it to clean the crumb tray?

Can AI improve on that? Reviews are by nature unwieldy in the mass individuality, but human readers have proven pretty adept at distilling what they need from the human chaos. In some sense, the brain of the average Amazon review reader is a pretty good large language model, skimming and picking out key words and key information, so an AI will have to do it at least as well as a human would if it’s to add to the customer experience as promised. 

AI and cognitive overload

AI does have its advantages. For one, it won’t experience cognitive overload when facing an electric tea kettle on Amazon with over 32,000 reviews. It can comb through the data, but can it gives us only what we need or want to know? That can still be tricky for AI to get right, and it may be inherently problematic to mix product catalogs, customer reviews, and web information into a single distillation – with the sources of this information each coming with its own set of intentions.

“It’s important to recognize where AI is currently strong, such as in automation and pattern recognition, and where it still falls short, like in judgment-heavy tasks,” said Ankur Edkie, CEO & co-founder, Murf AI, which creates AI voiceovers. “A key question is whether there’s a way to factor in customer context as an input while generating these summaries,” he said.

The value of AI, according to Edkie, is finding the proper problem-capability fit. If that isn’t achieved, the sense of gimmickry is likely to sneak in through a door left open for AI fatigue, which he says consumers are likely experiencing by now.

Amazon’s Hear the Highlights AI-audio summaries are currently the same for every user.

“The AI summary needs to capture nuance and context. For example, even a few negative reviews on safety can outweigh many positives on other features,” Edkie said, adding that if a customer is focused solely on product performance, then summaries that emphasize price might not be relevant. 

The ability to factor in context, and to make the review process more of a dialogue between consumer and machine is likely where the technology is headed – in other words, toward the agentic side of AI, where Amazon is also actively adding to its AI commerce tools, such as Rufus, and a shopping tool called Interests AI, which prompts users to describe an interest “using your own words,” and then it generates a curated selection of products. That feature, rolled out in the spring, is separate from the main search bar on Amazon’s website. 

Chat and audio summaries will remain among the ways to engage, but having real-time conversations with an AI voice agent — asking specific questions, clarifying concerns, and getting deeper insights from reviews — is what will shift the experience from one-way delivery to two-way discovery, making it far more personalized, according to Edkie. “Currently, you can interact with Rufus through text or by using voice input, but Rufus cannot talk back, its responses are text-only,” he said. “With voice agents, however, you can have a two-way conversation with a bot that speaks to you,” he added. 

For now, one segment of customers likely to see immediate benefits is visually impaired shoppers, making accessibility an intriguing aspect of the feature, but the voice must be high-quality and deliver the content accurately.

Brian Numainville, principal at consumer research firm Feedback Group, says by providing an audio-based alternative to visually presented information, these types of features have the potential to make shopping more accessible, converting detailed text into easily consumed audio summaries. However, for it to truly benefit people with visual impairments, the feature must be thoughtfully designed. According to Numainville, this would include ensuring full compatibility with screen readers and keyboard navigation, providing clear, structured, and concise summaries, and avoiding overly long or confusing audio presentations. The quality and clarity of the AI-generated voice will also significantly impact usability. 

“The shift from diverse human reviews to AI-generated summaries might mean losing important nuances, context, and personal touches,” Numainville said.

Risk of losing unique shopper insights

The tendency of AI to focus on common themes can dilute responses even as it distills them.

Human reviews tend to include digressive stories and details around highly specific use cases — think of what motivates someone to write a review in the first place, and how this pairs well with the shopper’s anxieties and decision-making process — all of which can help to make the shopping experience more personal and insightful. 

“AI might overlook unique insights or niche needs that don’t align with the majority of responses,” Numainville said. “Additionally, the ability to critically interpret reviews — like spotting biases or trusting certain reviewers — is diminished with AI summaries.”  

Scraping a great deal of content to generate the summaries also sacrifices the annotation that can distinguish among product descriptions, customer reviews, and web information.

“It’s not 100% clear how much of the product description versus reviews is used in the current implementation — Amazon hasn’t fully detailed this mix,” said Numainville. 

Amazon declined to comment, referring CNBC to publicly available information on the feature.

“Product descriptions are typically marketing-driven and emphasize positives, whereas reviews reflect real user experiences and include both pros and cons,” Numainville said. Mixed without attribution — which by definition flattens information sources — it could be difficult for customers to discern ad-speak from genuine reviews, with the results be something like native advertising.

“This blending, if it occurred, could unintentionally mislead consumers by lending factual authority to subjective opinions or disguising promotional material as unbiased reviews,” Numainville said. 

Research does show that customers are inclined to trust voice as a format of information delivery, regardless of how balanced the information being delivered is.

“It seems likely to me Amazon’s desire to sell products would weigh more highly than incorporating all critical perspectives on a product,” said Tama Leaver, professor of internet studies at Curtin University in Australia.

Another concern is that “AI might weigh average and overall scores, while buyers often look at the few negative reviews – the one stars – even if there isn’t a lot of them,” Leaver said.

Dr. Nauman Dawalatabad, a research scientist at Zoom Communications, said in his personal view the technology is moving in the direction of better customer experience. “I take it as technology helping us to make informed decisions,” he said, citing the cognitive fatigue and wasted time that can come from reading through customer reviews.

If a streamlined description ends up leading to impulse purchases and regrets, that is on the buyer, he says, and no different than the way things have always operated — the same charges of subtle coercion could be made against all marketing efforts. He thinks that as voice-based agentic AI continues to evolve and consumers start talking (instead of typing and searching) with an AI agent and describe what they want, “it will get you exactly what you need.”