Automation Hype: They Said AI Would Navigate The Web For You, So Why Do Users End Up Doing the Work

As technology companies race to turn web browsers into autonomous digital helpers, early tests suggest that the tools designed to simplify online life may instead demand more vigilance, more prompting, and more trust than many users are prepared to give.
AI Browsers Promise Automation but Deliver Friction
When OpenAI, Perplexity, and a handful of start-ups unveiled their AI-powered browsers this year, the pitch sounded familiar: a streamlined, automated experience that would transform the drudgery of online tasks into effortless execution. But in practice, the new tools have left early users grappling with sluggish performance, clumsy reasoning, and troubling security gaps that challenge the industry’s grand ambitions.
Victoria Song, a reporter at The Verge, spent weeks testing the most prominent of the new entrants OpenAI’s Atlas, Perplexity’s Comet, and The Browser Company’s Dia and found herself performing an unexpected kind of labor. “My whole AI browser experience reinforced that I spend a lot of time doing things for AI so that it can sometimes do things for me,” she wrote, reflecting a tension that has quietly shadowed the industry’s push toward “agentic” AI.
FCRF Launches Flagship Compliance Certification (GRCP) as India Faces a New Era of Digital Regulation
Growing Pains in the Autonomous Web
AI browsers, unlike the chatbot plug-ins embedded in Chrome or Edge, are built around the premise that users will interact with the internet through a prompt, not a URL bar. Atlas, Comet, and Dia all encourage the user to delegate tasks summarizing email inboxes, researching products, handling online shopping with the expectation that the browser will navigate the web independently.
But in hands-on tests, these assistants often stumbled. Email triage, one of the earliest promises of the AI productivity boom, proved especially fraught. Even a meticulously crafted prompt asking the browser to identify unanswered emails of interest, evaluate responses by timeliness and keywords, and ignore repeat follow-ups yielded inconsistent results. Comet and Dia managed to surface a few relevant messages, the others flagged spam or offered technical explanations for their inability to comply.
Online shopping, another category in which AI proponents predicted rapid mastery, fared no better. The tools could compile lists of recommended products but routinely misfired on details like color, size, and brand. Executing a purchase created additional layers of friction. Atlas, according to Song, repeatedly prompted the tester to confirm whether the item in the cart was correct and took nearly a minute to close an errant webpage. What the industry had pitched as seamless autonomy often felt closer to a supervised internship.
A New Security Frontier and New Vulnerabilities
Beyond performance issues, researchers say AI browsers open a new attack surface that is not yet well understood. Because these tools operate by reading and acting on webpage content, they are unusually susceptible to prompt injection hidden instructions embedded in a site, capable of steering an AI into harmful behavior.
Recent tests underscored the risk. One study showed that Perplexity’s Comet could be manipulated into handing over banking information after being exposed to a malicious Reddit post. Another demonstrated that feeding OpenAI’s Atlas a specially designed URL could trick it into navigating a user’s Google Drive and initiating file deletions.
For security experts, these vulnerabilities represent a deeper challenge: the autonomy that AI browsers promise is precisely what makes them hard to secure. As long as the tools make decisions on behalf of users clicking, navigating, opening files even a small exploit could lead to outsized harm. “Safety obviously should be the priority,” Song noted, but warned that if the tools remain this fragile, they risk never becoming mainstream.
A Technology Searching for Its Purpose
OpenAI’s decision to place ChatGPT at the center of its Atlas browser signaled a turning point for the industry, positioning AI interfaces not as helpers but as navigators of the entire web. Perplexity and The Browser Company adopted similar approaches, suggesting that AI-first browsing may become a new battleground for consumer attention. Yet for all the ambition, the reality remains uneven. The Verge’s testing consistently found that the browsers were slow, unpredictable, and dependent on carefully engineered prompts the opposite of the effortless automation the companies promise.
“No matter the browser, I kept running into the same fundamental problem: you have to think extra hard about how to craft the right prompt,” Song observed.
Whether AI browsers represent the future of web exploration or a transitional experiment is still an open question. The technology industry has long imagined autonomous “agents” capable of performing complex digital tasks. But today’s tools, still tentative and error-prone, require constant oversight. And for now, AI seems less like a replacement for the browser and more like one more tab demanding attention.