AI platform You.com announced today an upgraded AI research agent, ARI, or Advanced Research and Insights agent. ARI scans 400-plus online sources simultaneously to produce, in minutes, research reports with verified citations and interactive graphs, charts, and visualizations. According to You.com, its research agent solves the three fundamental challenges for automated research assistants—scale, speed, and synthesis. ARI facilitates simultaneous analysis of public and private data sources, and its early deployments have demonstrated its potential impact in highly regulated industries where comprehensive source verification is critical.
Established by AI research scientists Richard Socher and Bryan McCann as an online search competitor to Google, You.com has raised $99 million from Marc Benioff’s Time Ventures, Salesforce Ventures, NVIDIA, SBVA, Georgian Ventures, Radical Ventures, Day One Ventures, Breyer Capital, Norwest Venture Partners, DuckDuckGo, and others. You.com has positioned itself as an agnostic AI assistant, providing access to multiple large language models, and currently aims to become the foundation AI agent layer for enterprises.
ARI is the first of several planned specialized agents developed by You.com for its agent ecosystem, where users have built more than 50,000 custom agents since the fall of 2024. ARI is now available to enterprises to preview and launch as part of the new You.com Labs, a testing ground for building new enterprise AI solutions. In developing ARI, You.com worked with its enterprise customers, including Germany’s leading health media publisher Wort & Bild Verlag, and global advisory firm APCO, adding their private data sources and testing its accuracy and speed in professional-grade research.
2025 is the year of AI agents and the hottest segment right now is AI research agents. Last December, Google launched Deep Research, “your personal AI research assistant,” and earlier this month, OpenAI introduced its own Deep Research, “an agent that uses reasoning to synthesize large amounts of online information and complete multi-step research tasks for you.”
But these brand-name efforts are just the tip of the research automation pile. There are several agents on the market that offer, for free or for a fee, assistance to humans who conduct research for a living and to aspiring students. Elicit, for example, scans 125 million academic papers from the Semantic Scholar corpus to find relevant studies, extract key information from the papers, summarize findings, and suggest alternate questions to uncover additional relevant articles.
Specialized AI research agents are another development in this market, focusing on domains where literature searches are time-consuming. Lex Machina and LawGeex, for example, review cases and contracts, in addition to automating several time-consuming legal work activities.
Given the excitement regarding the new automated legal eagles, lawyer and journalist Bob Ambrogi decided to test OpenAI Deep Research against the “legal research prizefighters” Lexis+AI, Westlaw Precision AI, and vLex’s Vincent AI. While these services primarily rely on their massive proprietary databases, Deep Research does not have access to paywalled legal research, conducting its research using only public sources. Given this handicap, Ambrogi was surprised to find out that Deep Research, unlike the prizefighters, paid particular attention to a relevant legal doctrine (the “death knell” doctrine) and its practical implication (the need to file an immediate appeal).
“One can only wonder how Deep Research would do if it could access a more robust legal research database. Would Deep Research be the death knell for commercial legal research?” asks Ambrogi.
I would venture to speculate that AI research agents will become very popular with students, journalists, and anyone who would like to reduce the time spent researching a specific topic online. However, they will not be the death knell for any research activity that requires careful and critical data curation.
Here’s a counter-example: When Ben Thompson of Stratechery asked Deep Research for a report on an industry he is familiar with, it “completely missed” a major entity, a private company with little public information on it. Thompson: “[Deep Research] worst results are often, paradoxically, for the most popular topics, precisely because those are the topics that are the most likely to be contaminated by slop.”
There you have it. Little Web discussion leads to an incomplete result, and lots of Web discussion (e.g., about the “death knell doctrine”) leads to a more complete result than what’s provided by proprietary research services not relying on the Web.
The Web, the repository for publicly available digital information, is a crowd-sourced database. Google got it first, developing the most successful online search tool by focusing on what people did online. The relevance of a Web page to your search was determined by its popularity, by how many other Web pages (i.e., people) linked to it. Ever-improving its tracking of what people do online, Google today tells you what other, slightly different, search queries people use to search for what you are trying to find on the Web. This handy list has now become the “reasoning” and the “chain of thought,” the steps by which research agents “think” through your query and conveniently reveal to you.
There is a man behind the curtain or, in the case of research agents, state-of-the-art search engines that know what people do online and can match their behavior (e.g., what words they use when they ask a specific question) with the specific words of your query.
The knowledge of crowd behavior on the Web powers AI research agents. Add generative AI, and you get nicely packaged and time-saving reports that require careful scrutiny to separate the wheat from the chaff.
The downside of reports based only on the Web is the crowd, the source of the bountiful data and information. Think about Web documents published this year—do they provide the latest, most updated information about a specific topic? Not necessarily if they are based on old data and information. Think about exaggerated claims by a company trying to sell a product or a service—it’s called marketing. How about misleading information provided by competitors or against competitors? Are “predictions” and “estimates” by consulting or market research firms relevant to your research activity? Are they based on facts, or represent sloppy research, or are hype-driven, or maybe simply delusional?
I asked ARI about distinguishing between facts and opinions on the Web, and he answered: “The future of AI in distinguishing between facts and opinions, particularly when opinions are presented as facts, lies in the integration of advanced technologies with human expertise. By combining the processing power and pattern recognition capabilities of AI with human judgment and contextual understanding, we can create more effective systems for maintaining the integrity of information in our increasingly complex digital landscape.”
Indeed. The future of AI research agents is bright as long as they focus their work on carefully curated data.
