A general-purpose chatbot makes for a poor colleague in a newsroom. It hasn’t read your archive, it doesn’t know your style guide, and it can’t parse the massive document dump sitting on your desk. The AI agent truly worth having isn’t a generic web assistant—it’s a tool built directly inside the newsroom, grounded in the publication’s own trusted records, and designed with a human journalist firmly in the loop.
This exact philosophy took center stage on September 25, when roughly 150 journalists, researchers, and technologists gathered on the 15th floor of The New York Times Building for a forum on AI and investigative reporting, co-hosted by the Times’ AI Initiatives team and Hacks/Hackers. As Andrew Deck noted in Nieman Lab, eight Pulitzer awardees this year disclosed using AI to organize, summarize, analyze, and translate massive document sets. Juliana Castro Varón, the Times’ senior design editor of AI initiatives, clearly defined the editorial boundary: the goal isn’t “robots telling the news.” Instead, the real power of modern AI lies in its ability to understand complex images and unstructured text at scale—far more than its ability to generate content.
The necessity of purpose-built tools became obvious earlier this year when the Department of Justice released over 3.5 million pages of the Epstein files, alongside tens of thousands of images and videos. Duy Nguyen, lead AI science researcher at the Times, pointed out that stacked printed copies of those PDFs would reach as high as the Empire State Building, taking a single reporter 11 years to read. In response, NPR developed an internal search tool to classify documents and surface name matches reporters wouldn’t have known to search for, while the Times built the Epstein Files Engine—a self-hosted chat interface layered directly over the document corpus. Over 100 journalists asked it more than 5,000 questions, directly contributing to at least 20 published investigative pieces.
Building on that success, the Times team expanded the concept to create News Agent, an internal “uber agent” designed to replace off-the-shelf tools like ChatGPT or Gemini. Powered by the paper’s archives, internal documentation, style guides, and reporting files, it gives staff a secure, tailored interface with a simple prompt: “Ask anything.”
Crucially, these specialized systems do not replace editorial rigor. Afrooz Mosallaei of the Center for News, Technology & Innovation emphasized during the forum that while AI excels at surfacing leads and patterns, it is far less capable of establishing publishable evidence. Verification, source development, and editorial judgment remain strictly human responsibilities. A tool can assist in organizing data or highlighting a potential lead, but it cannot prove a claim.
This is the design pattern worth replicating across the media industry, and realizing its value does not require a Times-sized engineering department. The path forward is simple: point AI agents at materials your newsroom already owns and trusts, preserve total transparency around primary sources and human sign-off, and leave off-the-shelf chatbots on the open web.
Our mission is to build the software that brings this exact architecture to every newsroom. By equipping journalists with secure AI assistants tailored to their own archives and reporting workflows, we can also help editorial teams tackle complex investigative projects without ever compromising on accuracy or trust.