How Reuters Is Building AI Into a Newsroom of 2,600 Journalists

Hi, I’m media innovation journalist Ulrike Langer and you’re reading News Machines. This week, I’m in Chicago, attending the Online News Association Conference #ONA26. This is my 12th ONA conference, so I guess I count as a veteran. That comes with perks. I know how to get the most of this conference and I realize when sessions go deep enough for a case study. That’s what happened on Monday which why I chose Reuters for this issue.

But before we dive in, I have a request. You can support my newsletter in so many ways. Talk about it at the water cooler. Recommend it on LinkedIn. Post it in your Slack group. Become a paying subscriber. Or click on the ad. That’s like a tip jar, except that it only costs you a click. Thank you! I appreciate it. Ulrike

About a year ago, Andy Sullivan was staring at a thousand-page piece of legislation. Republicans were trying to pass the One Big Beautiful Bill through Congress, and Sullivan - Reuters' Acting Deputy Editor for Politics and Foreign Policy, a 25-year veteran of the wire - knew from experience that bills like this hide things. Somewhere in there were nuggets that would matter. Finding them meant reading the table of contents, making educated guesses, and hoping.

So instead, he built something. He fed the bill into an LLM prompt inside OpenArena, Reuters' internal AI platform, and started asking it questions. What's happening with electric vehicle subsidies? How does this affect rural hospitals? It worked. He kept going.

A year later, Sullivan runs 14 tools - monitoring bots, document analyzers, scheduling apps, story development aids - that serve dozens of Reuters colleagues. Not bad for someone who has a demanding day job. Sullivan is not a developer. He had tried to teach himself Python a decade ago, built one thing and then forgot it entirely. He is the journalist who keeps his engineering counterpart Paul J. Cifarelli up at night.

At the ONA Conference in Chicago, Sullivan presented alongside Cifarelli, VP for LSEG (the financial data group that owns Reuters' data division), and Editorial Tools, Jonathan Leff, Global Editor, Newsroom AI and Financial News Strategy, and Arlyn Gajilan, Global Editor, AI Development and Integration. Their session, "Making AI Real: What Leaders Need to Know Before Scaling," was less a presentation than a live diagnosis of what it actually takes to get AI working inside a major news organization. Reuters has 2,600 journalists in more than 100 bureaus. Here are five fundamental things they have learned.

Two years ago, Reuters launched OpenArena - an internal LLM environment where any journalist can build, test, save, and share prompts or basic chains using mainstream models. All data is protected under commercial agreements with AI providers. This year, 1,500 of Reuters' 2,600 journalists have used it, generating more than 600,000 requests. The tools that emerged include a custom German-language editor, a Brazilian fact-checker, and a Russian translation tool - all built by journalists, for journalists.

Without OpenArena, those tools would have been built on personal accounts, with unprotected data, invisible to anyone else in the organization. Sullivan's own tools currently live partly outside Reuters' official infrastructure, distributed through a personal website and a Gmail account that Reuters' spam filter routinely blocks. That problem - talented people building real things in a governance vacuum - is exactly what a new platform called Eden is being built to solve. Eden (Editorial Development Environment) will give journalist-built tools a governed home - compliance and security embedded from the start, not retrofitted after. It is still in development. 

In parallel with the grassroots work, Reuters has been integrating AI directly into Leon, its internal CMS - the tool every journalist opens every morning. Headline suggestions, bullet summaries, an error catcher, and a style guide prompt are already live inside the writing interface. The next phase, now in late-stage testing, is AI-assisted drafting of the first paragraph after an alert fires. Reuters publishes several thousand alerts a day globally.

The logic is blunt. "Building something that literally sits in the process that journalists already use," Leff said, "You're reaching a user where they are rather than expecting them to go craft something outside of it." A tool that requires a behavior change gets used by the 10 percent who seek novelty. A tool embedded in the existing workflow gets used by everyone. Sullivan's tools work the same way from the other direction - he built them around information flows his colleagues were already monitoring. The AI added a layer to something that already existed.

Sullivan's Federal Register Bot checks roughly 200 regulatory filings three times a day, filters them to the ones that matter across beats from telecom to healthcare to export controls, runs them through Claude for analysis, and delivers a digest at 8:47 every morning to about 25 to 30 journalists. "We've gotten a few scoops out of it," he said.

It was also the first tool he built, and the hardest. Getting the mechanics working was the easy part. Tuning the prompt so that it surfaced the right things, stopped ignoring what mattered, and stopped breaking every morning - that took months. His current estimate for a new prototype: a few hours. For a trustworthy tool: several months. 

Arlyn Gajilan's job - AI trust and governance - was supposed to be part of a broader role. It is now full-time. Reuters did not build its governance framework in response to a crisis. It built it in advance, which is why it is coherent rather than reactive.

The framework covers disclosure, accountability, and the stages process that takes a journalist's personal project from private experience to enterprise capability. The organizing principle runs through all of it: The journalist whose name is on the story owns the story, regardless of what tools produced it. "We didn't feel like we wanted to introduce disclosure that would seem somehow to change the accountability," Gajilan said.

Cifarelli's version of the same principle from the engineering side: “Trying to prevent journalists from doing things that will help them do their jobs better is a pointless endeavor." The governance system is not a gate. It is a path.

The newsroom currently has six different vibe-coded scheduling apps. Each was built by a journalist with a problem and a few hours. Each works for the 20 people who use it. None of them work for 2,600 people. Cifarelli's job, eventually, is to pick one and retire the other five.

This is not a Reuters-specific problem. It is a 2026 newsroom problem. The cost of building has dropped to near zero. The cost of maintaining six versions of the same thing has not. Before your newsroom hits the same wall, you need two things: a governed space for journalist-built tools to live, and a clear trigger for when a personal project needs a governance review. Neither requires a large engineering team. They require a decision.

At the session, I asked Leff directly whether Reuters was still free to experiment or whether pressure for hard numbers had arrived. Both, he said. The bottom-up work is valued and expanding. But the enterprise capabilities are tracked against specific outcomes: speed, error rates, coverage expansion. "We are not a research organization," Cifarelli said.

The goal Reuters has set for itself is not to produce more journalism. It is to use whatever efficiency AI creates to produce more original journalism - the scoop, the investigative piece, the photograph nobody else has. "Producing more of that makes us a better news organization," Leff said.