The approach is anchored in trust, with editorial oversight and a “human in the loop” ensuring that AI-generated content remains accurate and unbiased. 

To support this strategy, Reuters has developed a suite of newsroom tools that enhance speed, consistency, and editorial impact:

• Fact Genie, an AI-assisted summarisation tool.
• LEON, an AI-powered headline assistant.
• AVISTA (Automated Video/Image Sourcing, Tagging and Archiving) leverages machine learning to help journalists quickly find, tag, and edit photos and videos.

Reuters’ Bangalore newsroom – now its largest globally – has emerged as a key hub for AI-driven journalism, leading experiments in fact-checking and real-time news delivery.

AI adoption has accelerated across the organisation, with new tools moving rapidly from concept to daily use. Yet, trust remains central.

“We’re firmly committed to having the human in the loop to stand by any AI-assisted work,” said Euan Rocha, Bangalore Bureau Chief, Thomson Reuters.

Rocha was speaking at WAN-IFRA’s recent AI Study Tour, and later at the Bangalore AI Forum, where he interviewed his Reuters colleagues Chris Peters (Speed Editor), Sri Prasad (Senior Manager, Product Management), and Sagar L (Manager, Applied Research, TR Labs) about the agency’s work with AI technologies. 

The edited conversation follows below:

Euan Rocha: Thomson Reuters Labs works on multiple AI ideas. How do you filter and prioritise them?

Sagar L: We receive a continuous stream of ideas from both our product partners and the AI researchers within Labs. 

First, we evaluate the impact-versus-feasibility trade-off – how scalable is the idea?

Second, we internally assess technology trends and readiness. AI is evolving rapidly, but not every breakthrough translates into a robust, newsroom-ready solution. 

Finally, the most important factor is addressing real user pain points. We ensure AI solutions solve problems rather than add complexity.

We test extensively, scaling successful solutions gradually based on user adoption.

Rocha: Chris, from the editorial side, could you start by explaining the role of the Speed teams, what they do, and the unique challenges they’ve historically faced in processing large volumes of information quickly?

Chris Peters: Reuters has 250-300 journalists across bureaus and Speed teams, publishing around 100,000 business news alerts monthly.

While structured data like earnings can be automated, events like CEO firings or layoffs require ultra-fast human reporting – measured in fractions of seconds. Each region has its own reporting style and sources, adding to the challenge.

Despite this, our goal remains the same: delivering accurate, fast, and contextual news, making AI a key focus for us.

Rocha: How has Fact Genie changed your workflow?

Chris: Our goal is to publish key alerts within 30 seconds of receiving a press release. Fact Genie scans entire documents in under five seconds, suggesting newsworthy alerts.

Journalists review, cross-check, and decide whether to publish. The first alert often goes out within six seconds, significantly reducing cognitive load.

Fact Genie also enhances editing – clients need both speed and clarity. Simplifying a central bank statement quickly is risky, but Fact Genie does it more effectively than any human, making it invaluable.

Rocha: Clearly, Fact Genie’s been a bit of a game changer for the Speed teams. Curious, how long did it take you all to work on it from the experimenting stage to the testing stage to production deployment?

Sri Prasad: We moved quickly from prototype to productionising Fact Genie, taking about four months from idea to rollout.

Journalists were central to the workflow, with Fact Genie built around their needs. We started with senior users in structured markets like the US and UK. Over three months, we ran agile-driven iterations, refining models and prompts based on continuous feedback from Chris and his team.

Co-locating Labs, editorial, product, and development teams accelerated production. A continuous feedback loop, user squads, and regular catch-ups ensured constant iteration, collaboration, and efficiency.

Rocha: At the Lab side, you’ll have to work with teams and individuals very often who may not know too much about technology and product development. How have you all been able to bridge that gap?

Sagar: As we integrated more GenAI tools, we realised it’s easier to teach a senior journalist to write effective prompts than to train a data scientist in journalism.

Building on this, we introduced pair prompting where journalists and data scientists collaborate on solutions. This mutual learning ensures journalists master AI tools while AI developers understand journalistic thinking, improving integration.

Rocha: Sagar, you also collaborated with our legal editors and legal products, correct? While the workflow differs slightly, it still involves managing content in a similar way, in a sense. Could you share your direct experience with that collaboration and how that worked?

Sagar: We’ve learned a lot working with veteran editors – nuances often emerge beyond initial AI prompting.

In one project, we built a tool to generate AI synopses, expecting time savings. While junior editors worked faster, senior editors took longer, as they analysed AI choices and reread the original text.

This insight reshaped our deployment strategy, tool offerings for senior editors, and how we presented AI outputs.

Rocha: That’s fascinating. Chris, on the editorial side, have you noticed any differences in how experienced staff on the Speed team use the tool compared to others? Has it helped newer team members pick things up more quickly?

Chris: When we rolled out Fact Genie, our most senior journalists could match its speed on certain tasks. 

Our junior journalists were able to work much faster and meet Reuters’ standards more easily. In many ways, it made the work more accessible to them.

Rocha: Following up on my question to Sagar earlier – journalists we typically hire aren’t well-versed in tech speak or product design, right? Have there been any challenges or learnings for your team? 

Chris: We’ve worked with tech teams, including Sri, for a long time. We don’t try to be experts, but a key lesson from past projects is the need for early involvement. For Fact Genie, editorial shaped some requirements from the start, and we actively tested and provided feedback. Without using the tool and offering input, it’s hard to guide the project.

New tools may not always fit perfectly, but resisting change leads to a product that adds no value in the long run. So, we ensured active participation, even if it caused minor workflow disruptions.

Rocha: And since you have quite a few early adopters, have you noticed a multiplier effect, with more journalists experimenting with tools or suggesting workflow improvements?

Chris: There’s plenty of basic training to get people started. In a team of 100, at least 10 to 15 might be drafting their own prompts – both to refine Fact Genie’s output and enhance other workflows. This has reduced time and improved efficiency in many other parts of our alerting workflow.

Soon, we’ll be using AI to tackle challenges we haven’t addressed before.

Rocha: That sounds promising. Sri, bringing a product to life comes with unique challenges, especially with the often unrealistic demands from the editorial side on UI, specs, and more. What have you learned from that process?

Sri: Real-world scenarios often differ from data science predictions. We use AI to filter out non-newsworthy content before sending it to a large language model and generate content out of it, preventing irrelevant releases from reaching users.

Handling 100,000 daily releases, filtering was a major challenge. Early models like GPT-3.5 Turbo took a minute to generate alerts, while GPT-4o mini does it in 10 seconds. To improve speed, we limited content sent to the model.

Quality was also key – ensuring consistent attribution in alerts. Chris and his team collaborated with tech, handling manual edits until these improvements were achieved – an accomplishment in itself.

Rocha: Chris, developing the tool is one thing, but integrating it into daily workflows is another challenge. How did we handle that, and how successful has the integration been?

Chris: Adoption has been smooth overall, but initially, AI-driven workflow changes understandably caused fear – some vocal, some through cynicism. We ensured communicating the importance of embracing change with those involved.

For Fact Genie, we selected 10 senior journalists covering speed news, focusing on the US market – our most critical. Testing it live was a major risk, as this was Reuters’ first large-scale AI integration. 

Once these users demonstrated safe, reliable results, we expanded usage. The tool has been highly successful, taking on a life of its own, with new ideas coming in that could push it far beyond just alerting.

Rocha: From the Lab side, what are the key considerations when moving from experimentation to product development and deployment?

Sagar: These models are not deterministic, meaning – you use the model once, you might get the same response or a different one. 

To manage this, we continuously monitor their accuracy, assess performance daily, and implement feedback loops to refine them based on real-world use.

We also track user interactions to avoid surprises and measure their tangible impact.

Rocha: Over the past year, what has been your biggest learning or takeaway?

Sri: Journalists must be at the centre of AI workflows. Tools should enhance their speed and clarity without disrupting trust in reporting.

Sagar: AI should prioritise user needs over automation. Non-deterministic models require continuous monitoring and adaptation to real-world usage.

Chris: We’ve made significant progress, but AI is just beginning to reshape journalism. Fact Genie is just a starting point. Embracing and guiding these changes is crucial to staying ahead.