The European Correspondent is a pan-European, independent newsroom that covers news stories from all over Europe. Covering the continent comes with its challenges for their newsroom, namely in trying to ensure that its journalism is useful and relevant for all of its readers.

“Why should someone from Estonia need to know what is happening in Portugal? How do you tell a story from Norway that is important for someone from Italy? We don’t have the resources in terms of editors to manage that many stories, so a lot of times, you have to do the work on your own,” said Philippe Kramer, Co-founder at The European Correspondent.

The problem: Ambitious, but resource-constrained

At the same time, they realized that a lot of mistakes that journalists make in their stories are repeated by their peers in their newsroom. While the journalists may often realise that they’ve made these mistakes, they may need a “subtle ping” to help them notice that they’ve repeated the error, he added. 

“Every so often we have a text. In the beginning of the text, we make a promise that we will explain xyz and then we will actually not explain maybe one aspect of xyz that we promised our readers we would at the start of the story. That can happen because the day is hectic or resources are scarce,” said Kramer.

While many newsrooms currently use AI to write texts, Kramer said their AI works more like an editor, leaving comments in a document and through that “increasing the number of conscious decisions they make”. This led to creating Vika, an AI based companion editor that suggests comments and edits to reporters’ drafts, similar to what editors would do.

Building the solution: AI to assist and contextualise editing

The team building Vika consisted of Kramer, the editor-in-chief, and a developer. The remaining newsroom took part occasionally in brainstorming sessions. 

Kramer and his team were convinced that the only way the tool would work for their organisation would be if it mimicked the existing writing environment. The entire editor works alongside the organisation’s CMS. Vika’s primary capability lies in its “analyse” function which provides editorial suggestions and commentary throughout the draft of the article. Kramer demoed the tool to us in one article with multiple suggestions from Vika. 

Some of the use cases are quite simple. “In a sentence where there is a verb that’s quite boring, for example, it will ask if you want to turn it into something else. In another paragraph which you may have maintained as a placeholder, but forgot to remove, it would suggest that you add text or remove it,” he said.

It works through a system of what they call “mini AI agents”, where different agents conduct different functions that can then be performed across the draft document. For example, one mini AI agent checks for style and tone, another one that checks if the article meets The European Correspondent’s audience needs model, and yet another that simply checks the title of articles. Journalists in the newsroom can also custom create their own mini agents using Vika to detect different things in their texts according to their requirements. 

“At the core of everything that we do is helping the journalists write the story and the editors who write it, make a greater number of conscious decisions of what they want to look at. Vika directs them to what they should potentially look at and then they can use their true expertise and let it shine,” explained Kramer.  

Despite the use of the term “AI agents”, Vika does not actually use agents at all. In fact, Kramer said it was a “very long, very structured” prompt that led to its creation. They mainly used Gemini for building Vika, however Kramer noted that the structure could easily be adopted to other AI models.

“When it came to AI we realised we didn’t need something very fancy or autonomous. The true innovation for us came in making it really convenient for this one use case that we have,” he added. 

The opportunities: Building a seamless system for their users

There was a great amount of testing that went into creating and structuring Vika, according to Kramer. They took inspiration from other text editors that already exist including for structure and logic. They didn’t have to reinvent the wheel and just had to learn from it, he added. 

The feedback for Vika has been positive, he explained. They first launched Vika without the AI and just as a writing environment so that the journalists could learn to use it intuitively. This also helped them tune out any bugs that were associated with it. Later, the AI elements were added.

“What has really been interesting is how the newsroom started building a passion for it. They started coming up with their own suggestions on what they would like to see in it, about what should be added. They appreciated how they didn’t have to perform crazy document management; that everything is built into the system,” said Kramer. 

Creating Vika also had its share of challenges. One of them was in learning to prioritise the most important features. “Our project was really ambitious. Knowing where to stop and what’s good enough so that it works - was challenging,” said Kramer. They were also worried that, despite the positive feedback, their team of journalists would not use it and would circumvent the system being built, he said.

One of the bigger challenges they face is in ensuring the relevance of Vika’s suggestions. “If you have ten suggestions, and eight of them are not relevant, then you start getting annoyed with the AI and will stop using the tool pretty quickly. So the AI needs to know when to stop suggesting and when it’s not relevant anymore,” he added.

To solve this, they built a two-layered prioritisation system within Vika. This system ensures that smaller editing issues like punctuation errors are not suggested right at the beginning. Instead, things like structural changes to text are suggested first. To help with accuracy, they also have an internal feedback system tied to the suggestions. Moreover, since the AI doesn’t rewrite text and only highlights potential issues for journalists to rewrite at their discretion, most bias problems that come with an AI model can be avoided, said Kramer.

“It’s not just human-in-the-loop, but the human is the person doing the task. That solves the problem of hallucinations which may not be captured in time,” he added.

Lessons for newsrooms

To Kramer and his team, design and limitations are some of the key factors to consider while using AI.

• Design for precision through limitation: Avoid focusing on the "magic element" of AI (like simply generating text). Instead, limiting how you apply the AI makes it more precise and therefore truly useful for newsroom tasks.

• Focus on stronger decision-making: “If you want to use AI in your newsrooms, it helps in being able to plan and demonstrate that using AI will enable your journalists to make stronger and more impactful decisions. This is a powerful lesson for us.”