Viestimedia: Transforming journalist's daily work with AI

Viestimedia is the leading media company in agriculture, forest industries, and related businesses in Finland. It publishes the newspaper Maaseudun Tulevaisuus and two subscription magazines, Koneviesti and Aarre. Despite having a relatively small staff, Maaseudun Tulevaisuus is Finland's second most-read daily newspaper, reaching 550,000 people weekly across print and digital platforms. Online, Viestimedia's websites, mt.fi and koneviesti.fi, attract a combined total of over 500,000 visitors each week.

Viestimedia aimed to update its operational model and integrate AI tools efficiently with a minimal budget. The goal was to compete effectively with the industry's largest players using AI solutions despite having significantly fewer resources.

The company wanted to improve internal journalistic processes, which included time-consuming daily tasks like transcribing interviews, writing and editing articles, and, in the final stage, proofreading them before publishing. Additionally, journalists in different outlets spent extensive time searching for pictures to be featured in their articles. They also devoted a large chunk of time searching for further reading ("suggested articles") related to the material they planned to publish. Another cumbersome task that needed resolution was including video and audio files in articles without involving the photo department each time.

Furthermore, many news outlets in general, including Viestimedia, struggled to find the most efficient way to distribute their articles to the public via social media. For a long time, the people who wrote the articles (who knew the subject best) and the social media team worked separately, which could lead to discrepancies between the article content and the social media captions.

The project kicked off exceptionally without a proof-of-concept phase, moving directly into tool development. The idea was to create a solution that is deeply integrated into journalists and other media staff's everyday operations rather than standing as "a nice-to-have AI tool".

A fundamental premise was journalistic ethics and principles; the company didn't want the tool to start producing long text segments and doing the work on behalf of the journalists. It is essential for readers to trust that the journalists are accountable and responsible for the text, sources, and data used in the articles.The tool's name 'Renki' itself, implies that the tool is a helper and that a human is always needed to finish the work.

We started by integrating an existing proofreading system directly into Renki. We then expanded Renki to utilize OpenAI's LLMs, feeding it as much context as possible from the article content and offering journalists prompts such as 'create social media copy,' 'create a headline,' and 'create lead-in text.' Creating an eye-catching, informative, and accurate headline wasn't enough. We wanted to forecast how well the article's headline would perform. To achieve this, we accessed Viestimedia's database, analyzing a year's worth of data that included 45,916 different headlines and the corresponding readership numbers. From this, we developed an algorithm that predicts the potential performance of a headline. Based on this, a journalist can decide how they want to revise the headline.

The goal here is that Renki helps 80% of the way, and then the remaining 20% must always be done by the journalists themselves. Nothing generated by AI goes directly to the reader without consideration, but it helps to think about alternative ways to write some of the texts.

We also created a feature to search for images to accompany articles. The company maintains a large database of over 700,000 images, half of which lacked proper descriptions. Among those with descriptions, many were poorly written, particularly from a search perspective. For example, a description might read 'horse race (location) (date),' yet the image shows someone eating a sausage at the event. We resolved this issue using OpenAI's Vision model. Now, Renki aids in finding better pictures; for instance, it informs journalists if an image has been previously used by the media company's other outlets.

Data approaches used:

• Large Language Models (LLM)
  • Chat
  • Image recognition using multimodal LLM
• Speech recognition
• Supervised learning
  • Vector embeddings
  • kNN search
  • Regression analysis
• Database queries
• Collaborative filtering

We aimed to introduce AI to journalists as easily as possible. Initially, the adoption seemed challenging; only a handful had prior experience with any form of AI technology – with just roughly 2% of the staff showing interest towards new technology. However, Renki's impact became undeniable, changing the core work of media – how journalists at Viestimedia created, edited, reviewed their work.

Renki was quickly integrated into daily operations, with approximately two-thirds of journalists utilizing it daily. The high acceptance rate reflects its effectiveness in improving how teams work. The tool allows journalists to work faster and more efficiently – whether they're generating ideas, selecting the best images, editing titles, or proofreading content. It has reduced the number of errors and improved the overall quality of articles. Additionally, the improved process for drafting social media posts has enabled journalists to collaborate more effectively with social media teams, leading to better audience engagement. Moreover, with the introduction of Renki, journalists can now independently add audiovisual content, eliminating the need for extra support.

Overall, the introduction of Renki has allowed Viestimedia's editorial team to save about 155 hours per week, leading to a roughly 20% improvement in annual operational efficiency.