How does a very small Norwegian newsroom use agentic AI to help its reporters? With great clarity of thought — and great impact.

At my colleague Jodie Hopperton’s brilliant master class on agentic AI, I was delighted to watch a presentation by Rune Ytreberg, the data editor at iTromsø’s data lab, which uses AI to enhance journalism and develop new tools and services for 70 local newspapers in the Polaris Media group.

The 22-person newsroom recently undertook a hard-hitting investigation into the housing crisis in Tromsø, Ytreberg said. For the story, they analysed Airbnb data going back a decade, as well as numbers on tourist growth. 

“It was a lot of data. The challenge when you have so much data is the manual limits of the available journalists in the newsroom. So, the solution was to have a workflow redesign where we had to use agents to do a lot of the work because we wanted to do an advanced data journalism analysis using econometric models.

“We wanted to separate correlation from causation. We wanted to track market structure at scale. That’s hard to do when you don’t have enough journalists. That’s where agents come in.”

iTromsø put in place a layered agent research system with editorial supervision — where human experts oversaw and approved all the output from the agents and gave them tasks and subtasks — on research and analysis and looking for stories in the data.

Each layer of agents had a different set of tasks: Evidence infrastructure agents looked through all the data. Data collection agents went through the data, cleaned it, and assessed whether the data is “solid and valid.” Then, a database architect agent looked for data alignment for reproducibility. Then, a cleaning validation agent that paved the way for a journalist to determine whether the data was good enough to work on. 

After that, analytical agents parsed the data, and story intelligence agents looked for trends, system pressure points, and anomalies. Reporters and editors then examined the meaning of these.

The rationale was simple: “We didn’t only want to do anecdotal journalism. We wanted to do real investigative journalism and try to find and review the system complexity. That meant we had to go from traditional investigative data-driven journalism to redesigning our newsroom using agents,” Ytreberg  said.

Here are some of their key learnings:

Editorial governance is important

Humans with specific newsroom roles are monitoring and managing the agent systems. The data editor approves the models by validating statistical claims and ensuring the integrity of output. The news editor assesses the relevance and the social impact. The front editor examines the news and headlines and ensures clarity.

Agents can propose tools and queries, but it is the humans who make decisions. The system surfaces a proposed action “and we can ask that it explain this in simple terms, like a colleague would.”

You still need reporters to do what humans do wellAfter reviewing the data analysis, journalists went out and spoke to people, interviewing and photographing Airbnb hosts, families, politicians, and regulators. 

Why agents? 

“The sources we need are fragmented. They are inconsistent and very time sensitive, such as listing registries, scrape pages, and data sets that don’t share the same definitions. So these agents help us turn those sources into something the newsroom can actually use — in this case, an SQL database that anyone here can query and share and reproduce in their work,” iTromsø data analyst Gisle Johansen told the master class.

It is a cross-departmental team effort

Editorial staff steer much of this work, deciding which questions matter, which categories and definitions should be used, and what counts as interesting, while technical staff handle the technical pipeline, cleaning the data, and maintaining the database, and tailoring this to the local editorial context.

How did such a small team get so far so quickly? 

“I think we had to. We are so small and we are competing with a larger newspaper in our city. They have more journalists, more money, and more readers than us. So the way we can compete is to go for data that we can provide as exclusive news to our readers,” Ytreberg said.

Subscriptions have risen 5% every year for the past five or six years, he said: “So, it’s working.”

If you’d like to subscribe to my bi-weekly newsletter, INMA members can do so here.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .