The pattern across the six newsrooms is consistent. AI earned its place on tasks that were repetitive, structured and already well understood, including:
- Turning a weather bulletin into an article
- Drafting a meta title
- Cutting a long recording into clips
- Tracing where a photo first appeared
- Reading comments no human could.
Each solved a problem the team had named itself, and every tool had a human checking the output before it reached their respective audiences.
On build versus buy, the cohort mostly did neither. Building something new was judged too technical and too expensive at this scale, and off-the-shelf products missed local languages and formats. The working middle ground was configuring existing tools.
For example, some set up custom agents in plain language on consumer subscriptions, while others built on a free prototyping environment, deployed open-source automation, or added a button to an existing CMS.
Where a build needed skills the newsroom lacked, they hired help but learned the backend so the capability stayed in-house.
The common pitfalls were rarely technical. Prototypes over-scoped and stalled until the task was narrowed. End users joined too late. Working tools went quiet because no one owned them. Without a named owner and a routine check, even a successful build stops. And these were early builds, so the time spent training, prompting, testing and discarding was part of the cost.
The governance essentials run across every policy in this casebook. This includes:
- human review before anything is published
- Red lines on what AI must not touch
- Disclosure to the audience
- Sensitive data kept out of external tools.
For a small or mid-size newsroom, the starting point the cohort suggests is modest. Pick one specific task nobody wants, train first, configure rather than code, keep a human on every output, and name the person who will keep the tool alive.