Publishers can feel AI in the building and still fail to find it in the accounts. That was already the finding of WAN-IFRA’s sixth AI report, Where publishers are seeing AI’s real value – so far, based on a second-quarter 2025 survey of more than 100 media leaders. Six months on, it is still the right question to ask before buying another tool.
The gains were operational. Seventy-five percent reported efficiency improvements. Sixty-four percent said content production got better. Fifty-five percent published faster. Forty-four percent allocated people more usefully. Only 9 percent could point to direct revenue.
That split is easy to sneer at, and it should not be. A desk that files sooner, with fewer people stuck on transcription, has recovered hours. Hours are the raw material of reporting, editing, and the audience work that subscriptions actually depend on. The mistake is stopping at the stopwatch. If the hours are poured back into more undifferentiated copy, the newsroom has become a faster version of the same business, and the 9 percent figure will not move.
The report’s own emphasis was that AI’s clearest impact was behind the scenes. Audience-facing value was harder to show. That matches what journalists told the Reuters Institute later in 2025: weekly personal use, limited organisational integration, thin training. A newsroom can be busy with models and still have no product a reader would pay extra to keep.
So the test for any AI project this spring is blunt. Name the reader behaviour you expect to change, or name the reporting you will do with the time you get back. “We published faster” is a means. It is not an outcome. Faster production of stories nobody finishes does not fix news avoidance, and it does not fix a weak subscription.
Measure both clocks. One is inside the newsroom. The other is whether a reader stayed, returned, or paid.