Which media companies are making money with AI in real life, and how are they doing it?

Hi, I’m media innovation journalist Ulrike Langer and you’re reading my weekly AI in media newsletter. Thank you for being here! If someone has forwarded it to you please subscribe and never miss any future post.

Out of 1,047 newsroom AI initiatives tracked in AI For Newsroom, the most detailed public inventory of its kind, exactly 77 are about making money. Just over seven percent. And when I pulled all 77 records this week, I found a second number hiding inside the first: 70 percent of these revenue projects have no stated connection to the big generative AI systems that dominate every industry panel — no ChatGPT, no Gemini, no Claude. All summer, News Machines has examined how the publisher and AI business relationship should work in theory: exposure, licensing, payment, discovery. This newsletter is a reality check — who is earning money, and where, and with what. And because the database behind it is free and answers questions in plain English, the second half of this edition shows you how to interrogate it yourself — about monetization or anything else among its 1,047 entries.

Sergei Yakupov — founder of the media consultancy Novocean and instructor in the JournalismAI Strategy Lab — has been collecting newsroom AI implementations for years. His index, AI For Newsroom, holds 1,047 of them, reaching back roughly a decade to projects that predate the generative AI era. Last week he added something new: an MCP server, a standard interface that lets AI assistants query the database directly instead of a human clicking through a website.

On Monday I connected it to Claude and started asking questions. That night I sent follow-up questions to Yakupov in Portugal by LinkedIn DM. His answers — and a CSV export of all 77 monetization records — arrived before I finished my first cup of coffee Tuesday morning.

Entries reach the index by three routes: Either newsrooms submit them, Yakupov spots them in the RSS feeds and email newsletters his system subscribes to, or he runs research sweeps with Claude and Groq models. He cleans the results through scripts, then does a final pass by hand. His inclusion rules are strict: Initiatives must be scalable and tech-focused, "not a single usage." One-off experiments — "we used NotebookLM to chat with a dataset and wrote this article" — don't get in. So the number 77 is not missing casual dabbling but deliberately excluding it.

Full disclosure: News Machines is among the publications Yakupov tracks as sources for the index. None of the 77 monetization records derives from my own reporting. I checked every source field before writing this to avoid a self-referential post.

Geography first: Of the 77 money projects, 66 are based in Europe and North America. Asia has five entries, South America three, Oceania two, Africa one. Whatever the AI money layer of journalism is, it is so far a North Atlantic project.

The 77 records split into two very different kinds of money. About 30 are licensing and platform deals:

Thirty records sounds like a broad movement until you notice that almost all of them are participations in the seven programs listed above — the database counts each publisher separately. This is money from the machines: revenue that arrives because an AI company wrote a check, not because a newsroom operates anything.

The other 47 records are publishers using AI to earn, and nearly all of that work is prediction rather than generation.

The newest entries do the same jobs with better tools:

A small group of news publishers invent new products outright — Hearst's Houston Chronicle turned public county data into TX Tax, a property-tax protest tool sold to readers, and Dow Jones launched an AI-built French-language financial newswire.

Projects that make money by generating content — text, images, audio — barely appear. Which is what that 70 percent figure from the top measures. The database records which AI system each initiative runs on, and within the 77, that field names one of the big generative models only 23 times. The other 54 records either leave it empty or say "Custom": systems built in-house, overwhelmingly for prediction tasks like the ones above. Whether some of those custom builds use generative components internally, the field can't say. What it can say: The money side of newsroom AI is not being built on ChatGPT or Claude.

"I was surprised that so few examples are about monetization," Yakupov wrote me. His explanation: "It seems to me that we are trapped in this 'generative' nature of modern AI — it is very difficult to start thinking of it beyond 'create text,' 'generate image' tasks." The industry stares at the flashiest capability while, in his words, "the most underrated feature of AI is dealing with mess and chaos. And any datasets are mess and chaos. And AI in monetization should be built on data."

The numbers above back him up. The projects that earn are exactly the ones doing the unfashionable work: cleaning subscriber data, predicting churn, pricing access.

In a forecast the Reuters Institute published in its 2026 AI outlook, Yakupov predicted the industry's focus would shift from production toward distribution and monetization. I asked whether his own database shows that shift yet. His answer: not yet. The generation phase isn't over, he told me, though newsrooms have started looking past it. What holds the shift back, in his reading, is infrastructure and literacy: Implementing AI beyond a chat window still requires "some rocket science," and organizations without external funding rarely get further than the simplest use — generation.

If the shift comes, its first entries are already in the index. The newest monetization records are agentic:

The leading edge points exactly where his forecast does.

Everything in this issue came out of the database through plain-English questions, and you can do the same. The index itself is free at aifornewsroom.in. The MCP connection works like this:

Claude: This works on the free plan — free accounts get exactly one custom connector, and this only takes one. Go to Settings, then Connectors, then "Add custom connector." Name it, paste the server address https://aifornewsroom.in/mcp/server, click Add, and approve the sign-in. From then on, you ask questions and give directions in the chat: "Which monetization initiatives does the database track in my country?" — "Group all initiatives by region and type." — "Which projects name the AI model they run on?", etc.

ChatGPT: Custom connectors sit behind a beta feature OpenAI itself calls developer mode and labels as intended for developers. On individual paid plans (Plus and up), you enable it under Settings, then Connectors, then Advanced settings; after that you add the server address the same way. In Business and Enterprise workspaces, only an admin can switch developer mode on — members can't add custom connectors themselves. Free ChatGPT accounts can't use custom connectors at all.

Microsoft 365 Copilot: Relevant for everyone whose employer decides their tools — you cannot add it yourself. Custom MCP sources become available as "federated connectors," and only a Microsoft 365 administrator can set them up. The practical move: Send your IT admin the server address and forward this newsletter issue.

Important: Know what you're querying. AI for Newsroom is a curated index, not a census — the European overweight partly reflects where Yakupov's network is probably strongest. Individual queries return capped result sets, and the oldest entries carry year-only dates. For the full dataset, ask the curator Yakupov: He sent me the complete CSV within hours and told me he may soon build a data export for researchers directly into the server.

That last detail is my favorite thing about this story. A one-person database, a one-person research, questions and answers exchanged over 24 hours, and the tool will probably be improved because a journalist (me) wanted to check a whole data set. The money map of newsroom AI is thin, Northern, and largely non-generative — but the infrastructure for examining it in public is suddenly very good. Thanks, Sergei, for building this useful free tool.