Brand journalists have spent the past year trying to figure out where AI belongs in their editorial workflow. Many of the available answers they’ve found have not been particularly useful. The more instructive examples are coming from traditional newsrooms, where reporters work under the same constraints brand publishers face: small teams, tight budgets, and a publishing cadence that doesn’t slow down for experimentation. Most of what they’re doing is small, costs very little — and can be replicated in an afternoon.
The marquee experiments we often hear about are the ones brand journalists tend to read and file under “not for us.” The New York Times, for example, has an eight-person AI team that doesn’t write articles. It builds internal tools like Cheat Sheet, a spreadsheet that allows reporters to upload documents, audio, and video, then simply click a button for the job they want done: transcribe this, summarize that, or find every mention of a name. Cheat Sheet grew out of a tax-break investigation, and now powers the Times’ reporting on everything from election interference to the manosphere.
When the Assad regime fell, Reuters used AI to process tens of thousands of security force documents that reporters in Syria had photographed. Data journalist Allison Martell built infrastructure that could translate, index, and search the records — work that would have taken a team of researchers months to wade through. The system helped expose a plan to move a mass grave.
Both projects are remarkable. Both are also easy to read and conclude that AI in editorial is the remit of dedicated teams, custom-built archives, and big budgets.
It isn’t. The more replicable work is being done by reporters and editors solving specific editorial problems with tools that cost almost nothing. We’ll first look at how newsrooms are putting AI to work; then, at what brand editors can take from it.
How newsrooms are using AI
After Oracle’s blowout September earnings call — in which the company disclosed multibillion-dollar cloud contracts that sent its share price soaring and briefly made Larry Ellison the world’s richest person — Wall Street Journal technology reporter Sebastian Herrera was assigned a profile of Ellison. Herrera, who had never covered the Oracle co-founder, only had a couple of days to file. He loaded old interviews, articles, and YouTube appearances into Google’s NotebookLM — a research tool that answers only from the sources you give it — and used it to map Ellison’s career and surface new angles. Decades of material became searchable in minutes. Herrera still did the reporting and the writing; NotebookLM just got him to the starting line faster.
At the AP, reporters had a problem: they were drowning in podcasts and couldn’t keep up. Senior AI product manager Ernest Kung built them a system that scrapes RSS feeds, transcribes new episodes, identifies keywords, and alerts reporters to relevant content. It got its first major test when Kash Patel was nominated FBI director in January 2025 — Patel had appeared on more than 100 podcasts, and the system surfaced what he had said in all of them. The reporting that came out of it became the first AP story to credit an AI product manager. The AP’s system itself is straightforward to replicate — RSS scrapers, transcription APIs, and keyword alerts are all off-the-shelf tools now.
At Fortune, the experiment is more aggressive — and controversial. The publication launched “Fortune Intelligence,” bringing back former editor Nick Lichtenberg to test ways to use AI in breaking news. In an internal memo, editor-in-chief Alyson Shontell wrote: “We intend to surf this wave, not get pummeled by it," Semafor reported. In just six months, Lichtenberg produced more than 600 articles — more than any colleague managed in a full year. AI-assisted stories accounted for nearly 20% of Fortune’s web traffic in the second half of 2025. Lichtenberg’s workflow is simple: he uploads a press release or analyst note into Perplexity or NotebookLM, prompts the tool to draft a story based on a headline, and edits the output.
The model has drawn criticism from industry observers who see it as a step toward content-farm territory. Six hundred articles in six months is a different kind of experiment from Herrera’s: it’s about volume, not depth. It is also the version of AI in editorial that brand publishers should be most cautious about copying.
Still, the Cleveland Plain Dealer took a related approach. The paper hired an “AI rewrite specialist” to run a modern version of an old newsroom staple: a reporter covers an event or interviews a source, then files raw notes, quotes, and key facts to the rewrite specialist. In The Plain Dealer’s case, the rewrite specialist feeds the material into a generative AI tool, which produces a draft. The specialist reviews it, the reporter checks it, and the story is published. The reporter never sits down to write.
The approach has also drawn criticism from industry observers. The paper’s editor, Chris Quinn, argues the results justify the method: by separating reporting from writing, the newsroom has freed up an extra workday per week for reporters to spend in the field.
At the Pew Research Center, the AI use case is simpler — and arguably more instructive. Previously, the audience team was spending 95% of its time writing formulaic social media posts and only 5% engaging with audiences. So they built a WordPress plugin that auto-drafts the posts. Researchers review a few options, pick one, and move on. The team now spends more time reading comments, answering questions, and turning reader feedback into content ideas.
In 2025, The Texas Tribune trained an experimental AI assistant on its archive of school voucher coverage. It embedded the chatbot inside an explainer, so parents trying to understand a complex new policy could ask specific questions and get answers drawn from the Tribune’s own reporting. What the bot couldn’t answer turned out to be more valuable than what it could: every gap was a potential story, and reporters were notified each time one came in.
What brand publications can take from newsroom experiments
None of the suggestions below requires an eight-person AI team. Several cost almost nothing. The common thread is that AI should be applied to a specific editorial pain point, rather than bolted on as an ad hoc productivity tool.
Use NotebookLM as a research engine for your own content. Any brand publication with an article archive, SME interviews, or a proprietary dataset can load the information into NotebookLM and query it. An editor preparing a feature can ask the tool to surface contradictions, identify gaps, and find quotes in minutes, compressing weeks of research into hours.
Turn your proprietary data into stories. This is where brands have an advantage over traditional newsrooms: Glassdoor has salary intelligence. Upwork has millions of freelance contracts. Zillow has housing figures. Toast has transaction-level data across thousands of restaurants. The same kinds of AI analysis Reuters used on documents from Syria — pattern recognition, anomaly detection, and automated categorization — could surface stories from any of these datasets. And the barrier to entry is lower than it looks: tools such as Claude and NotebookLM can analyze uploaded datasets and flag anomalies without custom infrastructure.
Automate what’s formulaic, then reinvest the time. The Pew model is the easiest to replicate. Most editorial teams face the same problem — they spend time on tasks that are necessary, but not editorial: social post drafts, meta descriptions, newsletter subject lines, alt text copy, and so on. The time saved can be invested in audience engagement: reading and responding to comments, fielding reader questions, and mining feedback for story ideas.
Turn podcasts into a story pipeline. The AP’s podcast system is more replicable than it looks — RSS scraping, transcription, and keyword alerts can be stitched together in an afternoon by a reasonably technical editor. Pick the 20 or 30 podcasts where your category’s experts, customers, and competitors appear. Quote-worthy statements, contrarian takes, and recurring questions are all raw material for story ideas.
Turn your archive into a chatbot — then see what it can’t answer. The Guardian built an internal chatbot that searches its own article archive. Any brand publication can do something similar without writing code. Export your published pieces as PDFs or text files, load them into a Claude Project or a custom GPT, and start asking questions: “Which sources have we quoted on employee retention?” or “What have we published about remote work in the past year?” That alone gives your team a searchable institutional memory, which can prevent duplicate coverage, help writers find internal experts, and surface topic gaps. Point the same tool at your readers, as the Texas Tribune did, and it starts doing a second job: every question it can’t answer is a story that hasn’t been written yet, assigned by the audience you’re trying to reach.
Where to draw the line
A University of Maryland study found that over 9% of newspaper articles now contain AI-generated text, often without disclosure. In March, The New York Times cut ties with a freelance critic who used an AI editing tool on a book review and failed to catch that it had pulled passages from a Guardian review of the same title. The Times called it “a serious violation” of its standards.
For brand publishers, the consequences of a lapse don’t stop at the masthead. A single AI-assisted mistake can travel from the publication to the company that funds it, the product it sells, and the customers it’s trying to reach.
AI’s role in the editorial process needs to be defined explicitly. The tasks it can touch and the ones it can’t should be written into guidelines, communicated to contributors, and revisited as the tools change. Experimentation is how you find the line. Guidelines are how you hold it.
Florent Daudens, the co-founder of Mizal AI, observed in a recent International News Media Association webinar that media is moving beyond random prompts to practical systems — modular instructions that guide models through specific tasks such as fact-checking or chart formatting — that are shareable across a team and refined over time. That kind of maturity takes iteration. But the starting point is simpler than it looks: one tool, an afternoon, and an honest audit of editorial time.
.svg)