Q: How are publishers using AI in content transformation?
A: Content transformation covers a wider set of tasks than it might first appear, and AI is being applied to several of them at once.
One is atomisation: breaking an article, a report or an interview down into modular components (a headline, a lead, a data point, a self-contained paragraph) that can be stored, tagged and recombined independently of the original piece, rather than treated as one indivisible block of text.
A second is repurposing: taking a single piece of reporting and adapting it into the different shapes a newsroom actually needs, from a push notification to a newsletter blurb to a longer social post, without a journalist rewriting it from scratch each time.
A third, further upstream, is turning raw material, such as interview audio, press conferences, wire copy, into a structured, editable draft in the first place, so that transcription becomes a starting point for an article rather than an end product in itself.
None of this is conceptually new. Modular journalism, which means building articles from interchangeable blocks that can be reordered or swapped depending on what a given reader is interested in, has been discussed in publishing circles for the better part of a decade, and content repurposing has been a stated ambition of most digital newsrooms for almost as long.
What has changed is feasibility: manually atomising and repurposing content at scale was never realistic for most newsrooms, so it stayed a slide in a strategy deck rather than daily practice.
As AI models get better at understanding structure, context and editorial nuance, this stops being a theoretical nice-to-have and starts becoming an operational, strategic asset – one that determines how much of a newsroom’s existing output can be reused, personalised and distributed, rather than produced once and left to expire.
Q: What is a particularly illustrative use-case?
A: One instructive example of this shift is a project focused on turning interview audio into publishable article drafts.
The starting assumption was that transcription itself was the bottleneck; in practice, the more useful and harder problem turned out to be the step after transcription: structuring raw dialogue into a coherent, section-based article with a narrative arc, the way an experienced journalist would organise an hour of tape.
Building this required splitting the task into distinct phases rather than treating it as one generative step. A cleanup phase merges fragmented dialogue, resolves split sentences and corrects grammar and typos, with every change checked against the original recording rather than assumed.
Only then does an article-creation phase produce a structured draft, organised by topic, with suggested follow-up questions flagged where the interview left gaps for the journalist to close.
A final refinement and verification phase checks terminology and formatting and, importantly, traces every factual claim back to its place in the source audio, so factchecking is built into the pipeline rather than left as a manual step at the end.
The key learning was that the value of domain-specific AI here is not speed alone, but structure and traceability. A general-purpose model can produce fluent prose from a transcript quickly, but it tends to smooth over ambiguity and gives an editor no easy way to verify a sentence against what was actually said.
Separating the process into cleanup, drafting and verification made the output auditable, which mattered more to journalists piloting the tool than raw generation quality alone. It also clarified where the human role sits: not in typing up tape, but in the editorial judgement of what the interview means, which quotes carry the story, and how it should be framed – decisions no model, however well tuned, should be making alone.
Q: What are your three top tips?
- Optimise for structure, not just transcription. The real bottleneck in content transformation is rarely converting audio or raw material into text or other formats, but it is turning that material into something with an editorial shape: sections, a narrative arc, a lead. When evaluating AI tools, look past raw transcription accuracy and ask whether the output actually gets you closer to a usable draft, not just a cleaner recording.
- Build verification into the process, not around it. Any tool that transforms raw material into editorial copy should make it easy to trace a claim or quote back to its source. Treating factchecking as a built-in step, rather than a manual afterthought bolted on at the end, is what makes AI-assisted drafts trustworthy enough for journalists to actually use under deadline pressure.
- Treat model choice as an editorial decision, not only a technical one. General-purpose language models are a reasonable starting point, but domain-specific or fine-tuned models tend to hold editorial tone and nuance more reliably at scale, and are easier to constrain and audit. Keep journalists and editors involved in defining what “good” output looks like, alongside the engineers building the tool.
About us
Supporting publishers and media outlets to develop a sustainable business model through innovative technologies and services, Atex solutions allow today’s publishers to address and scale for tomorrow’s challenges. In the dynamic world of news, our goal is to innovate and create solutions that redefine how news is produced, shared, and monetised.
Email: info@atex.com
Web: www.atex.com
Sara and the other contributors to our AI Special will be taking part in an ‘AI Special – Q&A’ webinar on 24th November. Click here for more information and to register.
This article was included in the AI Special, published by InPublishing in October 2026. Click here to see the other articles in this special feature.
