It learns your desk's voice.
Wire copy arrives from dozens of sources; this gathers everything written about the same event into one place and drafts the story from it. Every correction an editor makes is studied and folded into how the next piece is written.
What the desk sees before anybody is committed to a story — the score, and the four things that produced it. The figures are an example of the panel, not a measurement.
A desk subscribed to a dozen agencies receives the same event fifteen times before lunch — in four languages, with different spellings of the same name, under headlines that share no words at all. The subeditor's real job has quietly become triage: working out which forty of the four hundred items on the screen are the same story, and which of those is the one worth reading.
Article Studio does the triage. It groups by event rather than by source or by wording, so one entry holds everything anyone has filed on a subject regardless of which language it arrived in. From there it can draft, and the draft is checked against those same sources before an editor ever sees it — because the fastest way to destroy the value of automated writing is to make an editor fact-check every sentence of it.
15 things it does. Open any one for how it works.
Everything filed about the same story is gathered together, regardless of the agency, the language it came in, or the words each writer chose. On a desk taking a dozen wires, four hundred filings a day collapse to around forty events.
A Hebrew filing and an English one about the same incident end up in the same place. Two stories that look similar but are about different events stay apart — merging them is worse than missing the connection.
A developing story stays as one entry for days. New filings join the thread rather than starting a rival one, so the desk never has three half-versions of the same running story.
Every story carries a score, and what drove it is shown alongside — how fast it is travelling, how widely it is carried — so you can see whether something is genuinely big or merely noisy.
Showing the components matters more than the number. A story with high velocity and one source is a single agency repeating itself; the same velocity across nine sources is the world picking something up. You can argue with the components; a bare number you could only accept or ignore.
The score resolves into three plain tiers a desk can act on: several independent sources within two hours, something building more slowly, and something still showing signs of life. A prompt for the morning meeting, not a running order.
Set a threshold and everything above it is drafted as it arrives, so the desk opens to finished drafts rather than to a list of things to start. Below the threshold nothing is written until somebody asks.
Most desks set the threshold high at first — breaking only — and lower it as they come to trust the drafts. Everything produced this way still lands as a draft in the review queue.
Headline, social headline, search title, meta description, standfirst, slug, category and tags — drafted alongside the article and held to the character limits each one has, so nothing is truncated badly at publication.
These are the fields that get filled badly at eleven at night, and the ones that decide whether a piece is ever found. Nothing comes back truncated mid-word, and you can regenerate any single field without touching the article or the others.
The article is written from the gathered sources in Arabic, following the conventions your desk actually uses rather than a generic register — sentence length, how names are introduced, how you handle attribution.
Before you see it, the draft is checked line by line against the sources it was written from. Claims that cannot be traced are corrected or removed, and what remains questionable is flagged for you rather than smoothed over.
Every quotation in a draft is matched back to a filing that contains it. A quote that cannot be traced is flagged before publication, not after a complaint.
Invented quotations are the failure mode that ends careers, and they are the easiest thing for fluent automated writing to produce — a plausible sentence in the right register attributed to the right person. Checking them is not optional here.
If a story has only a headline and two lines behind it, you get a short factual brief — or nothing at all, with an explanation. It will not pad a hundred words of fact into six hundred words of article.
When an editor rewrites a paragraph, the difference between what was drafted and what was published is studied. Patterns become written rules that shape the next story.
The rules it derives are readable text, not a hidden setting — you can open the file, see what it thinks your desk wants, disagree with a line and delete it. A newsroom's style guide should be something people can argue about.
Who and what a story is about is pulled out as you write, so the piece is filed correctly and connects to the rest of your coverage without anybody tagging it by hand.
Suggested imagery comes from your own agency subscriptions and from openly licensed sources, arriving with the credit and licence attached rather than as a link somebody has to check.
Three visible states, with a record of each transition. Articles hold for approval by default; authorised users can enable automatic publishing for a destination.
A quiet list of the subject areas with nothing published against them. Not an instruction — just the thing that is easy to miss on a busy day.
Agencies and feeds arrive continuously and are gathered by event as they land. Nothing waits for a run to be triggered.
Open the list and see events, each showing how many sources are behind it and how fast it is moving. That is the morning read.
Choose an event and a draft is written from everything gathered under it, checked against those sources, and handed to you with any doubts marked.
Rewrite what you want. Ask for a specific section to be redone rather than the whole piece. Add the paragraph only your correspondent could write.
Move it to published and it goes to your site or your own publishing system. The record of who wrote, who edited and who released it stays attached.
What you changed is compared with what was drafted, and the patterns become rules. You do not configure this; you just edit normally.
Whoever opens the desk normally spends the first hour finding out what happened overnight — on a desk taking a dozen wires, four hundred filings of it. Instead they open a list of around forty events, sorted by how many independent sources carried each one, and know within ten minutes what the morning meeting is about.
A negotiation drags on with a filing every few hours. Instead of twenty separate entries that nobody can reconcile, the desk has one thread that has been growing since the first day — so the update written on the nineteenth day is written from the whole story, not from the last three filings.
In the first week editors rewrite most opening paragraphs. By the fourth, the pattern in those rewrites — you front the local angle, you never open on a number, you attribute in the first sentence — has become a written rule, and the drafts arrive that way. Not because anyone configured it, but because forty editors corrected the same thing forty times.
Open it, read what the system believes your desk wants, and edit it directly. Nothing about your house style is hidden inside a setting nobody can see.
Gathering and scoring run continuously. Drafting happens when somebody asks, or automatically above a newsworthiness threshold you set — and everything written that way still lands as a draft.
If only the intro is wrong, ask for the intro. The rest of the article, including your own edits, is left exactly as it is.
Add a feed, mute one that is flooding the desk, or weight one you trust more. The mix of what comes in is an editorial decision.
Where the material runs out, the article stops. A hundred words that are all true is the correct output for a story with a hundred words of fact behind it.
Every quote is traced back to a filing that contains it. Anything that cannot be traced is flagged to the editor before publication.
Similar-looking stories about genuinely different events stay separate. A missed connection is recoverable; a merged pair of unrelated stories is a correction.
By default an article reaches readers because an editor moved it there. A desk that wants a category to publish on its own can set that up deliberately — what does not exist is an accidental route from the wire to your readers.
Most automated writing arrives fixed: it produces the same register on the first day and the thousandth, and any improvement means somebody rewriting prompts. Here the improvement comes from the work your desk is already doing. Every time an editor rewrites an opening, tightens an attribution or cuts a sentence, that becomes a rule you can read and argue with. Your best subeditor's judgement stops leaving the building at the end of their shift.
No, and most desks do not, at least at first. The gathering and the newsworthiness scoring are useful on their own, and many newsrooms run for months using only those while their editors write as they always have. Drafting is there when a story does not warrant a reporter’s morning — and whether drafts publish automatically is a separate setting you control per category.
Yes — filings arrive in several languages and are grouped by what they are about rather than what language they are in, then written up in Arabic. A story broken by a German paper and confirmed by an English agency lands as one entry.
The rules it learns are plain readable text that anyone on the desk can open, and a rule that is wrong can be deleted in a sentence. It is a style guide that writes itself and that you retain the right to overrule.
Yes. Articles can be published through the platform's own store or pushed to your existing publishing system, so the newsroom changes how it writes without changing where it publishes.
The useful demonstration is the one that runs on your footage and your wire feeds. Tell us how your desk works and we will show you the parts that matter — and say so where we do not help.