What problem does it solve?
Regional and local newsrooms receive far more material than reporters can turn into stories: press releases, council notices, community announcements, event listings and stories supplied by partner services. The BBC describes the same bottleneck for stories from the Local Democracy Reporting Service, which arrive at a rate that means not all can be rewritten into BBC style and published. Newsquest says this routine content is necessary but does not call for investigative skills or human insight, and that doing it quickly frees traditional journalists for more complex stories.
Readers add a second demand. The BBC says bite sized content is increasingly favoured by news readers, especially younger audiences, and it is piloting short summaries at the top of longer stories. Writing a summary for every article by hand is one more task for an editor.
How does it work?
- Gather verified material. The reporter collects the facts: the press release, the council document, quotes and figures. At Newsquest the reporter places the reliable and checked information in a notes field, which limits the draft to facts the journalist has checked, though the reviewer must still catch anything the model adds.
- Generate a draft or a summary. For a draft, the AI turns the notes into an article in the house style. The BBC's Style Assist writes new drafts in BBC style from supplied stories. For a summary, the AI reads the finished article and returns a short list of key points from one approved prompt.
- Review and edit. A journalist checks every claim against the source, fixes errors and tone. Gannett's summaries carry a disclaimer that they were reviewed by a journalist before publication.
- Publish with disclosure. The page states that AI was used for the summary or the draft and which part of the page it covers. The BBC says it will disclose where AI has been used.
- Monitor after publication. Editors read a sample of published items, collect corrections and track reader behavior on summaries.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Interactions handled | Not pooled | at least 10,000 | 1 | 1 organization |
Value drivers: Employee productivity, Speed and cycle time, Lower cost to serve.
Indicative value
A regional news publisher that drafts routine stories with AI
USD 12,000 to USD 200,000
Journalist time freed, valued at cost per year
How this is calculated
Formula: routineStories * minutesSaved / 60 * hourlyCost. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Routine stories drafted with AI per year routineStories, stories per year | 3,000 | 10,000 | Editorial assumption. Newsquest reports over 10,000 articles published with AI aid across its titles by June 2024. |
| Net minutes saved per story after journalist review minutesSaved, minutes per story | 8 | 20 | Editorial assumption. No publisher on this page has published a time saving per story; measure it in a pilot. |
| Loaded cost of a journalist hour hourlyCost, USD per hour | 30 | 60 | Editorial assumption, replace with your own fully loaded cost. |
What it leaves out: Gross value of time freed only. It leaves out the cost of the tools, the extra editing when drafts are weak, corrections after publication, and whether the freed hours are used for reporting that earns more than the routine stories they replace.
Who already uses it?
4 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Newsquest
United Kingdom · Media and entertainment · 2024
Newsquest, a UK regional publisher, equips AI assisted reporters with its own tool, News Creator. The reporter places verified information in a notes field, the AI turns it into a draft article, and a journalist reviews the draft for accuracy and quality before it is published. The content covers press release based stories, community updates and event listings. Jody Doherty-Cove, head of editorial AI, wrote in June 2024 that Newsquest had published over 10,000 articles this way. The same article describes a separate bot that finds the right freedom of information email address and drafts requests to authorities.
- Interactions handled: at least 10,000, cumulative to June 2024, articles published with the aid of AI
"We’ve now published over 10,000 articles with the aid of AI, reaching millions across the UK."
Claimed by: organization
BBC
United Kingdom · Media and entertainment · 2025
In June 2025 the BBC announced public trials of two generative AI tools in BBC News. "At a glance" produces a bullet point summary of a longer story from a single approved prompt, and the summary is reviewed and edited before and after publication. "BBC Style Assist" takes a story submitted through the Local Democracy Reporting Service and writes a new draft in BBC house style, which a journalist reviews before it is published online. The BBC says it will disclose where AI has been used and will assess whether to release the tools more widely.
No outcome disclosed.
Gannett
United States · Media and entertainment · 2024
In May 2024 Gannett began adding AI generated "Key Points", a short list of bullets, at the top of stories on USA TODAY, according to an internal memo reported by The Verge and MediaPost. The articles carry a disclaimer that the Key Points were created with the assistance of AI and reviewed by a journalist before publication, and that no other part of the article was generated with AI. MediaPost notes that Gannett had stopped an earlier AI experiment, AI written high school sports recaps, in August 2023 after the output was ridiculed.
No outcome disclosed.
Aftonbladet
Sweden · Media and entertainment · 2023
Aftonbladet, a Swedish newspaper, built AI generated article summaries into its content management system. They appear below the article lead as expandable text. The Online News Association reports a click through rate of 43% on the expandable summaries (53% among readers aged 19 to 36), says elsewhere in the same case study that 40% of readers who saw the summaries expanded them, and reports that readers who used them spent more time on the full article. The ONA also reports that AI generated headlines did not perform as well as hoped and that tools inside the CMS worked better than tools that needed a second platform.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A house style guide and a set of approved published stories as examples
- Verified input for every draft: the source document, quotes and figures
- An editorial policy for where AI may be used and how it is disclosed
- A log of corrections, linked to the tool that produced the text
Systems to integrate
- Content management system (draft creation, summary field, publish workflow)
- Wire or partner story feeds, for example local democracy reporting
- Analytics for reader behavior on summaries
- Corrections log and editorial audit trail
Complexity: Medium
The build needs a content management system that can host the tool inside the editing screen, an approved prompt or house style set per content type, a clear disclosure label, and an editorial policy that says which stories may use AI at all. Aftonbladet's account, as reported by the Online News Association, is that tools built into the CMS worked better than tools that required a second platform.
- 1
Pick content types where the facts are in the input
Start with stories whose facts are all in a supplied document: event listings, community updates, press release rewrites. Leave anything that needs judgment about what is true to reporters.
- 2
Put the tool inside the CMS
Add a button in the editing screen that fills a draft or a summary field. Aftonbladet's experience was that tools which needed a switch to another platform were far less successful.
- 3
Fix the prompt and the review rule
Use one approved prompt per content type, as the BBC does for its summaries, and write down what the reviewing journalist must check: names, numbers, quotes, dates and attribution.
- 4
Pilot with a named group and measure
Run the tool with a small group for a few weeks. Count drafts accepted, edits per draft, corrections after publication and minutes per story against a comparable group.
- 5
Disclose and widen
Publish a plain statement of where AI is used, add the label on each page, then widen to more desks only when corrections stay at or below the baseline.
Guardrails
- AI input limited to material the journalist has verified and entered
- A journalist reviews and approves every draft and every summary before publication
- A visible label on pages with AI assisted text or summaries
- No AI drafting of stories that need investigation, sensitive reporting or legal review
- Corrections traced back to the tool and reviewed monthly
KPIs to instrument
- Share of AI drafts and summaries published after edits, and edits per item
- Corrections per thousand AI assisted items against the baseline
- Minutes from input to published story for AI assisted stories
- Reader behavior on summaries: share who expand them and time on the full article
Human in the loop
The journalist owns the text: they check each claim, edit or discard the output and are accountable for what is published. Editors decide which content types may use the tool and review the corrections log. Readers can report errors in a summary through the usual corrections route.
Common failure modes
- A fluent error is published
- A summary states a date or a figure the article does not contain. Review against the article, not against memory, and keep the rule that no summary goes out unreviewed.
- Review becomes a glance
- Reviewers see hundreds of items and approve without checking. Sample reviews by a second editor and track the correction rate per reviewer group.
- Reporter resistance
- Reporters distrust a tool they did not choose. The Online News Association reports initial resistance to summaries at Aftonbladet; include reporters in the design.
- Style drift
- Drafts converge on one flat voice. Refresh the style examples and let editors tag drafts that read as generic.
What are the risks and rules?
EU AI Act
Limited risk (transparency)
Article 50(4) requires deployers to disclose text generated by AI that is published to inform the public on matters of public interest, unless the text has undergone human review or editorial control and a person holds editorial responsibility for it. A newsroom with a journalist who reviews and is responsible for every item meets the exception, but a label is still good practice. Newsroom drafting is not an Annex III use.
Guidance
- Article 50, transparency obligations for providers and deployers of certain AI systems (European Union, Europe). Sets the disclosure duty for AI generated text on matters of public interest and the exception for human review under editorial responsibility.
- Paris Charter on AI and Journalism (Reporters Without Borders and 16 partner organizations, Global). States that human decision making must remain central to daily editorial choices and that responsibilities for the use of AI systems should be assigned to humans.
Controls to put in place
- Editorial policy that lists permitted and forbidden uses of AI
- Journalist approval recorded in the CMS for every AI assisted item
- Label on pages with AI assisted text or summaries
- Corrections log that links each correction to the tool and the prompt version
When it went wrong elsewhere
- Gannett halts AI written high school sports recaps after the Columbus Dispatch. MediaPost reports that Gannett stopped an effort to insert AI into its journalism in August 2023 after AI written sports write ups from Lede AI drew ridicule for lines such as a scoreboard being "in hibernation until the fourth quarter".
Frequently asked questions
- Do publishers publish AI summaries without review?
Gannett, the BBC and Newsquest say a journalist reviews the output. Gannett's disclaimer says the Key Points were reviewed by a journalist before publication, the BBC says its summaries are reviewed and edited before and after publication, and Newsquest says journalists review each draft to check it is accurate and well written. The Aftonbladet case study does not describe a review step for its summaries.
- How many articles can AI aided reporters produce?
Newsquest reported in June 2024 that it had published over 10,000 articles with the aid of AI, mainly press release based stories, community updates and event listings. It gave no time saving or correction rate.
- Do readers use AI summaries?
The Online News Association reports a click through rate of 43% on Aftonbladet's expandable AI summaries (53% among readers aged 19 to 36), and says elsewhere in the same case study that 40% of readers who saw the summaries expanded them. It also reports that readers who used them spent more time on the full article. These are rates reported in a case study, not a controlled test.
- Does the EU AI Act require a label on AI assisted news?
Article 50(4) requires disclosure of AI generated text published to inform the public on matters of public interest, unless a person has reviewed it and holds editorial responsibility. Many publishers add a label anyway, as the BBC says it will.
How to cite this page
Blits.ai AI Use Case Library, "AI drafting and article summaries for newsrooms with editor review", last verified 13 October 2026, https://www.blits.ai/ai-use-cases/newsroom-drafting-and-article-summaries. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 13 October 2026: First published