Automated data-to-text journalism & newsroom assistant

Regional and national publishers have to cover thousands of events that each matter to a small audience: a county-league football match, a municipal election in one district, a quarterly result, a traffic jam on the night shift. Writers cannot cover them all. Data-to-text systems turn structured live feeds into short, accurate, localised articles in seconds and update them as the data changes. LLM-based newsroom assistants then help editors with headlines, teasers, SEO metadata, rewrites of agency copy and tagging. The benefit is reach in the long tail and time for journalists to do original reporting. Editorial responsibility, the press code and AI labelling still apply.

Problem & value

• Today: local desks select only the events they can staff, and agency copy is rewritten by hand for each edition. Night and weekend shifts cover traffic, weather and results with minimal staff. Headlines, teasers, SEO fields and tags are typed manually for every story.
• AI value: complete coverage of data-rich, formulaic events; instant publication; consistent style; editors freed for reporting; more search and recommendation traffic through better tagging.
• Evidence from deployments:
• Retresco reports that RP ONLINE generated or updated more than 3.7 million texts with textengine.io from July 2020. The texts covered traffic, Covid-19 figures (more than 2,000 updates) and municipal elections in 280 polling districts. Data came from public sources such as the state electoral office [1].
• Swiss traffic-news centre Viasuisse uses Retresco to generate real-time reports for more than 1,000 traffic event types, starting with the night shift [2].
• Die RHEINPFALZ rolled out Retresco's AI Content Assistant to more than 190 journalists in 13 local newsrooms, hosted on European servers. Retresco reports 13% daily use [3].
• Swedish vendor United Robots supplies automated local content to Mediahuis, Schibsted, Bonnier, Reach, DPG Media and McClatchy, among others. One customer, NDC Mediagroep, uses it to cover 60,000 football matches per season [5].
• AP has automated quarterly earnings stories with Automated Insights since 2014. This increased the number of earnings reports it offers more than tenfold, and AP says it freed about 20% of reporters' time [8].
• Industry view: in the Reuters Institute 2026 survey of 280 digital news leaders, 97% called back-end automation (transcription, metadata and similar) important. Only 13% called their AI initiatives so far "transformational" [12].

Variants

• Results journalism: elections (per municipality and district), amateur and professional sports (match reports, league tables, player stats), stock markets and company earnings, weather and severe weather warnings, traffic, real estate sales, public statistics (unemployment, fuel prices).
• Blaulicht / police reports: police and fire-brigade press releases are summarised and localised. This is sensitive (named persons, suspects, origin of suspects) and needs mandatory editorial review.
• Newsroom assistant: headline and teaser A/B variants, SEO title and meta description, social posts, shortening and rewriting agency copy for each edition, live-ticker drafting, CMS entity and topic tagging, related-article linking.
• Wire services and B2B information: agencies and financial-information providers producing alerts and short reports for customers. A logistics or trade publisher might cover freight indices or port data.
• Public sector: statistical offices and municipalities publishing plain-text summaries of data releases and election results (combine with LANG-010 for plain language).
• Private / small publishers: hyperlocal blogs and club websites can generate match reports from federation result feeds, but must respect data-licence terms.

How it works

1. live feed (API/XML/CSV, push or poll)
2. ingest & validate (schema, plausibility, deltas vs. last version) [streaming]
3. event detection & newsworthiness rules (upset win, record turnout, threshold breached)
4. content planning: facts to mention, ordering, local angle per edition
5. realisation: rule/template NLG and/or LLM constrained to supplied facts
6. fact check: every number/name in text must match the source record [guardrails]
7. tagging & metadata (entities, location, topic page, SEO fields) [classifier/LLM]
8. CMS: auto-publish (low-risk formats) | draft queue for editor (sensitive formats)
9. updates on new data; AI label; audit log

• Two generation styles:
• Deterministic template NLG (Retresco textengine, Arria, axite, SimpleNLG/RosaeNLG) is fully traceable and cannot invent facts. Formats with high legal risk still mostly use it.
• LLM realisation produces more varied language and handles unstructured inputs such as press releases. It needs strict grounding: pass only the source facts, use structured output, and run a post-hoc number and name check.
• Hybrid systems use an LLM for wording within a template-defined fact plan.
• Assistant functions run inside the editorial CMS: the editor triggers a rewrite, headline set or tag proposal and accepts or edits the result.
• Human-in-the-loop: editors approve formats and templates once, then spot-check. Sensitive formats (crime, accidents, minors, elections on election night) go to a draft queue. A named editor owns each format.

Data preparation

Data-to-text needs clean structured feeds (sports results, election numbers, financial data, weather, statistical tables) plus the house knowledge: style guide, past articles, archive and templates. These live in the CMS, agency feeds, spreadsheets, statistical office downloads and sometimes in journalists' personal folders or on paper.

Typical gaps: feeds with changing formats, missing field definitions, duplicate or conflicting figures between sources, unlabelled archive articles, and a style guide that exists only as a PDF or in people's heads. Preparation steps: (1) inventory feeds and set one validated entry point per source (see PREP-005); (2) restructure tables into tidy, machine-readable form (see PREP-011); (3) define terms and KPIs (see PREP-012); (4) tag archive articles by topic, date and type (see PREP-010); (5) collect gold-standard articles per format as few-shot examples (PREP-016) and write down terminology and house spelling (PREP-015); (6) convert PDF or paper archives where needed (DOC-014).

Tools: Python/pandas or n8n for feed validation, the CMS API, an open-source OCR for old archives. Rough share of effort: 30-40 percent, mainly stable feeds and good example texts.

Implementation paths

Which path for whom

• enterprise: Path B — publishers and media groups need the assistant wired into their CMS and data feeds, and volume makes seat licences costly.
• smb: Path A — mature newsroom and text-generation products exist at SMB prices for local publishers and corporate communications.
• developer: Path C — open-weight models with templates and feed adapters let you ship data-to-text without per-article API cost.
• public: Path C — press offices and public broadcasters value sovereignty and procurement rules, and a self-hosted model on HW2 handles structured-data reports well.

Path A — Buy (recommended for most publishers)

• Data-to-text platforms: Retresco textengine.io (Berlin, DE; EU), United Robots (Malmö, SE; EU), axite, formerly AX Semantics (Stuttgart, DE; EU), Arria NLG (UK/US). Each covers sports, elections, finance, traffic or real estate as ready-made "formats".
• Newsroom assistants: Retresco AI Content Assistant (formerly Rewrite) and the AI functions built into editorial CMSs such as Arc XP, CUE (Stibo DX) and WordPress plugins. Large German groups also build in-house tools; Ippen Media and RND are examples.
• Choose A when the feed is standard (football, elections, markets, weather) and you need a CMS connector quickly. Check the data licence, EU hosting, how templates can be edited, the update logic and the audit trail.

Path B — Build on cloud APIs

• Models: Mistral (Mistral Medium/Large, Mistral Small; EU provider), Claude via AWS Bedrock Frankfurt or Google Vertex EU, Gemini on Vertex AI EU regions, OpenAI models via Azure OpenAI EU Data Zone. German sovereign hosting is available through IONOS AI Model Hub and STACKIT AI Model Serving.
• Pattern: the feed goes to a fact-plan JSON, then to an LLM with a JSON-schema output, then to a validator that compares numbers and names with the source, then to the CMS API.
• Orchestration: n8n, Airflow or a Kafka consumer. Prompt templates are versioned per format.
• Choose B when the formats are your own (regional statistics, special data sets) or when you build assistant functions into your own CMS.

Path C — Self-host (open source)

• Models: open-weight Mistral Small 3.x, Qwen3 (dense 8B–32B or MoE), Llama 3.x and gpt-oss-20b, served with vLLM or Ollama using guided JSON decoding. German output quality is good from about 14B parameters upward.
• Deterministic NLG: RosaeNLG (supports German) and SimpleNLG.
• Tagging: a fine-tuned German BERT-class NER and topic classifier, with embeddings (e.g. multilingual E5, BGE-M3) in Qdrant or OpenSearch for related-article and topic-page assignment.
• Hardware: HW2 (one 24–32 GB GPU) is enough for a quantised 14–32B model generating thousands of short texts per hour. Use HW3 for larger models or many parallel assistant users. Ops effort is moderate: feed monitoring matters more than model ops.

Path D — Custom training

• Worth it for very high volumes of one format, for a distinctive house style, or for a small model that writes match or election reports cheaply on CPU or a small GPU. Fine-tune (LoRA) on past pairs of feed data and the published edited text; a few thousand pairs per format is typical.
• A tagging classifier fine-tuned on the publisher's own tag taxonomy is usually the most valuable custom model.

State of the art (2026)

• Hybrid NLG is best practice: deterministic fact selection plus LLM wording, with automatic fact verification. Research benchmarks for data-to-text (RotoWire for sports, WebNLG, ToTTo) consistently show that neural generators produce fluent text but add or alter facts. That is why production systems check facts after generation.
• Agentic newsroom tools: LLMs now draft live tickers from feeds, propose headline variants for testing, and optimise texts for AI answer engines (generative engine optimisation) as well as Google.
• Transcription and multimodal input (interviews, council meetings) are feeding newsroom assistants. See SPEECH-001 and SPEECH-021.
• Audience acceptance is limited. Reuters Institute's Digital News Report 2025 finds audiences still sceptical of news made mainly by AI, more accepting of AI supporting journalists, and Europe more sceptical than the US [13]. An EBU study cited by Reuters Institute found AI assistants misrepresent news content almost half the time [12]. Labelling and human ownership are therefore commercial as well as legal requirements.
• Limits:
• Feed errors are published at machine speed.
• Thin template texts can count as low-value content for search engines.
• LLMs over-dramatise ("sensational win") and invent quotes.
• Crime and accident reports risk identifying people and stereotyping.

Shared resources & combinations

• Same LLM serving stack (llm-general on HW2/HW3) as LANG-001 marketing copy, LANG-005 SEO content, LANG-007 social posts, LANG-011 style rewriting and LANG-010 plain language. A publisher can run one vLLM server for all of these.
• Tagging and entity layer (text-classifier, embedding-text, vector-db) is shared with LANG-015 NER, DATA-007 metadata enrichment, REC-002 news recommendations, KNOW-019 archive discovery and KNOW-016 multimodal archive search.
• Streaming and workflow layer is shared with DATA-002 automated reporting narratives, which uses the same data-to-text technique for internal audiences.
• Natural combinations:
• LANG-009 translation for multilingual editions
• LANG-021 fact-checking
• SPEECH-006 text-to-speech for audio versions
• SPEECH-004 live captioning
• VIS-025 and VIS-055 for sports video highlights alongside text reports
• PRED-017 weather forecasting as an input feed
• ANOM-015 provenance and AI-content marking (C2PA, Art. 50)
• LANG-018 media monitoring

Effort, cost & KPIs

• Path A: weeks per standard format once the CMS connector exists. Pricing is usually per format, per text volume or per seat for assistants.
• Path B: 1–3 months for a first custom format with validation and CMS integration. Token cost per short article is low compared with editing and data-licence cost.
• Path C: 2–4 months plus one GPU server and feed monitoring.
• Cost drivers: data licences (sports and market data are often the largest item), template and format design, editorial QA time.
• KPIs:
• formats live and texts published per day
• error rate (corrections per 1,000 texts)
• time from data event to publication
• page views and subscriptions from automated texts
• share of assistant suggestions accepted
• editor hours freed
• search and Discover reach of tagged articles

People & change

Involve the editor-in-chief or management (editorial guidelines, labelling policy), the desk heads as process owners, and the journalists who will use the tool daily, since acceptance depends on them. IT connects CMS and feeds; the legal team checks copyright, press law and labelling; the DPO is needed where reports name individuals or use reader data (AVV with each provider, DPIA only if profiling at scale).

Works council: depends. Where a council exists, it must be informed about AI-supported work processes (§90(1) no. 3 BetrVG), and tools that log individual use fall under §87(1) no. 6. Note that press companies are Tendenzbetriebe (§118 BetrVG), which can limit co-determination on editorial matters, so clarify early with legal counsel. In public broadcasters the staff council applies.

AI literacy: support staff with training on hallucinations, source checking and labelling duties (Art. 4 AI Act as amended; Art. 50 transparency applies since 2 Aug 2026). Communicate that AI drafts routine data stories so journalists have time for research. Pilot with one standardised format (e.g. match reports) and a small volunteer desk. Typical concerns are job loss and errors: address them with a rule that nothing is published without editorial review, visible corrections, and clear scope for what AI may write.

See GUIDE-008, GUIDE-006 and GUIDE-014.

Risks, regulation & compliance

• EU AI Act: typically limited risk.
• Art. 50(4), applicable from 2 August 2026, requires deployers to disclose AI-generated text published to inform the public on matters of public interest. The duty does not apply where the text has undergone human review or editorial control and a person holds editorial responsibility [9].
• Fully automated formats published without per-text review should therefore be labelled.
• Providers must machine-mark synthetic output under Art. 50(2), except for assistive standard editing.
• The Commission's final Code of Practice on transparency of AI-generated content (June 2026) is the reference for marking and labelling practice [10].
• No Annex III category applies. AI literacy (Art. 4) applies to editors.
• Press ethics (DE): the Pressekodex preamble (version of March 2025) states that editorial responsibility also covers artificially generated content [11]. The Pressekodex also requires due care on accuracy (Ziffer 2) and restraint in naming suspects' origin (Ziffer 12, Richtlinie 12.1). Similar self-regulation codes exist in AT and CH and in the IPSO and IMPRESS codes in the UK.
• GDPR: election and market data are mostly non-personal. Amateur sports (player names, often minors) and police reports are personal data.
• Journalistic processing benefits from the media privilege (Art. 85 GDPR, implemented in German state press laws and the Medienstaatsvertrag), but data security and processor agreements with vendors still apply.
• Prefer EU hosting.
• Data licences and copyright: league, federation and stock-exchange data are licensed. Check the rights to use the data for automated publication and for model training.
• Mitigations:
• plausibility checks on feeds and kill-switches per format
• number and name verification
• draft queues for sensitive formats
• visible AI labels and correction workflows
• an audit log linking each text to its source data and template version

Getting started

1. List formulaic, data-rich events you do not cover today and check which data feeds and licences exist for them (electoral offices, federations, DWD open weather data, market data).
2. Pick one low-risk format (weather, traffic, league results) and pilot a Path A vendor with your CMS.
3. Define the editorial owner, AI-label wording and correction process per format before going live.
4. Add a newsroom assistant for headlines, teasers and tags. Measure the acceptance rate and time saved.
5. Expand to elections and sports, and only then to sensitive formats such as police reports with mandatory review.

Sources

1. Retresco — Case study: automated reporting at RP ONLINE — https://www.retresco.de/case-study/automatisierte-berichterstattung-rponline/
2. Retresco — Case study: Viasuisse automated traffic news — https://www.retresco.de/case-study/viasuisse-automatisierte-nachrichten/
3. Retresco — Case study: Die RHEINPFALZ AI Content Assistant — https://www.retresco.de/case-study/lokal-zeitung-ki-die-rheinpfalz-ai-content-assistant/
4. Retresco — Products (textengine.io, AI Content Assistant) and case studies — https://www.retresco.de/
5. United Robots — automated local content for publishers — https://www.unitedrobots.ai/
6. Arria NLG — data-to-text and hybrid LLM authoring — https://www.arria.com/
7. axite (AX Semantics GmbH) — https://axite.io/
8. Wikipedia — Automated journalism (AP earnings and Automated Insights, Quakebot) — https://en.wikipedia.org/wiki/Automated_journalism
9. EU AI Act Article 50 — transparency obligations — https://artificialintelligenceact.eu/article/50/ ; Regulation (EU) 2024/1689 — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
10. European Commission — Code of Practice on transparency of AI-generated content — https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
11. Deutscher Presserat — Pressekodex — https://www.presserat.de/pressekodex.html
12. Reuters Institute — Journalism, media and technology trends and predictions 2026 — https://reutersinstitute.politics.ox.ac.uk/journalism-media-and-technology-trends-and-predictions-2026
13. Reuters Institute — Digital News Report 2025, executive summary — https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2025/dnr-executive-summary