Despite remarkable advances in large language models, generic AI tools continue to fall short in meeting the unique demands of journalism. Factual precision, exact text lengths, and proper quotation handling remain challenging for off-the-shelf AI solutions.

Instead of waiting for perfect solutions, a small team at Badische Zeitung developed BZ.echo — a purpose-built platform created initially to boost newsroom efficiency and free up journalistic resources for higher-value tasks.

Without requiring users to write complex prompts themselves, BZ.echo employs agentic workflows and prompt chaining with non-AI tools for specialised quality control, significantly improving factual accuracy to meet journalism’s standards.

The results speak for themselves: Editors save up to an hour daily, processing times are cut in half, and adoption has grown to hundreds of daily runs within months of launch — all while maintaining the editorial quality and human judgment.

Beyond generic AI: agentic workflows

The platform incorporates agentic workflows that act as intelligent orchestrators, routing content through appropriate AI and non-AI tools based on specific task requirements.

Instead of relying on complex single prompts that attempt to accomplish everything simultaneously, the platform breaks editorial tasks into smaller, more manageable steps. This architecture allows for targeted quality control at each stage and significantly reduces hallucinations and factual errors.

For example, when editing local freelance reports, BZ.echo first extracts quotations and numbers, calculates necessary adjustments to meet length requirements, and analyses text structure before commissioning edits. After editing, control loops verify quotations, numbers, and other facts before presenting the final text to the user.

In the output, all words not contained in the input text are highlighted in colour, enabling efficient review of the results together with BZ. echo’s side-by-side editor.

Designed for working journalists

BZ.echo’s interface prioritises usability for journalists with varying technical backgrounds. Rather than requiring prompt engineering skills, editors interact with task-oriented interfaces that seamlessly integrate with existing workflows.

The system supports numerous editorial tasks, such as:

• Generating headlines, teaser text, and SEO-optimised meta tags.
• Proofreading and standardising numbers, units, and abbreviations.
• Editing freelance content to standardised lengths.
• Creating draft articles from press releases and police reports.
• Developing Q&A articles from text inputs or archive URLs.
• Crafting social media posts and polls.
• Adapting content for vertical platforms.
• Transcribing and editing interviews.
• Drafting sports match reports from live tickers.
• Summarising articles for print front pages and subsequent reporting.

Cutting down factual errors

Since its December 2024 launch, BZ.echo has demonstrated remarkable adoption rates and efficiency gains. Usage quickly scaled from 100 workflow runs per day to approximately 800 daily runs as staff completed AI training.

The productivity improvements are significant: Editors can save up to an hour daily using BZ.echo or even more when transcribing and processing interviews.

Processing time for routine news like traffic accidents and minor crime reports has been reduced from 10 to 15 minutes to just five to seven minutes. While maintaining editorial control, the platform frees journalistic resources for higher-value tasks.

Quality improvements are equally impressive. The system’s prompt chaining techniques significantly reduce factual errors, while external tools ensure precise character and word counts. AI-edited articles preserve original writing styles while enhancing clarity and readability.

Building AI literacy

Implementation included comprehensive two-hour training sessions covering LLM fundamentals, limitations, editorial guidelines, and best practices. Support continues through an Editorial AI Playbook, video tutorials, and monthly feedback sessions, building essential generative AI literacy across the newsroom.

Badische Zeitung found that combining editorial expertise with advanced AI architecture effectively transforms general-purpose LLMs into specialised journalistic tools. While powerful, agentic workflows require regular simplification and optimisation to maintain performance.

Also, launching an AI tool is just the beginning; successful implementation across a newsroom of over 100 editors requires ongoing commitment.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .