Editorial management with AI: between automation and human judgment

Pànxing —the project I described on the busiest day in The busiest Saturday: editor, ingest and a DNS that went down— manages several local and regional news portals in Catalonia. Every week it receives documents of all kinds: Word, PDF, images with text, scanned posters. All of this has to be turned into publishable articles on WordPress, written in Catalan and designed for SEO (the work of ranking well in search engines).

The process was manual, slow, and repetitive. Copy, paste, fix, adapt, publish. Hour after hour, week after week.

The idea

Simple: use AI (artificial intelligence) to automate most of the process. Have the system receive the original document, process it, and generate an article ready to publish. It's the same philosophy I explain in How I work with AI agents.

It had to work. What wasn't so obvious was how far AI could go and which parts would need human supervision.

What was built

A tool that divides the work into three phases: preparation, editing, and export.

In the preparation phase, the system receives the documents, extracts the text, runs OCR (recognizing text inside an image) when needed, normalizes the images, and converts everything to Markdown (a plain-text format with very simple markup) with a frontmatter (a few initial lines that describe the document): category, source, and other metadata.

In the editing phase, the AI generates the final article from the prepared content, following Pànxing's editorial guidelines. This is where it does most of the work.

In the export phase, the generated article is converted into the format WordPress needs to publish it.

Each phase can be reviewed. The result of one phase feeds the next, and at any point the process can be stopped for corrections.

The first problem: AI invents

It appeared quickly. In agenda articles —fairs, festivals, markets, cultural events— the AI generated information that wasn't in the original sources. Schedules that didn't exist, invented prices, activities that appeared nowhere.

The AI wasn't making things up on purpose; it was simply filling in the information with what seemed logical to it. But in an agenda article, an incorrect schedule isn't a minor error: it's misinformation published under the portal's name.

The solution: add a validation system that compares the schedules and key data in the generated article with the data in the original source. If something doesn't match, it's flagged for review. It isn't published.

The second problem: Catalan

Catalan has been a constant challenge. The AI generates correct texts most of the time, but there are errors it doesn't catch automatically: misplaced apostrophes, calques from Spanish, expressions that look correct but aren't.

For a portal that publishes in Catalan every day, this matters. You can't publish an article with visible spelling errors under the outlet's name.

That's why the AI generates, but a person reviews. The system isn't designed to replace the editor, but to save them the heaviest part of the work and leave them with the final quality.

The third problem: formats

Not everything that arrives is a clean document. There are posters with text over complicated backgrounds, PDFs with tables that don't structure well, and images with handwritten text that OCR doesn't recognize well.

Classic OCR works with simple texts, but when the material is complex or the quality is low, it falls short. This is where visual AI helps: it can identify text in images that OCR can't recognize, but it isn't perfect either.

The solution has been to accept that not everything can be automated. When the material is too complex, the system flags it and a person does the extraction manually. It's a small percentage, but it has to be planned for.

Three levels of fidelity

One of the most important decisions: defining three levels of editorial fidelity.

The normal level allows the AI to reinterpret and adapt the content to produce a fluid article. It works well for news or opinion pieces.

The faithful level forces the AI to stay closer to the original text. It's used when fidelity to the source matters, such as interviews or statements.

The strict level only allows reorganizing and cleaning up the original text, without adding or modifying any information. Designed for structured data, schedules, or factual information.

This differentiation is what allows AI to be used confidently for different content types without the result being unpredictable.

The role of the human

After the whole process, the conclusion is clear: AI can't do everything on its own. It takes on most of the heavy lifting, structures information, generates reasonable texts, and speeds up the process significantly. But it can't replace editorial judgment.

A person reviews that the schedules are correct, that proper names are spelled right, that the tone is appropriate, and that the information reflects what the source actually says.

AI prepares, the human verifies. That's the combination that works.