MOESS

AI editorial system for publishers and magazines

AI writes. Your editorial team decides.

Agentic AI for your publication’s content operations. Seven specialized AI agents handle research, drafting and fact-checking, adapt the text to your magazine’s voice, format it and optimize it for Google and AI search. Only what your editorial team has approved at three points gets published.

Book an initial call See a newsroom day

Already in production at a Munich publisher

Draft proof · GuideVersion 2

Minijob alongside your pension: what you can earn in 2026

Since 2023, there have been no additional earnings limits for old-age pensions. This also applies to early old-age pensions.

For the Minijob (German marginal employment) itself, a limit of €603 per month (as of January 2026) applies. It is tied to the minimum wage and rises with it.

7specialized AI agents, each with a clearly defined task
3approvals by your editorial team before anything is published
0changes in your CMS that the system overwrites
< €1.50model costs for a guide with its own research

Measured model costs for a researched guide up to publication: US$1.53.

What it does for your publishing house

More topics with consistent editorial care

Routine tasks are handled by the agents; editorial decisions remain with your editorial team.

Speed

From topic to draft proof in minutes

The editorial assistant finds topics and gaps; the pipeline delivers research, briefing and first draft. Your editors work from a sourced draft from the outset.

Control

Nothing is published without approval

Human in the loop is built into the system at three points. Each approval is logged with name and timestamp (audit trail), so it can be proven who is responsible for what.

Costs

All costs in the log

Every model call is recorded with agent, duration and cost. This makes it possible to demonstrate cost-effectiveness instead of estimating it.

Workflow

The path of an article

The workflow is illustrated with a production run, a guide about Minijobs alongside a pension. Seven agents with clearly defined tasks work in sequence, optionally joined by a Picture editor. At every handover, the result is checked and saved; at three points, a human decides.

  1. In
    Your editorial team

    Topic and requirements

    You set the topic, key message and structural preferences. Whatever your editorial team sets as fact is adopted and not researched again. The editorial assistant already supports the topic meeting. It knows what your magazine has published so far and finds gaps and outdated articles.

    Set: Since 2023, no additional earnings limits for old-age pensions, including early ones.
    Avoid: Pressure, anglicisms, unexplained bureaucratic German.
  2. A1
    Agent 1 · Researcher

    Dossier and briefing

    The Researcher searches the web and your knowledge folder of your own material such as reader surveys or studies, always with page references. Documents marked confidential are left out. Every piece of information goes into the dossier with source and as-of date. This becomes the briefing with structure, target audience and dedicated safeguard rules for this article; in the example there are twelve. An excerpt from the dossier and briefing:

    €603 earnings limit · since 01/01/2026
    €42.52 pension value (Rentenwert) · as of July 2026
    Knowledge folder: Reader survey 2026, p. 4
    Excluded: Media kit 2026 · confidential
    Rule: Don’t mix up reduced earning capacity pension and old-age pension.
    Rule: No invented quotes.
  3. 1
    Approval 1 of 3 · Your editorial team

    Is the direction right?

    You approve the briefing or send it back with a note, such as “Shorter title, more focus on the widow’s pension”. The note is worked in. Without this approval, not a single word is written.

    Step in yourself
    You change individual points such as the headline or the order of sections directly by hand. On request, the AI suggests five alternative headlines.

  4. A2
    Agent 2 · Writer

    The first text

    The Writer writes the article from the briefing and dossier. It sticks to the structure and safeguard rules and uses only facts that are in the dossier.

  5. A3
    Agent 3 · Fact-checker

    Every statement against the sources

    The Fact-checker checks figures, as-of dates and statements against the dossier and rates the text. Anything not backed up goes back to the Writer with a reason. This quality loop runs before a human sees the text.

    verified €603 Minijob limit, as-of date given
    verified No additional earnings limit since 2023
    not verified “Around 40 percent in retail” → back to Writer
  6. A4
    Agent 4 · Tone of Voice

    Your voice

    Tone of Voice adapts the text to your magazine’s language. It works from tone, style rules, banned terms and your best texts as models.

  7. A5
    Agent 5 · Formatter

    Ready for the CMS

    The Formatter sets subheadings, info boxes, teasers and bullet lists and brings the article into your CMS’s format.

  8. +
    Optional · Picture editor

    Images with clear labeling

    On request, you can add the Picture editor. It picks images from your own image database or generates and edits images with AI. AI images are always labeled accordingly.

  9. A6
    Agent 6 · SEO / GEO

    Findable on Google and in AI answers

    SEO / GEO optimizes the title, meta description and structure so the article is found in Google Search and in answers from AI assistants such as ChatGPT or Perplexity.

  10. A7
    Agent 7 · Editor-in-chief

    The overall verdict

    The Editor-in-chief checks the finished article against the briefing and safeguard rules. If the quality is insufficient, the text goes back for revision. Your editorial team only sees texts that have passed all checks.

    passed Briefing and safeguard rules followed
    passed Fact-check with no open points
    passed No banned terms
    passed Sample calculations labeled as such
  11. 2
    Approval 2 of 3 · Your editorial team

    Sign-off on the text

    You read the finished article and sign it off. The approved version is then frozen. Every later change becomes a new version.

  12. 3
    Approval 3 of 3 · Your editorial team

    Publication in the CMS

    Only now does the article go into your CMS. If someone has changed something there in the meantime, the system overwrites nothing and reports the conflict.

All steps at a glance

Guardrails

Rules that can’t be switched off

These rules are built into the software itself, right down to the database, and are not merely described in a handbook.

Approval

No article without humans

The three approvals cannot be skipped, not even via detours in the system.

Versions

Approved means frozen

Signed-off versions remain unchanged. You can see at any time which text was approved when and by whom.

CMS

Never overwrite

If your editorial team has made edits in the CMS, the system detects the change, leaves it intact and asks how to proceed.

Facts

No figure without evidence

Every figure and every statement must be in the Researcher’s dossier. Anything not backed up fails at the Fact-checker.

Voice

Your tone, your rules

The editor-in-chief maintains the brand profile, including style rules, banned terms and reference texts.

Images

AI images always labeled

Images come from your own image database or are generated and edited with AI. AI images always carry a label.

Data protection

Data storage and data protection at a glance

Your data protection department will ask about the GDPR and where texts are transferred. The following overview describes the current state and the points I am still working on.

How it is today
  • Servers in Germany. Your instance runs in a data center in Nuremberg, and so do the daily backups.
  • A separate instance per publisher. There is no shared database with other customers. If you prefer, the system runs on your own premises.
  • Its own sign-in. Sign-in does not rely on a login service from a US provider. Passwords are never stored in plain text, and access expires automatically.
  • The language model is in the US. For research and writing, the briefing, dossier and text are sent to Claude by Anthropic. Anthropic does not use content sent through this interface to train its models.

Costs

What an article costs is in the log

For every call to a language model, the agent, duration and cost are recorded individually. These figures can be compared with the costs a researched guide incurs today.

Log · Excerptproduction run
AgentDurationCostStatus
Writer139.7 s$0.19aborted

Even an aborted attempt is in the log with its costs. Here, the text was longer than allowed. The cause was immediately visible and fixed the same evening. The language model provider bills in US dollars.

Guide with its own researchunder €1.50

The complete workflow with web research. For a guide about checking accounts, model costs of US$1.53 up to publication were measured.

Biggest item
Web research. In the Minijob example, research and briefing together cost US$0.87.
Processing time
About five minutes to the finished briefing, then a few minutes for the text.
Revision
If you send a text back, it is rewritten based on the existing dossier, without new research.
Article from a press releasesignificantly less

The press release is supplied as the source; the Researcher’s web search stays off. That largely eliminates the most expensive step.

Dropped
Web research and cross-checking against further sources on the web.
Retained
All three approvals and the fact-check, now against the press release.
Exact figure
Measured in the pilot using your own press releases.
Model per magazineselectable

Each magazine gets a fixed language model. It only changes when you deliberately switch it, so tone and quality don’t fluctuate unnoticed.

Compact
The fastest and cheapest model, suited to short news items and checklists.
Standard
The costs per guide stated above are based on this model.
Premium
The strongest writing model, about two and a half times as expensive per article.

Outlook

What comes next

All the features described are in use today, including the editorial assistant, knowledge folder and direct editing. Here is what I’m working on next. The editorial teams in the pilots help set the order.

Planned

Complete archive

The entire archive from your CMS is imported, including articles from before the system. New texts link to it.

Planned

WordPress and other CMSs

The interface is prepared. WordPress is next, other systems as the pilots require.

Idea

Fact store with expiry date

Verified facts are reused until they expire, for example when the minimum wage changes. Follow-up articles are produced faster and at lower cost, without carrying over outdated figures.

Pilot

Getting started with a pilot

You start with a manageable pilot using one of your titles and current topics from your planning.

  1. Adopt your voice

    Your best texts and style rules become the brand profile for Tone of Voice, optionally together with your image database. Your editor-in-chief reviews and adds to it.

  2. Connect your CMS

    Sanity is connected today, WordPress is in preparation. Your existing workflows remain unchanged.

  3. Five topics from your planning

    Your editorial team runs five of its own topics through the system and decides at every approval itself. Afterwards, the benefit for your editorial team can be assessed on a sound basis.

Frequently asked questions

What publishing executives ask first

Does the system replace my editors?

No. The agents handle research, the first draft, fact-checking, formatting and search engine optimization. Your editorial team decides at three points what gets published. The focus of their work thus shifts from fixing to deciding.

What happens if the AI invents facts?

A dedicated Fact-checker checks every statement against the Researcher’s dossier, with source and as-of date. Anything not backed up goes back to the Writer before a human sees the text. Hallucinations are thus caught within the pipeline.

Do I have to label AI texts?

The EU AI Act requires labeling for AI-generated texts on matters of public interest, unless a human has reviewed them editorially and holds responsibility for them. The system documents every approval with name and timestamp.

Assessing your specific case is a matter for your legal counsel.

How does Google view AI-assisted articles?

According to Google, it evaluates the quality and usefulness of content, regardless of how it was produced. Verified facts, clear as-of dates and editorial review meet these requirements. In addition, the SEO / GEO agent makes sure titles and structure are readable for search engines and AI assistants.

Which CMSs are supported?

Sanity is connected today, WordPress is in preparation. Other systems can be connected via a prepared interface.

Are we tied to one language model?

No. Today the system works with Claude by Anthropic, but the language model can be swapped without rebuilding the pipeline. You are therefore not tied to a single provider (no vendor lock-in).

Contact

Test it on one of your topics.

In the initial call, I’ll show you a run from topic idea to publication. On request, we use a topic from your current planning.

See a newsroom day first
Contact person Dieter Möss
  • Call of about 45 minutes, online
  • Demo of a complete run
  • No preparation needed on your side
Inquire by email WhatsApp

Or directly: dieter@moess.de · +49 152 02838168