AI Content Factory from the Inside: A Detailed Breakdown of How the System Works
This is a detailed breakdown of the product briefly described in the post «AI Content Factory». Here - how everything works on the inside: the route one news item takes through seven steps, where the human stays in the loop, numbers, and stack.
A Familiar Situation for a Content Director
A B2B company represents solutions from several dozen vendors across ten to fifteen countries. Each vendor regularly posts on LinkedIn: news, updates, client cases. Almost everything - in English, with a global focus.
To turn this into story ideas for a local audience, the team needs to monitor the flow, select what matters, translate it, rewrite it in the brand voice, design visuals, and distribute across websites, social media, and digests. In every language. In every country. Every working day.
Most companies do this by hand. For some, it takes up half the content team. For others, only what caught someone’s eye first gets published - the rest gets lost. Meanwhile, the local audience reads the original on the vendor’s page, not your content.
Where It Hurts and Why Doing It Manually Is No Longer an Option
These pain points repeat almost everywhere we go with this kind of request.
The source flow consumes the team. Twenty to thirty vendors, each publishing several times a week - that’s hundreds of pieces per month. Monitoring them is one job. Selecting what matters is another. Rewriting to fit the company voice and translating into N languages - two more. A marketer turns into an assembly line with no time to think about strategy.
Important news gets lost. In a heavy flow, whatever came first gets published. A month later the same story resurfaces as a client question: “why did we hear about this from LinkedIn and not from you?” The opportunity was there, a competitor covered it, you didn’t.
Reach is below potential. The vendor operates in a dozen countries, but your content goes out in one or two languages. Local audiences read the original at the vendor’s page. You lose SEO traffic in local search engines and reach in local social channels.
Digests are made “on weekends”. Once a week someone manually assembles them by country and language. Always longer than expected, always “we’ll finish it later.” The newsletter goes out late, and by then the top story is no longer top.
What We Built
The content factory takes your vendors’ news and delivers publication-ready materials: written in your company’s voice, in all your languages and channels, under your team’s control.
Let me address three common questions upfront. The final “publish” is always pressed by a human - no auto-posting on behalf of AI here. Texts are written in your name and your tone, not retelling the vendor in their own words. And this is not a boxed product: the architecture is the same for everyone, the configuration is per client.
One vendor news item produces a website article with full SEO markup, a post for LinkedIn and Facebook (style matched to the nature of the news), a block for the weekly digest, and branded banners for each language. All of this - immediately across all target languages, websites, and social channels.
The Route of One News Item: Seven Pipeline Stages
Here is the journey a specific news item takes from vendor to publication.
Step 1. Source. The system monitors your vendors’ LinkedIn pages at varying frequencies: active ones are polled daily, medium ones weekly, rare ones monthly. Sources with nothing happening are not triggered and we don’t pay for empty runs.
Step 2. Filter. The news item goes through a classifier. Webinar announcements, videos, event registrations, promotions - filtered out here. Only corporate news and topically relevant content moves forward. This saves time and AI processing cost: we only pay for what will actually become content.
Step 3. Scoring. Four AI agents evaluate the filtered news item: by content, geographic relevance, technical value, and commercial potential. Each gives a score and explains why. A coordinator aggregates the scores. The manager sees a card with the original post, a full translation of the vendor page text, and the final score - and can approve or lower the score with a comment. This is how the system calibrates to your brand voice.
Step 4. Generation. Three formats are produced from one news item. Each goes through a pipeline of three AI agents: the first sets the style, the second writes the marketing brief, the third - the final text. The output is a website article of ~20,000 characters with SEO fields, a social media post of ~1,500 characters, and a digest block of ~500 characters.
Step 5. Translation. Texts are translated into target languages - not by machine Google Translate, but by dedicated AI agents for each language, each with its own professional prompt. Coverage is ready for 13-14 languages; new countries are enabled by configuration, not development.
Step 6. Visuals. Branded banners are assembled in parallel. AI generates the background, then the pattern; the manager approves both, and until they do - the process doesn’t move forward. After approval the system pulls the vendor logo, picks one of the pre-created branded design templates, and assembles two banners per language: one for the website, one for social media. Consistent style, no designer needed for each publication.
Step 7. Publication. Finished materials appear as drafts across all your WordPress sites, and all your LinkedIn and Facebook pages at once. The social post includes a link to the article on your website, in the same language, so the reader comes to you and not to the original at the vendor’s page. The team reviews and adds to the calendar.
Where the Human Stays
This is the key question for anyone responsible for a brand: if AI writes everything, what do I do when it writes something wrong in my name?
The answer is simple - AI doesn’t write everything. AI does the work, the human makes the choices. There are four manual control points in the chain.
Every news item is approved or rejected by the manager before generation kicks off; they can lower the score and explain why - this feedback is recorded and influences future selections. The background, pattern, and sample banner go through manual approval: until the manager approves the visual, the system does not proceed. Texts are saved ready but are not published on their own - scheduling and timing are up to the team. Style, voice, prompts, and templates are configured once to match your brand and locked in. Not “generic B2B”, but your specific voice.
That’s the boundary: AI handles the routine, marketing manages the meaning. Human control is not a temporary crutch during a learning period - it is a permanent part of the design.
What’s in Production and What It Costs
This is not a concept or a lab pilot. The system is currently running in production for a client: around thirty vendors, publications in approximately fourteen countries in as many languages. Every working day - two to three finished stories, each expanded into the full set of websites, social channels, and a block in the weekly digest.
The full AI load for such a factory comes to approximately $250 per month. That is the total spend on neural networks: text generation, translations into all languages, content processing and classification, visual generation. A system that covers the work of a multi-person content team costs less than a monthly subscription to a mid-range corporate SaaS.
Why This Is Not a Raw Solution
Under the hood - an enterprise stack, nothing exotic. Microsoft 365 for approvals and archive (Teams, OneDrive, Excel). WordPress for articles. LinkedIn and Facebook API for social media. Microsoft Dynamics 365 Marketing for newsletters. AI models from Anthropic, OpenAI, and Together.ai, so there is no single-vendor dependency.
The automation itself runs on n8n: a dozen reusable modules, fan-out publishing by regional groups, async approvals via Teams, fault tolerance at every critical node. This is a system that already runs every day - not something we’ll build from scratch for your project.
What “Adapting to the Client” Means
We don’t sell a boxed product. The architecture is ours, the configuration is yours.
Before configuring anything, we sit down and go through your inputs: which vendors and how actively each one publishes; how many websites, which localizations and CMS; which social channels and who currently posts to them; which languages are priority and which are on the horizon; what your brand voice is - formal, friendly, technical; how approvals work and who signs off on what; whether you have banner templates or they need to be designed from scratch.
After that, the factory is configured to your setup. Prompts - to your style, approval points - to your hierarchy, templates - to your design guide, regions and languages - to your footprint. What you get is not another installation, but your own system built on a proven architecture.
What to Do Next
If you have a large content team drowning in the flow of vendor news; or a small one that physically can’t cover all languages and countries; or digests that are always made “on weekends” - let’s look at what can be solved.
We start with a free 60-minute audit of your content process. No obligations. In the meeting we break down where news currently comes from and how it’s selected, how much time each format and language takes, where exactly you’re losing reach and why, how your sources and languages will fit into the factory, and what effect you’ll see in the first month.
If the solution is obvious - we’ll say what it is and how long it takes. If automation won’t pay off in your case - we’ll say that too. Our studio principle: we don’t implement AI for AI’s sake.
Write to us - we’ll dig into your situation and show how the factory will work in your context. The first consultation is free.