AI Content Pipeline: What $0.30 Per Article Actually Automates (And What It Doesn’t)
An AI content pipeline is easy to sell with a percentage and hard to judge by one. “90% automated” gets thrown around until it means nothing. Here’s what the pipeline I built and run actually costs, what it does unattended and, just as important, where a human still has to step in.
Key takeaways
- Ask for the cost per article, not the automation rate. My pipeline drafts an article for roughly $0.30 in Claude cost. That figure can be checked against an invoice, “90% automated” can’t.
- The pipeline automates the middle of the work. Research synthesis, first-draft structure and on-brand phrasing at scale. Not the decision what to write, and not the decision whether it’s good enough to publish.
- Every draft lands as a WordPress draft, never as a live post. Nothing goes out until a human has opened it and said yes.
- The human layer is the part that doesn’t get cheaper. Brief, fact-check, final edit and the publish decision are what protect a brand’s credibility, especially in regulated markets.
In this article
- What is an AI content pipeline?
- What the AI content workflow actually does
- AI content automation: what it covers and what it doesn’t
- AI content pipeline cost: why $0.30 tells you more than “90% automated”
- Why regulated markets need a human review gate
- Four questions before you buy an AI content workflow
- Frequently asked questions
What is an AI content pipeline?
An AI content pipeline is a chain of automated steps that takes a topic and turns it into a finished draft: research, structure, writing and staging in the CMS, run on a schedule instead of by hand. The models do the labor-intensive middle. People decide what goes in at the start and what comes out at the end.
I run one for ExpatAndalucia, one of my own publishing properties. It’s a GitHub Actions pipeline built in Python that researches, drafts and stages articles as WordPress drafts on a schedule, no manual kickoff required. It orchestrates a small stack of models: Claude Sonnet for drafting and structuring, Perplexity for source research, and Gemini in a supporting role.
The Claude cost per drafted article runs at roughly $0.30. That number is real and specific, which is exactly why I’d rather lead with it than with a vague automation percentage.
What the AI content workflow actually does
On a schedule, the pipeline runs through the same four stages every time:
- Pick up a topic, without anyone pressing a button.
- Research it via Perplexity, so claims are grounded in current sources rather than in a model’s training data.
- Draft the article with Claude Sonnet against a fixed structure and style brief. Gemini handles a secondary pass in the workflow.
- Push the result into WordPress as a draft, never as a published post.
That last stage is the part most “AI content automation” pitches leave out. Nothing goes live without a human opening the draft and deciding whether it’s ready.
How the schedule works
The pipeline runs as a scheduled GitHub Actions workflow, which starts on a cron schedule by itself. GitHub sets the shortest interval at once every five minutes. A content pipeline needs nothing close to that, but it does need to run when nobody is watching.
AI content automation: what it covers and what it doesn’t
The pipeline automates the labor-intensive middle of content production. It does not automate the judgment calls at the start, what’s actually worth writing about, or at the end, whether the draft meets the bar to publish.
What the pipeline does unattended
- Research synthesis from current sources
- First-draft structure against a fixed brief
- On-brand phrasing at scale
- Staging the result in WordPress as a draft
What still needs a human editor
- The editorial brief: which topics are worth the reader’s time
- The fact-check pass on every claim the draft makes
- The final edit
- The publish decision
None of this is an argument against AI-assisted content operations. It’s the opposite. Systems like this genuinely change what a small content team can produce. The argument is against pretending the human layer is optional, because that’s the layer that decides whether the output is actually usable.
AI content pipeline cost: why $0.30 tells you more than “90% automated”
When a client asks what an AI-assisted content operation actually costs to run, “90% automated” doesn’t answer the question. $0.30 in model cost per drafted article does. It tells them exactly what scales linearly with volume and what doesn’t.
The rest of the cost structure is human: the editorial brief, the fact-check pass, the final edit, the publish decision. That’s the part that doesn’t get cheaper just because a model wrote the first draft, and it’s the part that actually protects a brand’s credibility.
A cost-per-article figure can be checked against an invoice. A “90% automated” claim usually can’t be checked against anything, and for content buyers who’ve been burned by thin AI output before, that’s a red flag, not a selling point.
What Google says about AI content
Google’s position is less dramatic than most vendor decks suggest. Appropriate use of AI or automation is not against its guidelines. What its spam policies call scaled content abuse is generating many pages for the primary purpose of manipulating rankings and not helping users, no matter how they were created. So the question is never whether a model wrote the draft. It’s whether someone checked that the page adds anything.
Why regulated markets need a human review gate
This distinction matters even more in regulated markets. Across a decade writing for regulated German-speaking markets, bwin, PartyCasino and Sportingbet under Entain among them, I’ve never seen a compliance team accept “the AI checked it” as a substitute for a qualified human review.
A content system built around AI drafting still needs that gate. Anyone selling you a fully unattended pipeline for regulated content either hasn’t shipped one at scale, or is leaving out the part where something eventually went wrong.
Lukas Thies
The same logic decides where automation helps and where it multiplies damage. Point a cheap pipeline at a site structure nobody has decided on, and it fills the wrong pages faster. I described that failure in the four-page trap on category pages.
Four questions before you buy an AI content workflow
If you’re evaluating an AI content vendor, these four questions separate a working system from a slide:
- Ask for a cost-per-article figure, not an automation percentage. One is auditable, the other is marketing.
- Ask where the human review gate sits. If the answer is “there isn’t one”, that’s the risk you’re actually buying.
- Ask what the research layer is grounded in. A model drafting from its own training data will confidently invent details a model working from live research won’t.
- Ask how compliance language is handled. In regulated categories, this should be a reviewed, reusable template, not something the model improvises per article.
If a vendor can answer all four without hesitating, you’re talking to someone who has run a pipeline, not just demoed one. Most of the time, the problem behind a failing content operation isn’t the tool anyway. It’s the system around it, which is what I wrote about in why most iGaming and SaaS content operations fail. If you want to see how I set that system up, here is how I work.
Frequently asked questions
What is an AI content pipeline?
An AI content pipeline is a chain of automated steps that turns a topic into a staged draft: research, structure, writing and handover to the CMS, run on a schedule. In my setup, Perplexity researches, Claude Sonnet drafts, Gemini handles a secondary pass, and the result lands in WordPress as a draft for human review.
What does an AI-written article cost?
In the pipeline I run, the Claude cost per drafted article is roughly $0.30. That covers the model drafting, not the human work: the editorial brief, the fact-check, the final edit and the publish decision still cost what they cost, and they don’t get cheaper with volume.
What can AI automate in content production?
AI can take over the labor-intensive middle: research synthesis from current sources, first-draft structure against a fixed brief, on-brand phrasing at scale and staging the draft in the CMS. It doesn’t decide what’s worth writing about or whether a draft is good enough to publish.
What still needs a human editor?
The editorial brief, the fact-check pass, the final edit and the publish decision. In regulated markets, compliance language should also come from a reviewed, reusable template rather than from the model, and no compliance team accepts “the AI checked it” as a substitute for qualified human review.
Work with me
Want a second opinion on your content stack before you buy?
Let’s walk through what should and shouldn’t be automated in your case, and where your human review gate belongs.
