In this blog post I'm going to walk you through the exact AI content marketing workflow I use in my own business, the 80/20 split between human work and AI work that triples output without triggering Google penalties. The version with the line drawn between safe and dangerous use.
Most "AI for content marketing" content in 2026 is one of two things. Either a vendor pitch promising 10x content production (which always ends in algorithm penalties). Or a vague exhortation to "use AI as a tool" without saying how.
This article is the working version. The exact split. The exact workflow steps. The exact deliverables you ship at each phase. And the explicit guardrails that kept me from getting penalised again after my 35% traffic crash in mid-2025.
I've been a marketing consultant for twenty-one years. I went all in on AI in 2024. I currently produce 3x more content than I did in 2024 at the same quality. None of it is fully AI-generated. None of it has triggered a Google penalty since my recovery in late 2025. The workflow below is what made the difference.
By the end of this blog you'll know which 80% of content work should stay human, which 20% should be AI, the exact workflow steps, and the kill-switches that prevent slow-creep into dangerous AI use.
TL;DR
The 80/20 split:
Human (80%):
- Topic selection
- The opinion / contrarian take
- Specific examples from real work
- Editorial decisions
- Voice and tone
- Final review
AI (20%):
- Research synthesis
- Structure validation
- FAQ extraction
- SEO metadata
- Schema generation
- Internal link suggestions
- Distribution scheduling
Get the split wrong (or worse, drift past 20% AI over time) and you get penalised.
Why the 80/20 split
Every successful AI content workflow I've audited or built has the same shape. AI is a multiplier on judgement. AI is a substitute for judgement only when the content is operational (meta data, schema, scheduling), never when the content is the visible product.
Three reasons:
1. Google penalises pattern, not provenance. Google's helpful content updates don't detect AI use directly. They detect the patterns that mass-AI-content produces: generic openings, predictable structures, lack of specific examples, absent first-person evidence, low engagement. Human-led writing avoids the pattern.
2. AI is bad at being interesting. AI defaults to averaging. The opinion that makes content worth reading is, by definition, not average. Humans hold opinions. AI predicts the median response. The median response doesn't earn citations.
3. Trust signals come from humans. Specific examples, named clients, real numbers, dates, locations, these are the trust markers Google specifically rewards in helpful content. AI generates plausible-but-fictional versions of these. Humans cite real ones.
The 11-step workflow
This is what I run every time I publish a blog post.
Step 1 (HUMAN, 15 minutes): Topic selection
Pick a topic based on:
- A real client question from this week
- A specific gap in your existing content
- A search query you've seen and have a strong take on
Skip AI here. The topic decision is what makes the difference between "another AI marketing post" and "a piece worth reading."
Step 2 (AI, 5 minutes): Research synthesis
Feed the topic to Claude/ChatGPT with this structure:
- "I'm writing about [topic]"
- "Find: the 3-5 most cited positions on this, the typical counter-arguments, the data points referenced most often"
- "Constraint: cite source URLs where possible. Don't invent statistics."
You're not asking AI to write. You're asking it to compress 3-4 hours of research into a 5-minute briefing.
Step 3 (HUMAN, 10 minutes): Opinion crystallisation
Based on the research, decide YOUR position:
- What do you agree with the consensus on?
- What do you disagree with?
- What's the contrarian angle you can defend?
- What do you know from direct experience that nobody else writing about this topic does?
Write your position in 2-3 sentences. This is the anchor for the whole piece.
Step 4 (AI, 10 minutes): Structure validation
Now ask AI: "Given my position [paste], propose 6-8 H2 sections that build the argument. For each, one sentence on what it covers."
You'll get a reasonable outline. Adjust it manually, usually 2-3 changes. AI's instinct toward "comprehensive coverage" needs to be reined back to "argued position."
Step 5 (HUMAN, 60-90 minutes): Write the opening, the contrarian take, the close
The opening hooks. The contrarian take is the section that makes the piece worth reading. The close calls the reader to action.
These three sections are 100% human. They're the parts a careful reader reads. They're also the parts that get screenshotted, quoted, and shared.
Step 6 (HUMAN + AI, 60-90 minutes): Write the body
For each H2:
- Write the first paragraph yourself (sets the voice)
- Optionally use AI to draft the supporting paragraphs
- Edit AI drafts to add: specific examples, real numbers, actual client work, dates, named tools
The edits are what make the difference. AI's first draft will read smoothly but be generic. Your edits add the substance.
Step 7 (HUMAN, 30 minutes): Specific example injection
For every claim in the article, ask: do I have a real example? If yes, add it. If no, either find one (preferred) or remove the claim.
This is where AI-led content fails. It makes claims without evidence. Replacing claims with examples is the single highest-use editorial pass.
Step 8 (AI, 10 minutes): FAQ extraction + structuring
Ask AI: "Based on this article, generate 6-8 FAQs the reader is likely to have, with answers under 100 words each. Constraint: don't invent information not in the article."
Review and edit. This becomes the FAQ section AND the FAQPage schema source.
Step 9 (AI, 5 minutes): SEO metadata
Generate Yoast title (under 60 chars), meta description (under 155 chars), URL slug, focus keyphrase, image alt text.
Human-review each. AI's defaults are often too generic or too cute. A small adjustment turns them into something that earns clicks.
Step 10 (AI, 5 minutes): Schema generation
For most blog posts, you want FAQPage schema. For procedural content, add HowTo. For checklist content, add ItemList. For voice-search readiness, add Speakable.
AI generates the JSON-LD blocks. Validate them via Google's Rich Results Test.
Step 11 (HUMAN, 15 minutes): Editorial pass + publish
Read the full piece end-to-end. Check:
- Every paragraph either teaches something specific OR moves the argument forward
- No generic AI tells ("In today's fast-paced world...", "use", "Transformative")
- British or American English consistent throughout
- No em-dashes if your voice avoids them
- Specific numbers, dates, and examples where claims are made
If any paragraph fails the test, cut or rewrite. Then publish.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Total time: 4-5 hours per substantial blog post. Down from 8-12 hours when I wrote everything manually.
The kill-switches
Drift creep is real. The temptation to let AI do more grows. These kill-switches stop the drift.
Kill-switch 1: The 80/20 ratio check
At the end of every piece, calculate roughly: what percentage of the words on the page would I have written without AI? If under 60%, scrap and rewrite.
Kill-switch 2: The first-paragraph test
Read your first paragraph. Could a generic AI have written it? If yes, rewrite. The first paragraph is the audition.
Kill-switch 3: The example check
Count specific examples per 500 words. Target: 2-3 per 500 words. Below 1 per 500 words, the piece reads as theoretical and fails Google's helpful content signals.
Kill-switch 4: The opinion test
Read the piece. Can you describe the author's position in one sentence? If no, the piece doesn't have one, and pieces without positions don't earn citations.
Kill-switch 5: The bounce-rate review
After publishing, monitor the bounce rate weekly. Bounce rates climbing on AI-assisted pieces? Pull back the AI percentage on subsequent pieces.
What I no longer do
After my recovery from the 2025 traffic drop, I removed these from my workflow permanently:
Mass content production. No more "10 posts a week from AI." I publish 1-3 substantial pieces per week. Quality compounded faster than quantity.
AI-led drafts I edit. I used to ask AI to write the full piece, then I'd edit. The editing took longer than writing from scratch AND produced worse results. Now I write the structural skeleton and AI fills supporting paragraphs.
Programmatic SEO at scale. No more "1,000 location pages." Google catches this. Even at lower volumes (50-100 pages), the pattern flags.
AI for outbound emails. Personalised AI cold outreach destroys reply rates. AI helps with research; humans write outreach.
Auto-generated social posts from blog content. Reduces social engagement because the posts look templated. I write originals for each platform.
What works for high-volume content businesses
Some businesses need higher volume than 1-3 pieces per week. The patterns that work:
Multiple writers. Hire 2-4 writers, each producing 2-3 pieces per week, using the 80/20 workflow individually. Total: 8-12 pieces per week, all human-led with AI support.
Topic clusters around expertise. Don't try to cover everything. Cover 3-5 deep topic clusters where you have real authority.
AI for the long tail. For pages targeting low-volume long-tail queries (under 500 monthly searches), AI assistance can be higher, but still human-reviewed for accuracy.
Strict quality bar regardless of volume. Volume is fine if quality holds. Volume at lower quality is the trap.
The vendor traps to avoid
Three "AI content marketing" vendor pitches I tell clients to avoid:
Trap 1: "AI content at scale" platforms
Marketed as "10x your content output." Reality: they produce content that gets penalised. Some platforms now market themselves as "Google-friendly AI content", still produces content that pattern-matches as AI-generated to humans, which kills engagement metrics, which kills rankings.
Trap 2: AI SEO tools that auto-publish
The promise: AI finds keywords, writes posts, publishes them. The reality: every site I've audited that uses these has either taken a penalty or is heading toward one.
Trap 3: "Humanise AI content" services
The premise: AI writes, then a service edits to make it sound human. Reality: the editorial effort to humanise AI output is more than the effort to write from scratch with AI assistance. The "humanised" output is detectable.
Frequently asked questions
Can I use ChatGPT or Claude for this workflow? Either works. I use Claude for long-form drafting (better at extended reasoning) and ChatGPT for quick tasks (better at one-shot generation). Both are fine.
What about Google Gemini or other newer models? Gemini's content output is competitive. Use whatever you prefer. The workflow matters more than the model.
How long does it take to get comfortable with this workflow? 4-6 weeks. The first 5-10 pieces feel awkward. After that the 80/20 split becomes intuitive.
Can I delegate this to a junior marketer? Yes, once you've documented the kill-switches. Without explicit kill-switches, junior marketers drift toward higher AI percentages over time.
What about images? Should I AI-generate them too? For featured images, AI works fine (with brand consistency rules). For images embedded in content, real screenshots and real diagrams perform better than AI illustrations.
Want help setting up this workflow?
If the conclusion is "your content needs strategic work before AI helps," that's the conclusion. You walk away with the diagnosis and no further pitch.
I'm Lilach Bullock. I've been a marketing consultant for twenty-one years. I went all in on AI in 2024. I work with founders and marketing leaders who want AI to move their numbers, not just their tool stack.
Related: running a guest posting campaign
This builds on my main AI marketing guide, my main guide on the topic.
