How we automated our content marketing pipeline with AI

Before our last blog post had a title, an AI agent ran through 40 possible topics, pulled search volume and difficulty from Ahrefs (marketing automation platform), checked the top 10 results on google, and scored each idea 1 to 10 on a mix of opportunity and intent match. Once our content manager wrote the outline, a second AI agent handed her a table of cross-link suggestions with anchor text and placement.

In the past we usually spent 2-3 hours per blog post on technical research, cross-linking and other mundane work.

This time is now saved thanks to the AI automation.

Here is the marketing content pipeline, where AI helps, and the parts it does not touch.

Pillar 1: SEO topic research

We keep a running list of topic ideas on our marketing board in Birdview. Some weeks it’s 10 ideas, some weeks it’s 40. Before building the automation, scoring them meant opening Ahrefs, running each query by hand, eyeballing the SERP, and making a subjective call about which one was worth writing. It took a couple of hours per round, and the content manager did it maybe once a month, because it was boring enough to put off.

Now it’s automated. She feeds the ideas into our content agent, and it calls the Ahrefs MCP to fetch search volume and difficulty, calls another API to pull the top 10 results for each query, and produces a ranked table in about 90 seconds. Each row has a score, a short rationale, and a link to the live SERP so she can sanity-check the top-10 composition herself.

What our content manager still does: override the scores when she knows something the agent doesn’t. A topic with low search volume and a high score might be perfect for us because we have a real story to tell, and the agent has no way to know that. A topic with a high score might be wrong for us because every top-10 result is a step-by-step tutorial and we don’t write tutorials. The agent produces the ranking. The content manager makes the call.

Pillar 2: Outline and first draft

This is the most overhyped part of marketing automation. Every LinkedIn thread will tell you an LLM can write your blog for you. It can, and the output reads like an LLM wrote it, which is the opposite of what thought leadership content needs.

Our outline step is more modest. Once a topic is picked, the content manager writes a one-paragraph framing and a section list. The agent then expands each section into a rough draft, drawing on context we feed it: the content strategy doc, the voice guide, and the five previous posts so the draft doesn’t contradict itself.

What comes back is maybe 60% usable as scaffolding. The structure is often fine. A lot of the specific examples are generic or wrong because the agent doesn’t know what we actually built last month. Every sentence that sounds like it came from a consulting slide gets deleted. Every tidy three-beat list gets rewritten as a normal paragraph. Every place where the agent tried to summarize what the post is about gets cut.

Pillar 3: Internal linking

This is the pillar where we were most surprised at how much time it saves. After each post is drafted, a special AI skill takes the Google Doc, reads the published blog, and returns a table of suggested internal links with anchor text, destination URL, and a one-line rationale for each. It considers whether the anchor text reads naturally in its surrounding sentence, not just whether the keyword matches.

Before automation, internal linking was an afterthought. The content manager would add two or three links back to earlier posts and call it done. The agent now suggests eight to twelve per post and she accepts most of them. SEO benefits from the density of internal links. Readers benefit from the cross-references. Both improvements came from a tool that runs in 30 seconds.

What the content marketer still does: reject links that make the paragraph worse to read. The agent optimizes for SEO signal. She optimizes for flow. When those conflict, flow wins. The “rationale” column makes rejection fast: she can see why the agent suggested a link and decide in three seconds whether it earns its place.

What the pipeline can’t do

It cannot tell you the health scoring system flagged three accounts our CS team thought were fine. Only someone inside that project knows that detail and can frame it with the right weight. The agent, given the same subject matter, will write “the system caught several risk signals,” and the sentence will be technically accurate and emotionally dead.

It cannot admit that one of our reps missed a demo because of a power outage and a colleague covered for him, and that the coaching agent ended up using the colleague’s call as a teaching moment for the person who missed it. That moment ended up in the Week 3 post because it happened. Nothing in an agent’s context window tells it that small real events like that are the ones that make a post feel human.

It cannot form opinions with actual edges. “This failure mode is worse than no automation at all” is the kind of sentence an agent will soften into “this approach has certain drawbacks.” The softening is the problem. A reader doesn’t learn anything from hedged opinions.

The economics

A topic research round used to be a couple of hours. Now it’s a prompt and 90 seconds. Internal linking on a 1,500-word post used to take the content manager 15 minutes of half-engaged work. That’s down to two minutes of accepting or rejecting the agent’s suggestions. First-draft outlines that used to eat an afternoon now come back in under a minute, ready to be torn apart.

Call it three hours saved per post, conservatively. Across a 40-post quarterly cadence, that’s 120 hours. About three weeks of one person’s time, redirected from grunt work to the part of writing that only a human can do.

The API cost is a few dollars per post across Claude, and other services. The savings are not dramatic in absolute terms, because a blog isn’t our biggest cost center. They are dramatic as a ratio. Spend a few dollars, save three hours, write better posts because the boring parts are out of the content manager’s hands and she has more energy for the 70% that matters.

Try this at your company

You don’t need our exact stack. The principles carry.

1. Automate the grunt research, not the voice. Topic scoring and internal linking are the safe places for agents. Drafting in your brand voice is not.
2. Feed your agents a written style guide. An explicit list of tells to avoid and signals to keep is the difference between an agent that helps you rewrite and an agent that produces something unusable.
3. Keep the content honest. If the finished draft doesn’t feel like a person wrote it, the pipeline failed regardless of how much time it saved.
4. Put the content manager in the last-mile loop. An agent-suggested internal link that damages flow is a net loss. Humans make the final call on every change that touches the reader’s experience.

What’s next

This is Part 6 of “Becoming AI-Native,” a weekly series from the Birdview PSA team on our AI transformation journey. Follow along here on Birdview’s blog, on Vadim’s LinkedIn