# Building an AI content pipeline that cut a marketing agency's cost per blog post > A marketing agency used contract writers to produce its clients' blog posts, one post at a time. I proposed using AI for the first draft, with a person reviewing and approving each post. Then I built the system to do the research, suggest internal links, handle the SEO work and write the draft. **In brief:** A marketing agency used contract writers to produce its clients' blog posts, at roughly $250 a post by our own estimate. In early 2025 I proposed using AI for the first draft, with a person still reviewing and approving every post. I built each part separately, covering keyword research, analysis of the pages already ranking, internal-link suggestions from the client's own site, and sourced statistics and examples. The system also accounted for the client's brand voice, added images and SEO metadata, edited the draft and handled publishing to the client's CMS. I combined those parts into one workflow that takes a topic and produces a draft for review. By May 2026 it was producing more than 100 drafts a week, at roughly $25 a post including the human review. **Marketing agency · AI content pipeline · Python · LangGraph · Next.js** ## Before → After Before: Every client blog post started from scratch with a contract writer, at roughly $250 each. - **100+**: drafts a week by May 2026 - **$25**: per post, down from about $250 (rough estimate) - **6**: CMS platforms it publishes to - **Reviewed**: by a person before publication ## The stakes The agency wrote blog posts for its clients, and a contract writer started every post from scratch. By our napkin math that came to about $250 a post. The number of posts the agency could deliver depended on how many writers it had. In early 2025 I proposed having AI write the first draft, which a person would then review and edit before approving it. That would let the writers spend their time judging the draft instead of typing it. ## Why a chatbot draft isn't enough, and what a content pipeline needs **Most of the work on a good blog post is the research a writer does before writing it.** Giving a chatbot a topic gets you a draft that reads fine but does little. It doesn't know which keywords are worth targeting or what the pages already ranking for them cover. It also doesn't know which of the client's pages to link to, or how the brand sounds. Without a check, it will invent statistics too. I needed the system to do that work for every post before it started the draft. ## What I did I built it in two layers, starting with separate pieces I could test on their own. These covered keyword research, analysis of the pages already ranking, and internal-link suggestions from the client's own site. For statistics and examples, I added research and a check that each one matched its source. Other pieces handled the client's brand and voice profile, images, titles, meta descriptions and structured data. I also built an editing pass to remove the usual signs of AI writing, and the connections for publishing to the client's CMS. Then I combined the pieces into one workflow. Someone gives it a topic, and it does the research and writes a draft. It optimizes the post for search and saves it in the client's CMS as a draft for review. Nothing publishes on its own. Keeping the pieces separate also meant I could use them in other workflows later, including refreshing older posts and adding internal links to pages that needed them. ## The result By the time I left in May 2026, the pipeline was producing more than 100 drafts a week. By the same napkin math, the cost per post had gone from about $250 with a contract writer to about $25 with an AI draft and a person reviewing it. The agency used it mainly for its own client work, and a person still reviewed every draft before it went out. ## What I learned I would talk to the people using it earlier and more often. I built much of it on my own before getting their feedback, and some pieces would have come out smaller and better if the editors had shaped them sooner. Next time I would put a very small version in front of real users first, then build from what they actually do with it. **What this proves:** The pipeline could take over the first draft because it did the research a good writer would do, with a person still signing off. That made the drafts cheap enough to produce at a larger scale and good enough for the agency to use. Building an AI content pipeline? Let's talk it through. Book a call: https://calendly.com/nerijus-masikonis/intro-call ## More case studies - [Stabilizing a WordPress news network that kept crashing under match-day traffic](https://masikonis.lt/case-studies/sports-news-network-migration/): Down on every Denmark match, and the owner had no direct access to his own servers. - [Recovering a publisher whose Google Discover traffic dropped to zero](https://masikonis.lt/case-studies/google-discover-recovery/): Dropped from Google Discover after a core update, with its main traffic channel sitting at zero. - [Restoring a publisher's lost archive after a CMS migration](https://masikonis.lt/case-studies/lost-archive-restoration/): Most of the archive missing after a CMS migration, with Google still requesting the vanished URLs.