Building an AI Content Pipeline From Scratch
I spent the last three months building a complete AI content pipeline for a B2B fintech blog. Not a hypothetical one. Not a "here's what you could do" thought experiment. An actual pipeline that takes us from zero to published in about four hours per post, where the old process took two to three days.
The total monthly cost for the tools involved: roughly $380. The time savings: about 60% compared to our previous workflow. And the quality? Honestly, better on average than what we were producing before, because the AI handles the parts humans are bad at (consistency, formatting, SEO optimization) while humans handle the parts AI is bad at (original thinking, industry nuance, not being boring).
Here's the full pipeline, stage by stage, with the exact tools, costs, and time estimates I've tracked over 40+ posts.
Stage 1: Research and Ideation
This is the stage most people skip or do badly. They open ChatGPT, type "give me 10 blog post ideas about fintech," and get garbage. Research and ideation is where your content either earns the right to exist or doesn't.
I use two tools here, and they serve very different purposes.
Gemini Advanced
Gemini's strength is its connection to Google's search index and its ability to synthesize information from multiple sources. I use it to map out what already exists on a topic, find angles that haven't been covered, and identify data points worth referencing. I'll give it prompts like "analyze the top 10 ranking articles for [keyword] and identify gaps in coverage." It's not perfect, but it saves me an hour of manual SERP analysis per post.
Claude Pro
Claude is better than any other model I've tested at taking a vague idea and turning it into a structured argument. I use it for the creative side of ideation: "Here's a topic and three existing angles. What's a fourth angle that would be genuinely useful to a CFO evaluating this?" Claude's reasoning is sharper than GPT-4o for this kind of work, and it pushes back on weak ideas instead of just validating everything you say.
The output of this stage is a one-page research brief: the angle we're taking, 3-5 key points, the target audience, and a list of data points or studies to reference. This document is the single most important artifact in the pipeline. If the brief is weak, everything downstream is weak.
Stage 2: Content Briefs and SEO Planning
Surfer SEO
Once I have a topic and angle, I run it through Surfer to get the SEO scaffolding. Surfer analyzes what's ranking, tells you the keywords to hit, suggests header structures, and gives you a target word count. Is it perfect? No. Does it catch things I'd miss? Every single time.
I pull the keyword clusters and suggested headers from Surfer, then manually merge them with my research brief. The AI doesn't do this merge step. I've tried automating it and the results were consistently worse than spending 10 minutes doing it myself. Some things are still faster by hand.
The output here is a detailed content brief: target keyword, secondary keywords, header structure (H2s and H3s), word count target, and internal linking opportunities. This brief goes into the next stage as the writing prompt.
Stage 3: First Draft
Here's where I'm going to say something that might be controversial: I use different AI tools for different types of content, and the difference in output quality is significant enough to justify the hassle.
Claude Pro (long-form)
For anything over 1,000 words that needs to sound like a human with opinions wrote it, Claude is the clear winner. I feed it the content brief, a few examples of our brand voice, and specific instructions about what not to do (no "in today's fast-paced world," no "let's dive in," no filler paragraphs). The output typically needs 30-40% editing, which is good. If it needs less than that, the content is probably too generic.
Jasper
Jasper is better than Claude for short-form marketing content. Controversial take, I know. But its templates are specifically trained on marketing patterns, and for things like ad variations, email subject lines, and social media posts, it produces more usable first drafts. I don't use it for anything over 300 words. That's not its strength.
My actual process: I paste the content brief into Claude, ask for a first draft, then go through it section by section. I rewrite the intro completely (AI intros are almost always terrible), keep the middle sections that work, and rewrite the conclusion. Total hands-on time: about 45 minutes of active editing on top of the 15 minutes of prompting and setup.
Stage 4: Editing and Optimization
This is the stage that separates "we use AI for content" from "our content is obviously AI-generated." Two tools, used in sequence.
Surfer SEO Content Editor
I paste the edited draft back into Surfer's content editor and check the SEO score. Typically the first draft scores around 60-70 out of 100. The optimization pass — adding missing keywords, adjusting header distribution, tweaking word count — brings it to 80-90. I never aim for 100. A perfect SEO score usually means the content reads like it was written by a keyword-stuffing robot from 2015.
Grammarly Business
Grammarly catches things that AI-generated text is particularly prone to: passive voice overuse, unnecessarily complex sentences, and subtle tense inconsistencies. I also use it to check the readability score. Our target is grade 8-9 reading level. AI drafts tend to come in at grade 11-12, which means they need to be simplified. Grammarly's suggestions aren't always right, but they flag the right sentences to look at.
I also run every post through a final "AI detection" pass. Not because I care about AI detection per se, but because the detectors are actually decent at flagging passages that sound generic and robotic. If a paragraph triggers the detector, it usually needs to be rewritten regardless.
Stage 5: Images and Visual Assets
I'll be blunt: AI image generation for marketing content is still more hassle than it should be. But it's gotten good enough that I've eliminated stock photo subscriptions entirely.
Midjourney
Midjourney v6 produces the most consistently usable marketing images. I use it for hero images and conceptual illustrations. The trick is being extremely specific in your prompts and always generating at least 4 variations. My success rate (image is usable without significant editing) is about 60%. That sounds low, but each generation takes 30 seconds, so the iteration cycle is fast.
Canva Pro
Canva's AI features have gotten surprisingly good. I use it to resize Midjourney outputs for different platforms, add text overlays, and create simple infographics from data in the post. Their "Magic Resize" alone saves me 20 minutes per post in creating social media variants. And their background remover is better than most dedicated tools I've tested.
For technical posts, I sometimes use DALL-E 3 through ChatGPT Plus instead of Midjourney. DALL-E is better at diagrams and technical illustrations, while Midjourney is better at aesthetic, editorial-style images. Pick the tool that matches the visual style you need.
Stage 6: Publishing and CMS Integration
This is the least interesting stage but one where small automations add up. We use WordPress, and I've set up a workflow in Make (formerly Integromat) that takes a Google Doc (our editing environment), formats it into WordPress-ready HTML, uploads images to the media library, sets featured images, applies the right categories and tags, and creates a draft post. The whole thing runs in about 90 seconds.
Could I publish directly from the AI tools? Sure. But having the Google Doc as an intermediate step means there's always a human-readable version of every post that anyone on the team can review without needing access to WordPress or any of the AI tools.
Stage 7: Distribution
Buffer
Buffer handles the social distribution. For each blog post, I generate 5-8 social media variants using Jasper (a mix of LinkedIn posts, tweets, and short-form snippets), schedule them across two weeks in Buffer, and let it handle the posting. Buffer's AI assistant can also generate social posts directly, but I've found Jasper produces better variants when given the full blog post as context.
Klaviyo
Every blog post gets summarized into a newsletter segment. I use Claude to generate the email version — shorter, punchier, with a clear CTA to read the full post. Klaviyo's AI features help with subject line optimization and send time optimization. Our open rates improved 15% after we started using their send time AI, which is the most clearly measurable ROI of any AI feature in this entire pipeline.
The Full Cost Breakdown
| Tool | Monthly Cost | Stage |
|---|---|---|
| Gemini Advanced | $20 | Research |
| Claude Pro | $20 | Research + Writing |
| Surfer SEO | $89 | Briefs + Editing |
| Jasper | $49 | Short-form + Social |
| Grammarly Business | $25 | Editing |
| Midjourney | $30 | Images |
| Canva Pro | $13 | Graphics |
| Buffer | $36 | Social distribution |
| Klaviyo | $25 | |
| Total | $307/mo |
Call it $380/mo if you count the Make subscription ($29/mo for the automation workflows) and occasional overages on Midjourney. That's the real number. Not cheap, but compared to what agencies charge for the same output volume, it's a fraction.
Time Per Post: Before and After
| Stage | Before AI | With AI Pipeline |
|---|---|---|
| Research & Ideation | 2-3 hours | 45 min |
| Content Brief | 45 min | 20 min |
| First Draft | 4-6 hours | 60 min |
| Editing | 1-2 hours | 30 min |
| Images | 1 hour | 30 min |
| Publishing | 30 min | 15 min |
| Distribution | 1 hour | 15 min |
| Total | 10-14 hours | ~4 hours |
What I'd Cut If Budget Were Tight
If you can only spend $100/mo, here's what I'd keep:
- Claude Pro ($20/mo) — handles research, writing, and email drafting. The most versatile tool in the pipeline.
- Surfer SEO ($89/mo) — the SEO data is irreplaceable. You can't prompt your way to keyword research.
That's it. $109/mo gets you 70% of the value. Everything else is optimization. Jasper is nice but Claude can do short-form too. Midjourney is nice but Canva's free tier plus DALL-E in ChatGPT gets you close. Buffer and Klaviyo are nice but you can post manually and send emails from Mailchimp's free plan.
The bottom line: An AI content pipeline isn't a magic button. It's a system that saves time on the mechanical parts of content creation so you can spend more time on the parts that actually matter: having something worth saying. The tools handle the how. You still have to figure out the what and the why.
If you're building something similar and want to compare notes, I'm always happy to talk shop. Just don't ask me to recommend tools I haven't personally used — the list is shorter than you'd think.