I make content about AI products and industry trends, and honestly, a big chunk of the job is repetitive. I collect recent info, look for an angle, draft a script, dump everything into a spreadsheet, and then turn whatever survives into a video.
Doing that by hand every day got old fast, because half my time went into moving the same information between tools before I could even start the actual creative part. Most of those steps felt predictable, so at some point I figured I might as well connect them with n8n. The goal was simple enough: collect data, generate topics, write a script, and get everything ready for video production. And the workflow worked — it's just that the content didn't always work with it.
What I built
The first part of the workflow pulls information from the sources I normally use for research, and that data goes to an AI model which picks out possible topics and drafts an angle for each one before everything lands in a spreadsheet. So instead of opening ten websites and copy-pasting into a doc, I just scroll through one sheet where each row already has the source material plus a possible direction. Once I pick a topic, the next stage drafts a script, and originally I wanted that draft to go straight into production — the dream pipeline looked like this: research → topic selection → script → video.
I never fully automated the final editing part, since I still used separate tools for visuals and tweaked things by hand, but I did expect the workflow to hand me material that was basically production-ready. And technically, it did. The scripts had intros, explanations, and conclusions, structured enough to become videos, and I was definitely producing faster than before. Then I watched the finished videos, and I didn't want to publish them.
The workflow produced content, but it didn't find the story
One of the videos explained how businesses can show up better in AI-generated answers. It was a relevant topic and the script covered the basics correctly，and being online doesn't mean an AI model actually understands or recommends you — so nothing in it was exactly wrong, it just wasn't very interesting. The script opened with the concept and then stacked examples on top, which made the whole thing feel like a short lesson. It told people why the topic mattered, but never gave them a real reason to keep watching.
Around the same time, I made another video, mostly by hand, where I asked a few AI models to recommend a milk tea franchise for someone with 300,000 RMB and zero food industry experience. Most models suggested big, familiar brands, but ChatGPT went somewhere different, and when I kept asking why, it turned out the models were interpreting the budget differently. Some only counted the franchise fee and equipment, while ChatGPT was also factoring in rent, deposits, initial inventory, and working capital.
That video didn't come from a clean automated pipeline; it started with one specific question and grew through an actual comparison, and the models disagreeing with each other was the story right there. Viewers got it instantly with same person, same question, different answers, and there were real stakes, because if you follow a recommendation without checking what the budget actually covers, you've made a very expensive mistake. The automated video looked cleaner, sure, but the milk tea one had the thing that matters more: a discovery.
I automated the wrong decision
At first I blamed the video production tool, so I changed the visuals, blew up the key text, and added more motion. It got slightly better, but it still felt empty, and it took me a while to realize the problem started way earlier in the pipeline. The model can summarize information and shape it into a reasonable script, but what it can't do reliably is tell which result has real tension in it — and when every topic goes through the same structure, a weak idea comes out looking almost as polished as a strong one.
That actually makes reviewing harder, because a nicely formatted script tricks you into thinking the idea is ready even when the story underneath is too general to carry anything. The workflow saved me from the blank page, but it also nudged me to keep going with ideas I would've dropped much earlier if I'd just looked at them properly. I had treated topic selection like another repetitive task, and it turns out it's the part that needs the most judgment.
Where n8n actually earns its keep
None of this means the workflow failed. The research stage got way easier, the spreadsheet gave me one place to compare topics, and I stopped losing time shuffling info between tools — plus, once I'd picked a promising idea, it was decent at producing a rough first draft. The trouble only started when I expected it to handle the whole editorial process. A model can tell that a topic is AI-related, and it can explain why that topic might matter to some audience, but neither of those things means there's a story in it worth telling.
For my content, the strong ideas usually start from something concrete: two models disagreeing, a company missing from every single recommendation, or an AI answer that flips when you add one small detail. Those situations create a question the viewer actually wants answered. Something broad like "why AI visibility matters" might be commercially relevant, but it still needs a real case behind it, otherwise you end up with a script full of correct explanations and zero investigation.
How I'm changing it
I'm not trying to automate research-to-finished-video anymore. n8n still collects and organizes information, and it can still throw out initial questions and rough scripts, but before anything goes into production, I run the experiment myself and decide whether the result is worth showing.
For example, instead of asking AI for a general script about manufacturers in AI search, I could ask several models to recommend Chinese steel suppliers for an overseas buyer. If every one of them skips a real factory, that's worth digging into — maybe the company has no clear English product info, no export records, nothing the models can verify. Now there's an actual question behind the video: why did a real supplier vanish from every answer? The workflow can collect the responses and organize the comparison, it just shouldn't get to decide the comparison is meaningful before I've even seen it.
I'm also simplifying the format. My team recently moved toward image-based videos with narration, plus text-based educational content, which are faster to produce and leave more time for research. A fancy animation doesn't rescue a weak idea, and a specific finding stays interesting even with simple visuals.
Automation should remove work, not judgment
When I first built this, I saw content production as a chain of steps, and I assumed that automating every step would make the whole thing more efficient. That was only half right. Some steps are just moving information around, while others are about deciding why the information matters — n8n is genuinely great at the first kind, and my mistake was assuming the second kind works the same way.
I'm still using the workflow, but its job is smaller and clearer now: it prepares the material so I can spend my time deciding what deserves to become content. The best automation didn't replace the creative process, it just bought me more time for it.
Have you ever built a workflow that ran perfectly and produced something you didn't actually want?
Top comments (1)
I think this is the biggest lesson in AI automation. We keep trying to automate judgment, but judgment is usually the product itself.
n8n is excellent at collecting, organizing, and preparing information. But deciding whether there’s actually a story, a contradiction, or something worth investigating still needs a human.
I’ve built workflows that were technically flawless yet consistently produced content nobody would remember. The pipeline wasn’t broken—the editorial decision was. Automation should reduce repetitive work so we can spend more time asking better questions, not fewer.