I opened the client's 'Drafts' folder. There were 47 posts in it
A story about one of the most typical deaths of a content factory — in week three. About what a 'working automation' that produces nothing looks like.
There was a call. The client — an online school, 40 people on the team, living off regular Telegram and Instagram content. The CMO called and said: "our content factory broke. Help me fix it."
I sat down to look. An hour in I opened their Notion folder called "Drafts." There were 47 posts in it. Of those, older than three weeks — 31. Older than two months — 14. One, the oldest, was marked "ready to publish" — a year and a half ago.
That post was about a trend which, by the time I was looking, was no longer a trend. Presumably it had been a trend when written. Maybe. Nobody remembered.
— So, — says the CMO. — We do publish, though. About one post a week.
— Does your SMM person write one a week?
— No. AI writes about fifteen a week.
Right. So fifteen a week were being written. One a week — published. The other fourteen — settled into "Drafts," where they quietly died of old age.
That was their "content factory." The AI worked. The factory didn't.
I've seen this picture at eight or nine clients since. Different scale (one a week vs five a day), but the same structure: generation is automated, everything else isn't. The result is a bottleneck in the manual part, and as soon as that shows up, the whole pipeline stops. Drafts pile up, pain piles up, and in a month or two the team stuffs the thing in a corner so they don't have to look.
The typical response: "well, the AI must write badly." Wrong. The AI writes well enough. The problem is that nobody called "content factory" all the other stuff needed for normal operation. And the other stuff is roughly 80% of the work.
A factory isn't a machine. It's a workshop. A workshop has ten machines, a raw-material warehouse, an assembly floor, QC, transport, and a person who mops the floor once a week. If you only have a machine, that's a stall, not a factory. A stall scales linearly in effort, not by two orders.
Here's what usually is and isn't automated in an "AI content factory."
Automated: draft generation. Maybe headline generation. Sometimes — visual generation. That's it.
Not automated:
Idea sourcing. Topic research. Editorial judgement ("is this actually what we want to say?"). Fact-checking. Adaptation to platform format. Approvals — with legal, for instance, if the product is in a regulated space. Publication scheduling. Actual publication. Response collection. Analysis of what worked. Feeding analysis back into the next cycle.
Add it up — ten steps. Automated — one or two. And I haven't even added the minor ones: "write the post in three tones for three channels," "rework to match our visual standards," "comply with our internal terminology rules." Three more steps.
Three things I see at every client like this.
First — "I have lots of ideas, I'll process them later." No. You won't. Ideas go stale. They're tied to the moment — a trend, a news cycle, a season, an audience mood. An idea that sat for three weeks is no longer your idea. It's something that resembles it.
The idea funnel must be wired to a publishing schedule. Not "we collect ideas, then decide." Instead: idea → topic → draft → publication, with dates locked to each other. An idea without a date is a museum.
Second — "AI writes, we'll fix it up." This works exactly until "fix it up" starts to equal "rewrite." That moment arrives faster than you think, and it arrives because you framed the prompt as "write a post about X in our brand voice" without specifying what "our brand voice" is and how it differs from "the voice of any similar brand."
I have a hard rule: if editing takes more than 30% of the time it would take to write from scratch, the prompt is broken. You fix it by changing the prompt, not by adding "another editorial pass."
Third — "we use five tools, not one system." Notion for ideas. Google Docs for drafts. Telegram for approvals. Linear for tickets. A spreadsheet for analytics. Each tool — good, on its own.
Together it's five extra clicks per unit of content. Five times twenty publications a month — a hundred clicks. A hundred clicks is an hour of your attention every day.
In a content factory a key metric is friction per unit. If it's high, the team gets tired. A tired team loses quality. Lost quality means content stops doing its job. Back to the "we publish but we get no leads" place.
What I did with the client and their 47 drafts.
First — shut down the AI generation. Fully. Two weeks. Because that's the source of the blockage. You can only clear a backlog by stopping adding to it.
Second — cleared drafts. Not "edited 47." Deleted 31. Because 31 had lost relevance. Left 16. Of 16 — 10 published inside two weeks. 6 — kept as a buffer, but with concrete dates.
Third — rewrote the process. From idea to publication, one board (they ended up on Airtable, but anything works). One record = one publication = all stages on one row. Idea, author, draft, editor, approval, date, channel, results. No "check Notion, then Docs, then Telegram, then the spreadsheet." One screen.
Fourth — the prompt became part of the system. Previously each writer called ChatGPT with "write a post about X in our brand voice" their own way. Now the prompt is an artefact, living in Airtable next to templates. Update it — everyone writes in the new way. A small-seeming detail, but it's what gives the team consistency, without which a content factory can't be a factory.
Fifth — manual-work budget per unit. No more than twenty minutes per post from draft to publication. More than that, investigate why. Budget includes editing. If editing eats fifteen minutes — that's five for everything else, time to change the pipeline.
Two months later, instead of "one post a week," this client was publishing four a week. Not fifteen, the way the old factory "produced." Four. But four published and doing their job, not fifteen rotting in Notion.
That's the key metric no one tracks: ratio of published to generated. Before the intervention theirs was 1 to 15. Two months in — 4 to 5. Not scale. Recovery.
A content factory isn't "AI that generates posts." It's a workshop where AI is one machine. If you bought the machine and didn't build the rest, you'll end up where my client did: 47 drafts in a folder and a CMO who doesn't understand why volume goes up but reach doesn't.
Machines are cheap. A workshop is expensive. But only the workshop produces. A machine on its own produces parts, which nobody then assembles.