An AI writing pipeline that learns from the editor

Why does most AI writing fail?

Because it is measured against the wrong thing. A prompt produces text that is roughly right, and roughly right still needs a full rewrite, so the editor's time stays where it was and the model bill is added on top.

The company I built this for had already tried the obvious version and stopped. What they had was not a writing problem, it was a feedback problem: nothing in the system noticed what the editor kept changing.

How does the pipeline work?

Five steps, and the last one is the part that matters.

• The brief goes in: topic, audience, the facts to use and the ones to leave out.
• The model writes a draft against that brief, with the house rules attached.
• The editor changes the draft and publishes their version, as they always did.
• The system compares the two versions and records what changed: structure, tone, the words that always get cut, the claims that always get softened.
• Those differences become part of the next brief, so each draft starts closer to what this editor wants.

What changed the numbers

The drafts got shorter to edit. Not because the model got cleverer, but because the instruction it received each time described this editor, this audience and the mistakes it had already made.

The cost per article went from about $900 to about $100. The saving is in the editor's time, which is the expensive part, and it keeps growing while the pipeline keeps running.

What still needs a person

Judgement. The editor decides what is true, what is worth saying and what the company is prepared to put its name to. Nothing in the pipeline should be allowed to make those calls.

That is why every article still goes through the same approval it always did. The pipeline only removed the mechanical rewriting around it.

Where else does this pattern work?

Anywhere one expert keeps correcting the same things in output from a machine.

• Support replies that a senior person rewrites before they go out.
• Product descriptions corrected for tone across a catalogue.
• Reports and summaries checked against a house format.
• Proposals and quotes that follow rules nobody wrote down.

When AI writing is the wrong idea

I turn down more of these than I build. Four signs that it is the wrong idea:

• You publish a few times a year, so the pipeline never sees enough corrections to learn from.
• Nobody edits the output before it goes out. An unedited model draft is a way to publish mistakes at scale.
• The facts would come from the model instead of from your work.
• Every piece is a different shape, so the same corrections never repeat.

What to remember

• Measure the editor's time, not the word count.
• Compare the draft with the final version and feed the difference back.
• Keep a person approving everything that gets published.
• Only build this if you publish often enough for it to learn.