Cheap production creates expensive momentum
The old process filtered ideas through cost. Research, writing, editing, design, and a calendar slot had to be justified before production began. That friction rejected good ideas along with bad ones, but it forced the team to choose.
AI removes much of the useful friction. Every keyword, sales question, competitor gap, and executive thought can become a plausible article. Drafting is cheaper. Attention, evidence, and reader patience remain inconveniently finite.
Once a draft exists, the team begins to defend the fact that it exists. Someone requested it. Someone added examples. The article has a URL waiting. Rejecting it now feels like throwing away work, even when creating it cost less than the meeting where the calendar was approved.
This is sunk-cost psychology with excellent formatting.
The first edit is whether there is an idea
Most editors enter too late and receive sentence-level authority. Fix the voice. Check the grammar. Shorten the opening. Make the CTA match the campaign.
That job assumes the article has earned publication. In an AI system, that assumption is the risk.
The first editorial pass should ask whether the draft contains an idea rather than a topic. Imagine a polished article comparing five project-management tools. The tables are accurate. The prose is clean. The conclusion recommends choosing based on team size, integrations, and budget, which is the conclusion every buyer knew before arriving.
The draft may be complete. It hasn't changed a decision.
An editor needs permission to stop it before line editing. Maybe the useful piece is narrower: which tool becomes a maintenance burden for a small operations team? Maybe the product data cannot answer that yet. Maybe the category needs no ninth competent comparison from us.
We have a whole article on why the topic isn't the point. A production system that skips that distinction merely writes the expected article faster.
A score can hide the fatal flaw
Rubrics are attractive because numbers make judgment look portable. Relevance: eight. Evidence: seven. Voice: nine. Links: five out of five. Total: publish.
The arithmetic lets a fatal weakness average out. An article with no defensible claim can pass because its metadata is complete and the sentences are pleasant.
Use a kill memo before the score. It needs five answers:
What is the claim in one sentence?
Which evidence would disappear from the category if we did not publish this?
What reader decision changes?
What could a reasonable expert reject?
Why does this belong to us rather than any competent competitor?
These aren't five equal points. A blank answer can matter more than four strong ones. The memo exists to expose the weakness a total score would politely absorb.
Give the draft four possible endings
Binary workflows create bad incentives. If the only choices are not ready and publish, every edit pushes toward publication.
Use four outcomes instead:
Publish means the idea has earned the final edit. Revise means the idea is sound but the frame, evidence, or execution failed. Strip for parts means an example or dataset belongs elsewhere. Kill means the topic produced nothing the company needs to own.
The last outcome is not a pipeline error. In a system that reports success too easily, stopping looks broken because movement is the only thing being measured. An editorial system should count a well-reasoned rejection as evidence that someone used the filter.
That decision needs a note, not a funeral. No distinct claim is enough. So is evidence cannot support headline or buyer question already answered better elsewhere. The note prevents the same draft from returning next week with a fresher adjective.
Approval expires when the argument changes
Cheap revision creates one more way around the editor. A draft passes review, changes materially, and keeps the old approval.
The editorial decision has to belong to the artifact that was reviewed. Change the headline, central claim, evidence, or conclusion and the kill memo expires. The next stage stays closed until someone approves the new version.
This is not paperwork for its own sake. A green check attached to work nobody read is not a control. It is set decoration.
Optionality is not inventory
Generating 20 ideas can be useful. The system can try opposing claims, narrower readers, stranger structures, and evidence the team might not have commissioned under the old economics.
The value arrives only when most of those possibilities are allowed to lose.
An editor who rejects 17 drafts hasn't slowed production. The editor has converted cheap generation into selection. Without that refusal, the system didn't create optionality. It created inventory, and inventory has a way of demanding publication simply because somebody can see it.
AI can make almost anything look ready. The editor's job is to notice when ready is beside the point.