Using AI in Content Production Without Publishing Filler
Why the failure is always the same shape
A model can draft faster than any writer, and the capacity it adds is entirely on the production side. Review capacity does not change. So the predictable outcome of adopting one without deciding anything else is that output rises tenfold, the review queue grows, and the queue is eventually resolved by reviewing less.
Six months later there are several hundred pages nobody has read, some containing confident invented numbers, and the cleanup costs more than the pages were ever worth. Nothing about that sequence is a tooling problem. It is a capacity decision that was never made.
Where the tools help, where they do not, and what has to stay human.
Where they help
Research summaries, outline options, rewriting for clarity, finding gaps against competitors, and first drafts of sections you already know well.
They are fast at structure and competent at tidying. That is genuinely useful.
Where they fail
They do not have your experience, your clients or your data. They hallucinate confidently. They produce prose that reads exactly like every competitor using the same tool.
Published unedited, that is the definition of content with nothing only you could add.
A workable process
Research with sources captured. Human outline. Draft with assistance. Human edit that adds experience and examples. Fact check every claim. Named author. Review date.
The time saved goes into thinking, not into publishing more of the same.
Where a model helps and where it does not
The reliable uses are the ones where the model is transforming something you already have: turning research notes into a first draft, summarising a transcript, converting a table into prose, finding the inconsistencies across forty pages. In all of those the facts come from you and the model does the typing, which is both safe and genuinely faster.
The unreliable use is asking it for the facts. A model will produce a confident statistic with a plausible source that does not exist, and it will do it most often on exactly the subjects where a specific number would be most persuasive. Any workflow that does not have a human verifying every number and every citation against a primary source will publish fiction, and the cost of that lands on the business rather than on the tool.
A workflow that holds up
- Gather sources and facts first, by hand, with links.
- Give the model the facts and ask for structure and draft prose, not for research.
- Verify every number, date, name and citation against the primary source.
- Edit for voice, because an unedited draft reads like every other unedited draft.
- Have a named person review and take responsibility, and say who on the page.
Scaling before the review step exists
The failure is always the same shape: production capacity rises tenfold, review capacity stays where it was, and the backlog is resolved by reviewing less. Six months later there are four hundred pages nobody has read, several of which are wrong, and the clean-up costs more than the pages were worth. Decide the review capacity first and let it cap the volume.
Common questions
Will search engines penalise AI-assisted content?
The stated position is that the method does not matter and the quality does. Unhelpful content performs badly whoever typed it.
Should we disclose AI assistance?
Our own policy is to say who is responsible for a page rather than which tools were used, which is the part a reader can act on.
What is the realistic time saving?
Substantial on drafting and restructuring, near zero on research once verification is included. Plans that assume the second are the ones that fail.
Can a model do the verification?
Not dependably. Asking a model to check a model’s claim produces agreement more often than accuracy.
