An AI-assisted content workflow

The argument about whether to use AI in content production is over in practice. Almost everybody is using it, including most of the people arguing against it.

The useful question is where in the process it belongs, and what has to be true before anything publishes.

What it's good at and bad at

Good: outlining, drafting from material you hand it, rewriting dense paragraphs into something liftable, summarizing long documents, generating title and description options, spotting what a section is missing, translating internal jargon into what clients actually say, checking whether your opening sentence stands up alone.

Bad: producing a statistic, writing the experience section, deciding what a regulation means, taking a position on something contested, and generating pages in bulk to cover a keyword list — which is the one Google's spam policies name outright as scaled content abuse.

The test is whether the task needs information the model doesn't have. If it does, it won't tell you it's missing. It'll generate something plausible instead.

That property is the entire reason the gate exists.

The gate

One human. Named. Accountable for the thing being correct. Reads it before it publishes.

What they're checking for is specific, and it isn't writing quality:

1. Every factual claim, especially the ones that sound right. Generated text produces plausible figures with no source and real-looking citations to things that don't say what's claimed.
2. Every number, against a primary source rather than another article.
3. The experience sections. Is this something we did, or something that sounds like what a firm like us would do?
4. The positions. Does this commit to anything, and do we agree with it?
5. The omissions. What got left out because the training data under-represents it?

A copy editor cannot do this. Fluency isn't the failure mode — it's the camouflage. The gate has to be somebody who knows the subject well enough to notice that a confident paragraph is wrong.

Make it structural

A process that relies on discipline fails in the quarter when everybody's busy, which is exactly the quarter when volume goes up. So the gate should block publication rather than be something people are supposed to remember.

In practice:

• A required field recording whether AI was used, that can't be left blank.
• A required reviewer, by name, not "the team."
• A publishing step that refuses when the reviewer is missing, rather than warning about it.
• Tighter rules for YMYL, where the reviewer has to be a different person from the author and competent in the subject.
• A published editorial policy describing all of it, linked from every page.

Disclosure

Google suggests considering a disclosure where a reader might reasonably wonder how something was made, and advises against giving AI an author byline.

So: say who's accountable, and describe the process somewhere findable. Naming a model as the author communicates exactly the wrong thing — it implies nobody is responsible.

What this costs

Honestly? The review gate is most of the cost of the content, and it doesn't shrink with volume.

A team that adopts AI drafting without adding review capacity hasn't reduced its cost per page. It's moved the bottleneck, and usually quietly lowered its standard.

The firms getting real value here are the ones that spent the time saved on drafting buying more review, more source-checking, and more firsthand material. That's a smaller productivity gain than the sales pitch promised. It's also the one that survives a quality update.