What does a human actually have to review on AI-drafted content?
Four things: every factual claim, every number, every statement about what your business does or promises, and anything with legal or licensing exposure. Grammar and style are the least important checks, because models rarely fail there. The reviewer's job is verification and accountability — someone has to be willing to put their name on the claims — not polish.
The failure modes are not where people look
Reviewers instinctively read for awkward phrasing, because that is what bad writing used to look like. Model output is smooth, so this instinct catches almost nothing while consuming most of the review time.
The real failures are confident and wrong: an inverted causal relationship, a code requirement that applies in one state and not yours, a warranty term stated as universal, a capability attributed to your company that you do not offer. All of it reads perfectly, which is why the review step in a content automation workflow has to be designed rather than assumed.
A review pass that fits in a few minutes
- Numbers first. Highlight every numeral on the page. Each one either traces to a supplied source or comes out. This single rule removes most of the risk.
- Claims about your business. Anything beginning "we" or describing your process, coverage, guarantees or credentials. Models fill these in from context and get them plausibly wrong.
- Regulatory and safety statements. Anything about permits, code, licensing, refrigerant handling, electrical work or consent requirements. These vary by jurisdiction and must be written generally or verified locally.
- The one-thing test. Can the reviewer point to a sentence a competitor could not have written? If not, send it back for a specific.
- Cross-page consistency. Does this page contradict something you published three months ago? Contradictions inside a corpus are corrosive.
Split the review between two kinds of people
Subject matter review and editorial review are different jobs and rarely live in one person. The technician or sales lead checks whether it is true. The editor checks whether it is clear, non-duplicative, correctly linked and consistent with everything else in the hub.
In small companies this works best as an asymmetric split: the expert spends a few minutes per page on truth, the marketing owner spends the rest on everything else. Trying to make the expert do a full editorial pass is how content programs stall.
Record the review, not just the result
Store who reviewed each page and when, alongside the source material used to draft it. This is not bureaucracy. When you later discover that a source fact was wrong, that record is what lets you find every page that inherited it instead of re-reading the whole corpus.
That is the same discipline we apply to any operational AI workflow: the output is only trustworthy if you can trace what produced it and who signed off. Content is no exception, and it is one of the few places where the audit trail is cheap to build from the start.
Topics: editorial review · fact checking · workflow · quality control
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.