We have been running AI writing tools in production across three content sites for a year. Not as an experiment — as part of the actual workflow, with real traffic on the line.
Here is what held up.
Where they genuinely earn their keep
Structural first drafts
Give a good model a detailed outline and it returns a serviceable skeleton in seconds. The prose needs rewriting; the structure usually does not. That is a real hour saved per article.
Format conversion
Turning a 3,000-word guide into a newsletter, a script, and eight social posts is mechanical work, and models are excellent at mechanical work. This is the highest-ROI use we have found.
Editing pressure
Asking "what did this article fail to answer?" surfaces gaps a writer staring at their own draft will miss. Used as a critic rather than an author, the tools are consistently useful.
Where they still lose
Anything requiring first-hand experience
A model cannot tell you that a tool's export silently truncates at 10,000 rows, because it has never hit the limit. This is precisely the material that differentiates content worth ranking, and it is exactly what models cannot fabricate honestly.
Anything where being wrong is expensive
Pricing, specifications, legal detail. Models state outdated figures with total confidence. Every number in a published piece needs a human check against a primary source — which, on pricing-heavy content, erases most of the time saved.
The honest summary
AI writing tools moved our drafting time down meaningfully and our editing time up slightly. Net, they are worth it. What they have not done — for us or, as far as we can tell, for anyone — is remove the need for someone who has actually used the thing they are writing about.