Where AI genuinely helps a content team — and where it does not

Most proposals to add AI to a CMS are solutions looking for a problem. A few are genuinely useful. The difference is whether a human still owns the output.

Tasks that suit automation

• Draft alt text for images, queued for a human to approve rather than published blind.
• Summarise long documents into a standfirst an editor then edits.
• Suggest taxonomy terms from existing vocabulary, never inventing new ones.
• Flag content for review — pages that contradict a newer page, or have not been touched in two years.
• Normalise messy imports, mapping inconsistent legacy values onto a clean set.

What these share: the model proposes, a person disposes, and being wrong is cheap and visible.

Tasks that do not

• Publishing generated copy with no review. It is the fastest way to put a confident falsehood under your own logo.
• Anything legally binding — pricing, eligibility, safety guidance, medical or financial information.
• Replacing search with a chatbot that cannot say “I don't know”.

Build it so it can be switched off

Treat the model as a third-party dependency that will be slow, change behaviour, or be unavailable. Keep calls asynchronous, cache aggressively, log inputs and outputs for review, and make sure the editorial workflow still functions when the integration is disabled. If turning it off breaks publishing, it was never an assistant.