AI in Your Content Pipeline?: What to Automate and What Never To

The question every small publisher is being asked right now is where AI fits in the workflow. The honest answer is that it fits in a narrow band — genuinely useful for structured, checkable, repetitive work, and actively harmful in the places where your reputation actually lives.

Here is the line as we draw it.

Where it works

Metadata and catalogue admin. Generating keyword sets, BISAC or Thema code suggestions, retailer description variants at different lengths from an existing approved blurb. This is reformatting of something already written and approved, and it is checkable in seconds.

First-pass transcription and indexing support. Interview transcripts, OCR cleanup on archival material, candidate index terms for a human indexer to accept or reject. The human still decides; the machine removes the typing.

Translation first drafts for internal comprehension. Reading a foreign-language source to decide whether it is worth commissioning a proper translation. Not for anything published.

Marketing variants. Subject lines, ad copy variations, social versions of an existing paragraph. Low stakes, easily judged, and you were going to write six versions anyway.

Internal summarisation. Long submissions, contracts, reports — a summary to decide what deserves your full attention. The decision is still yours.

Where it does not

Anything factual in a published book. Fabricated dates, invented citations, and confidently wrong details are the known failure mode, and in specialist non-fiction your readers are the people most likely to catch it. One invented source in a serious history title damages the imprint permanently.

Final translation of published text. Specialist terminology, period language, and the difference between a technically correct rendering and the right one are exactly what the tools handle worst.

Author or editorial voice. Readers of a small press are buying a sensibility. Generated prose reads as no sensibility at all, which is worse than a flawed one.

Correspondence with authors and rights holders. These are relationships. A generated rejection or a generated permissions request is detectable and it is remembered.

Image generation for anything historical or documentary. In a field where readers value provenance, a generated image on a cover is a credibility problem regardless of how good it looks.

The rule underneath

Automate where output is verifiable in less time than it would take to produce by hand. Metadata passes that test — you can check a description against the book in thirty seconds. A claimed source in a history book fails it, because verifying properly takes longer than writing it yourself.

Say what you do

Decide your position on AI use and state it publicly. Not because anyone is demanding it yet, but because specialist audiences care about provenance, and being clear early is much cheaper than being asked later and appearing to have been caught.

A short line on your site is enough. Something to the effect of: our books are written, edited, translated, and illustrated by people; we use software tools for administrative work like metadata and transcription. That sentence costs you nothing and it removes a doubt that is going to become more common, not less.