Catalog, editorial and merchandising content was bottlenecked behind a small group of senior copy editors. Time-to-publish for a new SKU averaged 23 days, against a target of 5.
The group's reputation was built on brand voice. Any system that produced 'good enough' but generic copy was an existential risk - not a productivity opportunity.
Fourteen markets, six languages, three brands - each with its own editorial style guide, market positioning and merchandising calendar.
Built a brand-voice corpus from 12 years of editorial output across the group's three brands, encoded as structured rubrics rather than as raw training data.
Designed a constrained-generation pipeline: the model produced multiple variants, an automated brand-voice score filtered the top quartile, and a human editor picked from a shortlist of 3-5. Editors retained authorship; the system industrialised drafting and localisation.
Generated localised SKU content across the six languages in parallel, with a market-by-market merchandising overlay (price positioning, story emphasis, comparable-product set).
Continuous-quality assurance ran daily, with a measured brand-voice score per piece of published content. Three pieces were rejected for re-work in the first month - and none after month three.
We had stopped believing AI could write our brand. Moweb's framing - that the system drafts and we author - turned out to be exactly right. Editors are now doing the work they were hired for.