AI Catalogue Enrichment at Scale: From Supplier Chaos to Channel-Ready in Days

Product content is the least glamorous bottleneck in commerce: every new season arrives as supplier spreadsheets in someone else's schema plus a folder of photography, and between that pile and a channel-ready listing sits weeks of human copywriting, attribute-hunting and translation. This is now one of the most reliably automatable workflows we build — twelve thousand products in nine weeks, in one client's case — but the architecture that works is specific, and the naive version ("ChatGPT wrote our listings") fails in ways that cost marketplace rankings.

The pipeline that works

1. Extraction from every source, reconciled. Vision models read the photography (silhouette, closure, pattern, material texture); language models parse the supplier sheet (fields, footnotes, the sizing table embedded as prose). The two disagree regularly — the sheet says cotton, the weave says blend — and the system's job is to flag contradictions, not silently pick one. Silent resolution is how catalogue lies are born.

2. Attribute mapping into one canonical model. Extracted facts land in your normalised product schema — units converted, vocabularies controlled, GTINs validated — from which every channel rendering derives. Enriching per-channel instead of per-catalogue is the classic architecture mistake: five channels, five drifting truths.

3. Generation per channel, per market. Titles to each marketplace's grammar; copy written natively per market rather than translated (localisation depth shows — German buyers read machine-translation instantly); claims constrained to a compliance vocabulary (cosmetics, toys, supplements each carry per-market claim law).

4. Confidence gates the humans. Every field carries a score; high-confidence output publishes, the uncertain tail routes to reviewers with the model's reasoning attached. The ratio is the economics: at launch expect one product in five needing eyes; with a feedback loop folding corrections into the eval set, mature pipelines run below one in ten. Reviewers move from writing to judging — the same human-in-the-loop shape as every agent system we ship.

The accuracy economics

The build pays back on three lines: copywriting spend displaced (the visible one), time-to-channel (a season live in days instead of ten weeks — revenue that simply did not exist before), and attribute completeness lifting filtered-search visibility across every connected surface. Against that: model spend (trivial — fractions of a cent per field with routing discipline) and the review team you keep, smaller and more senior.

Where it breaks

Regulated claims without a constrained vocabulary (the model will cheerfully promise dermatological miracles); photography too poor to extract from (garbage in, confidently-described garbage out); category taxonomies mapped by vibes instead of validated against each channel's schema; and skipping the eval set, which converts every model update into silent catalogue-wide drift. Each failure is preventable, and each prevention is architecture, not heroics.

Catalogue enrichment is the rare AI project with clean ROI math, bounded risk and a two-month path to production. If your next season is still a copywriting project, the audit version of this article takes a week.