The business problem
Many mature catalogs predate Shopify’s current product taxonomy or contain category data shaped for another system. Classifying every product by hand is expensive, while fully automatic assignment can turn ambiguous products into confident mistakes.
What DSWD delivered
DSWD built a catalog-wide classification workflow. It handles clear cases with client-specific deterministic rules, sends only the genuinely ambiguous remainder through targeted AI-assisted analysis, separates uncertain results for review, and writes approved classifications back to Shopify.
The operator experience
An operator can run the work without writing changes, inspect clear and uncertain results separately, review exceptions, and approve what is ready. Each run produces an auditable summary of its outcomes.
Safeguards and operating qualities
- Dry-run and staged execution modes keep review ahead of mutation.
- Confidence states remain explicit rather than being hidden behind a single output.
- Validation occurs before Shopify write-back.
- Store-specific rules and configuration remain separate from the classification engine.
- Every run records counts and review splits for later inspection.
Project-specific technical choices
For this project, we used a rules-first classification pipeline with limited AI assistance because obvious matches should be fast, repeatable, and inexpensive, while ambiguous products benefit from additional context. Human review remains the control point for uncertain writes.
Evidence and claim boundary
The workflow has run against a live catalog of more than 500,000 products and supports ongoing production work. Results depend on the catalog, the reviewed rules, and the quality of the source data; no universal accuracy claim is made.