Ten years of supplier imports left an office-supplies catalog with chaotic attributes, wrong categories and 1,900 duplicates. We ran an AI normalization pipeline with human review on every change. A cleanup that would take months of manual entry finishes in about three weeks.
Typical results
12,000
SKUs processed by AI
97%
SKUs with complete attributes
+17%
On-site search conversion
Figures reflect typical results for this kind of automation.
The challenge
Twelve thousand SKUs imported from dozens of suppliers over a decade: 'blue', 'BLU' and 'navy' as three colors, paper sizes in four formats, printers filed under furniture. Only 48% of SKUs had complete attributes, search results were unreliable, and an agency quoted six months of manual data entry to fix it.
Before and after
Cleanup timeline
−88%SKUs with complete attributes
2×Duplicate listings in catalog
−100%The automation
We built an AI pipeline over the full catalog. OpenAI-powered passes standardized every attribute into one schema, re-assigned wrong categories, clustered near-duplicates for merging and flagged SKUs missing critical data. Nothing published blind: every proposed change landed in a review sheet where the team approved in bulk or line by line. The same checks now run on every new supplier import, so the catalog can't drift back.
How it works
- 01
Full catalog scan
All 12,000 SKUs are exported with every attribute, description and category assignment.
- 02
AI normalization
OpenAI-powered passes standardize attribute names and values — sizes, colors, units — into one clean schema.
- 03
Duplicates & gaps
Near-duplicate listings are clustered for merging, and SKUs missing critical data are flagged.
- 04
Human review
Every proposed change lands in a review sheet; the team approves in bulk or line by line.
- 05
Publish & guard
Approved changes publish to the store, and new supplier imports pass the same checks at intake.
Starts from
- 12,000-SKU catalog
- Supplier data files
Delivers
- Change review sheet
- Approved updates published
- Duplicate & gap flags
The impact
A full cleanup typically takes about three weeks instead of months of manual entry. Attribute completeness can rise from 48% to 97% of SKUs, around 1,900 duplicate listings are merged away, and on-site search conversion tends to rise by about 17% as customers find what they type.
Questions about this automation
How does the AI decide what a clean attribute looks like?
The team first agrees on one target schema: attribute names, allowed values and units for each category. The AI maps every messy value to that schema, so spelling variants of the same color or paper size resolve to one agreed value. Each proposed change carries the original value, so reviewers see exactly what would change.
Does anything publish to the store without human review?
No. Every proposed change, from a renamed attribute to a merged duplicate, lands in a review sheet first. The team can approve whole groups in bulk when the pattern is clear, or go line by line on sensitive items. Only approved rows are written back to the store, and rejected suggestions stay logged for reference.
How does the catalog stay clean after the first cleanup?
The same checks run on every new supplier import. Incoming rows are normalized to the schema, tested against existing listings for near-duplicates and flagged when critical attributes are missing, before they reach the live catalog. It is a multi-step build, usually rolled out in stages: the one-time cleanup first, then the intake guard.