Can AI merge and deduplicate a parts catalog?

AI can merge a parts catalog, but the hard part is not matching similar text. It is knowing when similar-looking items are not the same item.

PAIR-20's July 2026 T06 task used three synthetic MRO distributor catalogs. The job was to normalize every price to per-piece cost, find genuine duplicates, and refuse to merge false duplicates where the specification changed. The task file called out the risk directly: glove thickness, bearing clearance class, bolt grade, and blank pack quantities can change the buying answer.

The headline result looked good. ChatGPT, Claude, and Gemini treated the nitrile gloves as not directly comparable and flagged Atlas's missing box count as a field to confirm before pricing. On the consolidation question, all four platforms agreed on Summit.

The failure was narrower. Copilot wrote "Likely 100ct" for Atlas's unknown box quantity. The case file said Atlas's box quantity varies by item and must be confirmed, not assumed. That turns a blank field into a denominator for a per-piece comparison.

Text matches are not enough. A catalog merge has to match function, dimensions, material, grade, compliance rating, manufacturer part number, pack quantity, and footnotes. Description similarity is only the starting point.

Blank fields should stay visible. In a procurement table, "confirm with supplier" is a usable answer. A guessed quantity is worse because it looks like data and flows into the price ranking.

Review the rows that drive consolidation. AI is useful for building the first merged table and finding likely duplicates. Before using it to award a primary supplier, check the false-duplicate list, the pack-quantity source, and any row where an unknown field would change the per-piece price.