Why similar wording is risky
An invoice, price list, or purchase order may describe a product differently from the internal catalog. The accounting system expects a specific SKU, but a similar text string does not prove that it is the same item. For M6×30 and M6×35 fasteners, a few millimeters can make a product unusable; packaging needs volume and unit checks, and a chemical product may require grade and concentration.
An incorrect automatic match can cause wrong prices, inventory, orders, or postings. The system should therefore act as a candidate finder, not as an authority to change the catalog.
An architecture with human control
A small workflow can use four steps:
1. Read the document and preserve the original line, document number, supplier, and item.
2. Normalize safe variations: spacing, case, common abbreviations, units, and size formats.
3. Find several candidates using codes, text, and required attributes.
4. Show the buyer the options and reasons for each score; write the selected SKU to the ERP only after approval.
A local language model can help parse free-text descriptions, extract attributes, and explain why two names look similar. But numeric parameters, codes, dimensions, and units should be checked with separate rules. A plausible language-model explanation is not a catalog validation.
What data is needed
Start with a catalog export and a few representative documents. For each product category, define required attributes:
- manufacturer part number or code;
- size and unit of measure;
- material, grade, or concentration;
- packaging and quantity;
- compatibility or other properties that affect use.
Matching can combine exact codes, normalized strings, and semantic search over descriptions. External reference sources require quality checks for their identifiers. OpenRefine calls a similar process reconciliation: semi-automated matching of records to an external dataset, where the system proposes matches and a person reviews and approves them.
Store three values separately: the supplier's original description, the chosen SKU, and the reason for the decision. This makes it possible to investigate uncertain matches, improve rules, and retain what the source document actually said.
A pilot without an expensive integration
Collect 100–200 lines from real price lists, removing commercially sensitive fields if they are not needed for the test. Ask an experienced buyer to mark the correct SKU and important differences. Then compare the manual process with the automation's suggestions.
Measure:
- exact-match rate and dangerous false substitutions;
- the share of lines where the model abstains or asks for confirmation;
- review time per line;
- incorrect entries, returns to source documents, and ERP corrections.
The pilot can begin in a spreadsheet or separate interface without direct ERP writes. Only after checking edge cases should the team connect ERP through a separate action that revalidates the SKU and user's permissions before writing.
Limits and economics
If the catalog is incomplete or attributes are inconsistent, a model will not repair the source of truth. Complex categories may require manufacturer reference data, unit rules, or a procurement specialist. Automation pays off through reduced processing time without more incorrect orders or postings, not through a percentage of similar names.
For a small business, a safe first step is to test how the system ranks 100–200 real lines and measure exception-review time. Until quality is stable, let automation suggest, not post.
