Ask a capable AI model what bouclé is and it can usually explain the textile. It may describe the looped yarn, the nubby surface and its use in upholstery and fashion.
Ask whether your store treats bouclé as an upholstery material, a finish, a texture or all three in different contexts, and the problem changes.
General understanding has met local policy.
Meaning is only the first decision
An AI model can infer what a product probably is from its title, description and image. That is valuable.
A retailer still needs to decide:
- Which internal product type to use
- Which standard platform category is the closest fit
- Which attributes are mandatory
- Which values should be canonical
- Which synonyms should remain searchable
- Which related products must stay distinct
- What should happen when confidence is low
These are not merely questions of language. They are questions of governance.
Your catalogue has a constitution
Most retailers do not call it that. Yet every coherent catalogue contains principles about what belongs where.
Perhaps oak means solid oak only when construction confirms it. Perhaps oak effect is a finish, never a material. Perhaps gold describes colour while brass-plated describes construction. Perhaps a chair can be suitable for dining without being classified as a dining chair.
These rules are the constitution of the catalogue. They should not be rewritten with every product. Rewriting them on every call is exactly what makes AI tagging frustrating at scale.
Consistency is not sameness
Logical application does not mean flattening every difference.
It means that the same approved distinction is preserved wherever it appears. If solid oak, oak veneer and oak effect are different in the merchant's taxonomy, consistency protects the difference. It does not erase it.
The better the rule, the less generic the output can become.
Explainability matters when a decision affects the catalogue
When a suggestion is wrong, a merchant needs more than another suggestion.
They need to know what evidence was used, what rule applied and what will change if they correct it. Otherwise, review becomes a sequence of isolated approvals with no visible learning.
A dependable workflow should be able to say:
Suggested type: Upholstered dining chair
Reason: title and specifications identify dining use, seat height and upholstered seat
Mapped store category: Dining Furniture
Mapped platform category: Furniture > Chairs > Dining Chairs
Confidence: High
The merchant can then approve, amend or reject the classification.
Human control is not a concession
The merchant knows what the product means commercially. That knowledge is not an inconvenience automation must work around. It is the source of truth the system should learn from.
NIST's AI risk guidance emphasises the importance of managing, measuring and governing AI systems rather than treating generated outputs as self-validating. The commercial stakes here may be modest compared with critical systems, but the principle travels well: an output should be tested against the purpose for which it will be used. NIST AI Risk Management Framework
Intelligence where it matters. Logic where it counts.
Semantikal uses AI for the part that benefits from interpretation: learning the vocabulary, relationships and implied structure of a catalogue.
Then the retailer calibrates that understanding. Once approved, logic applies it repeatedly, visibly and at scale.
AI can understand a product. Semantikal makes that understanding dependable enough to run a catalogue.