Can AI Sell Jewelry? What We Check Before Trusting Assistants With Customers
AI assistants are starting to answer buyers' questions about sizing, materials and care. Here is where they help, where they hallucinate, and the test we run before trusting one.
Shops are wiring AI assistants into their storefronts to answer the questions that used to fill their inboxes: Will 8 mm fit my septum? Is 14k okay for sensitive skin? Can I shower with it? The appeal is obvious — these questions arrive at 23:40 and the sale closes or dies by 23:45.
The risk is just as concrete: jewelry answers are precision answers, and language models are confident even when wrong.
Where assistants genuinely help
- Catalog navigation. “Show me gold hoops under 200 €” is retrieval, not expertise. Low risk, real convenience.
- Policy questions. Shipping, returns, engraving options — anything with a written source the assistant can be anchored to.
- Care basics. General guidance (remove jewelry before chlorine pools) is stable, well-documented knowledge.
Where they fail in ways that cost money
- Sizing. The difference between 6, 8 and 10 mm inner diameter is the difference between a keeper and a return. Models routinely blur size guidance or invent “standard” sizes that aren’t standard for the piece in question.
- Materials. “Is this hypoallergenic?” needs the actual alloy and plating data of the actual SKU. An assistant answering from general knowledge answers about jewelry in general — with your brand’s authority attached to it.
- Availability and price. Ungrounded assistants will happily state stock, delivery dates and discounts that don’t exist. Every one of those is a customer-service incident with a receipt.
- Medical-adjacent advice. Healing piercings are a health topic. “You can change it after two weeks” is the kind of sentence that must come from your policy, not a model’s average of the internet.
The test we run
Before an assistant faces customers, we score it on a fixed question set built from real inbox history: sizing edge cases, material and skin questions, care during healing, policy traps. Every answer is graded by a human who knows the right one, on three axes — factually correct, grounded in this shop’s data, safe when unsure. The passing behavior for an unanswerable question is not a fluent guess; it is “I’ll connect you with us directly.”
The pattern from testing: assistants are ready to route and retrieve long before they are ready to advise. Deploy them in that order, and wire every precision topic — sizes, materials, stock — to your product data, not to the model’s memory.
Where this leads
This is the thinking behind our Audit work. If you want it applied to your shop, start where we start: the Jewelry Commerce Diagnostic is free — we analyze your shop, brand and market, and tell you what to fix first, in plain language.
Work with StudioFormel
