Grounded, not generated: how Cartfilo avoids AI hallucination in your cart

The single riskiest failure mode for an AI shopping assistant isn’t a bad recommendation — it’s a confident, wrong one. A price that’s out of date. A product that’s actually out of stock. An item that doesn’t exist at all.

Cartfilo is built so that class of error can’t happen structurally, not just “usually doesn’t happen.”

The language model never sees your database directly

When a shopper’s message comes in, the model’s only job is to decide intent and search terms — what category, what quantity, what constraints. It’s handed a numbered list of real candidate products and asked to pick indices from that list.

It never sees a raw price, a raw stock count, or a product ID. Those get filled in afterward, from your live Shopify data, after the model has made its selection. If a candidate goes out of stock between the search and the reply, the reply reflects that — the model can’t contradict it, because it never had the chance to state a number in the first place.

Cart mutations aren’t an AI decision either

Once a basket is assembled, adding it to cart happens client-side, through Shopify’s own Cart AJAX API — the same mechanism your theme already uses. The assistant proposes; Shopify’s own commerce engine executes. Checkout, tax calculation, and payment never pass through the AI layer at all.

This is also why Cartfilo can degrade gracefully under load or budget limits instead of failing open with a made-up answer — see pricing for how usage and degraded-mode notices work.