AI & Visual
AI Shopping Assistants in Fashion
A conversational assistant is only as trustworthy as the catalogue it reads from, which is why data quality decides whether it helps or misleads.
Updated 2026-07-01 · 7 min read
Ask an assistant to "find me a black leather Chelsea boot that ships to the Netherlands this week" and you are testing three things at once: whether it understood you, whether it has current data, and whether it will admit what it does not know. The first is largely solved. The second and third are where these tools live or die.
Shopping assistants are among the most visible expressions of AI-assisted discovery, and they are genuinely useful. But the enthusiasm around them often skips the boring precondition: without clean, fresh, structured product data underneath, a fluent assistant simply produces fluent errors.
What they do well
The clearest benefit is conversation. Shoppers can describe what they want in their own words, refine across several turns ("no, more casual", "same but in suede") and get a coherent shortlist without translating their taste into filter checkboxes. For discovery — surfacing options a shopper had not considered — this is a real improvement over faceted search.
Assistants also compress research. Instead of opening ten tabs to compare a shoe across retailers, a shopper can ask one question and get a synthesised answer. Where the assistant sits on top of a strong product-data layer, that synthesis is fast and accurate. The value is real; it is just conditional.
The data-freshness problem
Fashion catalogues change constantly. Prices move with promotions, sizes sell out within hours, and seasonal ranges rotate. An assistant that answers from a stale snapshot will recommend products that are unavailable, mispriced, or out of stock in the requested size — and it will do so with total confidence, which makes the error worse.
This is not a flaw in the language model; it is a flaw in the pipeline feeding it. Freshness depends on how often offers are refreshed and how quickly stock and price changes propagate, the subject of our pages on merchant offer quality and price tracking. An assistant reading a two-day-old feed is answering yesterday's question.
Hallucination and how to contain it
Generative models can invent. Asked about a product they lack data on, they may fabricate a plausible detail — a material, a colourway, a shipping estimate — rather than declining. In casual chat this is a curiosity; in commerce it erodes trust and, in cases like fabricated availability or sizing, can cause real harm.
The mitigation is architectural, not merely a matter of a better model. Assistants that ground their answers in a retrieved, structured catalogue — and that are constrained to say "I don't have current data on that" — hallucinate far less than open-ended ones. Grounding trades a little fluency for a lot of reliability, which in shopping is the right trade.
Merchant availability and coverage
An assistant's usefulness is bounded by which merchants it can actually see. If it only reads from a handful of retailers, its "best option" is best only within that slice, and a shopper may never learn that a better price or the right size exists elsewhere. Coverage gaps are invisible to the shopper, which makes them especially misleading.
Availability also varies by market. An assistant strong in one country may be thin in another because merchant coverage, language and local stock differ — a theme we develop in localization and international search. Breadth of merchant data is as important as the intelligence applied to it.
Why structured data is the real dependency
Everything above points to the same conclusion: the assistant is the visible layer, but the value is created underneath. To answer well, a system needs products resolved to canonical records, offers normalised across merchants, and attributes — size, colour, material, availability — that are structured rather than buried in free text.
Business of Fashion and McKinsey have both framed conversational commerce as promising but immature, with execution gated by data readiness rather than model quality. That framing matches what the technology shows in practice: fix the data, and the assistant gets good almost for free; leave it dirty, and no model can rescue it.
A measured verdict
AI shopping assistants are worth building and worth using, provided expectations are set honestly. They excel at conversation and shortlisting and struggle with freshness, hallucination and coverage — all of which trace back to data rather than to the model. There is also a quieter risk worth naming: an assistant's tone of confident fluency can make shoppers trust it more than its underlying data warrants, so the interface has a duty to signal uncertainty rather than mask it.
The right posture is neither dismissal nor hype but disciplined investment in the foundation, paired with candid design that admits what the system does not know. Bias belongs on that list too — an assistant trained mostly on mainstream styles, sizes and price points will serve those shoppers best and others less well, which is a data and design problem rather than an unavoidable one. For the broader picture, return to the discovery hub.
Sources & further reading
- Business of Fashion, “The State of Fashion” (2024)
- McKinsey, “The State of Fashion” (2024)
Sources are attributed to their publishers and link to each publisher's own site. Figures reflect general market direction rather than point-in-time precision; consult the linked publishers for their current data.