AI in ecommerce: what's actually working in production in 2026

Most "AI in ecommerce" writing is either a vendor pitch or a list of things nobody has actually run at scale. This is the opposite: a survey of where AI has genuinely earned a permanent place in a live multi-vendor marketplace we operate, where it has not, and the single decision — data before models — that determines whether any of it returns more than it costs. Five areas pay off in production today. One category we now refuse to build.

The build order: data before models

The teams that get poor results from AI in commerce almost always inverted the order: they bought a model feature before their catalog was clean, their search was instrumented, or their data had a contract. A large model reasoning over a messy catalog produces confident, well-written nonsense. The same model over a clean, structured catalog is quietly excellent.

So the order that works is: fix the data, expose it through tools with a stable schema, then add models on top — and only where a deterministic system genuinely can't do the job cheaper. Roughly a dozen enrichment steps in our pipeline are pure code (unit parsing, price maths, variant grouping); the model is invited only to fill the gaps code can't. That ordering is why our economics work at all, and it is the theme running through every section below.

Catalog enrichment: the highest-leverage AI in the stack

If you do one AI thing, do this. Vendor feeds arrive messy — attributes buried in prose, languages mixed, variants ungrouped, brands sometimes exported as numeric option ids. A deterministic-first pipeline normalizes what it can, then a model fills the remaining gaps at temperature zero against a strict schema, generating native Estonian, English and Russian rather than translating, plus a handful of search synonyms per product per language. The payoff is compounding: better titles, better facets, better search, better assistant grounding — all from one pass. We cover the mechanics in AI product enrichment, and the reason it stays cheap in LLM cost optimization.

Two details separate enrichment that works from enrichment that embarrasses you. First, the model runs behind category-aware guardrails and post-generation validators, so a supplement gets supplement-shaped fields and a piece of equipment does not — and anything that fails validation fails closed to an empty field rather than a confident guess. Second, vision models read nutrition and spec information straight off product photos when the feed omits it, behind a sanity gate that rejects incoherent extractions. The output is not "AI-written descriptions"; it is structured, validated data the rest of the stack can trust.

Search: hybrid, not magic

Search is where shoppers with intent go, so it is where AI has to be careful rather than clever. What works is a hybrid: a fast keyword engine (typo tolerance, facets, synonyms) fused with vector search that triggers selectively — on long queries, intent phrasing, cross-language input, or when lexical recall is weak. Always-on semantic search is a downgrade; it blurs precise queries and burns compute. The full argument is in marketplace search, and the retrieval principles carry straight over to grounding an assistant, covered in RAG over product data.

Assistants: grounded or don't bother

A shopping assistant is worth building only if it can never recommend a product you can't sell. That means the model answers from live tool calls — search, stock, price, delivery — never from memory, and an empty tool result becomes an honest "we don't carry that". Wrap that grounding in a guardrail chain and typed refusals and you have a feature; skip it and you have a liability. The architecture is in building an AI shopping assistant, the guardrail layer in LLM guardrails in production.

In practice the assistant is worth shipping in two shapes on the same foundation: a structured guided flow that turns a short goal quiz into a plan of real in-stock products with a reason per item, and a free-form chat that handles the specific long-tail questions. The guided flow converts the shopper who arrives thinking "I don't know what I need"; the chat answers the one who knows exactly what to ask. Both call the same tools and pass the same guardrails, so you build the hard part once.

Content and operations: AI authors, systems execute

Behind the storefront, AI runs the unglamorous work: drafting articles behind an evidence-first citation gate, scoring image relevance, watching for broken links and index drift, and summarizing a weekly health sweep into a findings log a human acts on. None of it is autonomous in the scary sense — every AI output passes a deterministic check before it touches production. The pattern generalizes far beyond commerce, and we detail it in the white-label marketplace guide.

What we'd skip

The gimmick chatbot — a general model in a widget with no access to your catalog — is the one we refuse to ship. It hallucinates products, invents policies, leaks to competitors, and generates support tickets instead of removing them. The tell is simple: if the feature can answer a question your data can't verify, it will eventually answer wrong in public. We also skip "AI" that dresses up a hardcoded rule engine, and any personalization that can't be measured against a real conversion signal. When the underlying system is honest, the AI on top is an asset; when it isn't, the AI is a faster way to be wrong.

Frequently asked questions