Where AI Actually Works in Ecommerce (And Where It Doesn't)

There is no shortage of AI features you can add to an ecommerce store, and a shrinking amount of patience for the ones that do not pay for themselves. The useful distinction is not between clever and simple. It is between features where being occasionally wrong is cheap and features where being occasionally wrong is expensive.

That single test explains most of what follows. Language models are probabilistic. Deploy them where a wrong answer costs a shrug, and they are transformative. Deploy them where a wrong answer costs a refund, a compliance breach or a customer, and you have built a liability with a chat interface.

Standard storefront search matches keywords. A customer searching for "something warm for hiking in november" gets nothing, because none of those words appear in your product titles. Semantic search — embedding your catalogue and matching on meaning — returns the insulated jackets.

It works because the failure mode is benign. A slightly odd result is a mild annoyance, identical to the mild annoyance of a keyword search returning nothing, and the upside is that a whole class of natural-language queries starts converting instead of bouncing.

It is also unusually easy to measure. Zero-result search rate and search-to-purchase conversion are both already in your analytics. You can prove the value in a fortnight, which is more than can be said for most AI features.

Fully autonomous support agents are where most ecommerce AI projects go wrong. The model does not know your returns policy is different for sale items, confidently says otherwise, and now you are honouring a policy you do not have — or arguing with a customer holding a screenshot.

Triage is where the same technology earns its keep quietly. Classify the incoming ticket, pull the relevant order, draft a reply, and put it in front of a human who approves or edits it. The human stays accountable and the throughput roughly doubles.

Retrieval matters more than the model here. An assistant grounded in your actual policy documents and this customer's actual order history is useful; one relying on the model's general knowledge of how returns usually work is a hazard. If you take one architectural point from this piece, make it that: ground the answer in your data, and make it cite what it used.

Generating descriptions for four thousand SKUs imported from a supplier feed is a genuinely good use of a language model, with two conditions. The input has to be real structured attributes rather than the product title alone, and the output has to be reviewed before it goes live.

Without the first, the model invents specifications, and invented specifications on a product page are a returns problem and, depending on the category and jurisdiction, a legal one. Without the second, you have published four thousand pages of unverified claims under your own brand.

The realistic gain is not zero-effort content. It is turning a task that would take a copywriter six weeks into one that takes them one week of editing. That is still an excellent trade — it is just a different claim from the one usually made.

Inference cost is usually the smallest line. The real ongoing costs are evaluation, monitoring and the quiet drift that comes from a catalogue changing underneath a set of embeddings nobody scheduled a refresh for.

There is also a latency budget that gets ignored until launch. A recommendation call adding four hundred milliseconds to a product page is a conversion cost, and it can easily exceed whatever the recommendation earned. Anything on the critical rendering path needs to be cached, precomputed or moved off it.

Build the measurement before the feature. An AI feature with no evaluation set is not a feature you own, it is one you hope about.

Search, support triage and catalogue content are where we see AI reliably pay for itself in ecommerce. They share a shape: high volume, tolerant of imperfection, easy to measure, and with a human in the loop wherever a mistake would be expensive.

The features that disappoint are the ones sold as replacing a person rather than accelerating one. Aim for the acceleration, measure it honestly, and the technology is genuinely worth the effort.