Use Cases of AI in Retail and E-Commerce: The Complete Catalogue

What are the main use cases of AI in retail today?

The main use cases of AI in retail and e-commerce today fall into six groups: discovery and search, personalisation and recommendations, conversational commerce, product data and catalogue enrichment, forecasting and pricing, and in-store computer vision — backed up by back-office automation across supply chain, returns, support and marketing. Each needs a different technique and, more importantly, a different dataset, which determines whether you can start this quarter or next year.

Most retail AI lists are flat, mixing a chatbot that answers order-status queries with a forecasting programme that restructures a planning organisation. This catalogue is organised by where each use case sits in the business, and every entry states the technique it requires and the data it needs. If you can't name the dataset, you don't have a project — you have an aspiration.

Discovery and search: why keyword search leaks revenue

On-site search is where intent is at its highest, and it's where most catalogues quietly fail. An empty result set is a lost sale from your most motivated visitor. Keyword search matches tokens: it works when the shopper knows your vocabulary and fails when they describe things the way people actually do. "Something warm for a wedding in October" and "a quiet washing machine that fits under a counter" are perfectly normal ways to shop, and neither is a token match against a product title.

Semantic and hybrid retrieval

Embedding the catalogue lets you search by meaning. Pure vector search isn't enough on its own, because it blurs the exact tokens shoppers rely on — SKUs, model numbers, sizes. The production answer is hybrid: semantic and keyword retrieval together, fused and then reranked — the same pattern that powers retrieval-augmented generation, applied to product discovery.

• Technique. Embeddings on titles, descriptions and attributes; a BM25 index in parallel; a cross-encoder reranker on the fused candidates; query rewriting.
• Data required. A catalogue with descriptions worth embedding. Thin, templated text produces weak embeddings, which is why enrichment comes first.
• Watch for. Zero-result queries — the cheapest diagnostic in retail.

Visual and multimodal search

Multimodal models map images and text into a shared space, which makes "find products that look like this" a retrieval problem rather than a research project. For fashion, homeware and spare parts, this removes the hardest step in the funnel: describing the thing you're after in words. The engineering is mostly about the imagery — consistent, background-clean photography retrieves very well; mixed crops don't. See vision-language model applications.

Personalisation and recommendations — and their honest limits

A recommender is not a ranker

A recommender predicts affinity: what is this customer likely to want? A ranker decides what to show in a given slot, weighing affinity against margin, stock position, delivery promise and diversity. Teams that use the raw recommender score as the display order end up with the classic problems: the same three items everywhere, and bestsellers crowding out everything else. The fix is two-stage — candidate generation for recall, then a learned ranking layer that explicitly carries the business constraints.

Cold start is a permanent condition, not a launch problem

Retail catalogues are always cold-starting: new products arrive weekly, seasonal ranges have no history, and most visitors are anonymous.

• Content-based fallbacks. Embeddings from image and attribute data give a brand-new SKU a sensible neighbourhood before it has a single interaction — another dividend of good catalogue data.
• Session-based models. For anonymous traffic, the session is the signal, and sequence models on in-session behaviour beat anything that requires a stable user identity.
• Explicit exploration. A recommender trained solely on what it has already shown will keep narrowing. Set aside a portion of impressions for exploration.

Over-personalisation is a real failure mode

Personalisation that's too tight turns the catalogue into a corridor: a shopper who once bought running shoes sees nothing but running shoes for months. Worse, personalisation based on sensitive inferences — health, financial pressure, pregnancy, religion — comes across as surveillance and creates regulatory exposure. Define what you will not infer, and enforce it in the feature pipeline rather than on a slide.

Conversational commerce: assistants that can finish a task

The previous generation of retail chatbots answered questions. The current generation completes tasks. That difference is architectural, not a matter of prompting. A bot that can only talk is a search box with worse ergonomics; one that can check live stock at a named store, apply a promotion, change a delivery slot, start a return and issue the label, or reorder a previous basket is doing work the customer would otherwise have to queue for.

• Technique. A tool-using agent over your order, catalogue, inventory and promotions APIs, with grounded retrieval for policy questions and a deterministic confirmation step before anything that moves money or stock.
• Data required. Documented, stable APIs, more than data. If order status is only reachable by screen-scraping a legacy portal, then that integration is the project.
• Where it breaks. Unbounded autonomy. Agents that issue refunds without a policy check create liabilities very quickly. Constrain the action space and log every tool call.

Voice matters more in retail than most teams realise, particularly for order status and returns. We build these as AI chatbot development engagements, with explicit escalation rules and a logged action space.

Product data and catalogue enrichment: the highest-leverage use case

If you take one thing from this catalogue, take this section. Catalogue enrichment is the least discussed retail AI application and consistently delivers the highest return, because product data is an input to everything else: search quality, cold start, ad feed performance and AI-assistant visibility all sit downstream of it.

Large catalogues are a familiar mess: missing attributes, inconsistent taxonomy, duplicate listings, no alt text. You can't fix it manually at scale, which is why it never gets fixed. Vision-language models change the economics, because they read the image and the supplier text together.

• Attribute extraction. Derive colour, material, pattern, capacity, fit and compatibility from the images plus the text, then write them back as structured fields. That's what makes faceted filtering possible.
• Taxonomy mapping. Map supplier categories to your internal taxonomy and to the external taxonomies that marketplaces and feeds require — classification with a review queue.
• Deduplication. Find the same product listed several times under different titles. Image and text embeddings with a blocking strategy, not string similarity.
• Alt text. Generate descriptive alt text for every image: an accessibility obligation, and a machine-readable signal that search and assistants consume.
• Description generation. Draft descriptions from verified attributes, never from the model's imagination, or you'll end up publishing claims that aren't true.

Two things determine whether this works. First, confidence thresholds with a human review queue, because a pipeline that writes low-confidence guesses into the catalogue degrades it. Second, treating it as a versioned, evaluated batch pipeline. That's ordinary generative AI development discipline applied to a data asset.

Pricing, markdown optimisation, and demand forecasting

This group has the biggest financial impact and the highest data bar. These aren't language-model use cases at all; they're classical machine learning and optimisation problems that live or die on data quality.

Pricing and markdown

Price optimisation estimates the demand response and picks prices subject to constraints: margin floors, competitor position, price-family consistency, brand rules. Markdown optimisation is the same engine applied to end-of-life stock. The failure mode is governance rather than modelling: an engine with no guardrails will find some very embarrassing prices, and coordinated algorithmic pricing raises competition-law questions. Constraints belong in the optimiser; a human owns the policy.

Demand forecasting and allocation

Forecasting underpins replenishment, allocation and labour planning. Current practice uses global models trained across the entire SKU-location grid rather than one model per series, which lets a new item borrow the seasonal shape of its category.

• Technique. Gradient-boosted trees or global neural forecasters on hierarchical SKU-store-week series, reconciled across the hierarchy, with probabilistic rather than point outputs.
• Data required. Several years of sales history at the forecast grain, plus promotions, price history, calendar features and — critically — stock-out flags. Without those, the model learns that demand for out-of-stock items is zero and under-orders forever.
• What to optimise. Not accuracy — service level against holding cost. A forecast that is slightly worse on average but considerably better in the tail is usually the more profitable one.

In-store computer vision: shelves, queues, and loss prevention

Physical retail produces a constant stream of visual data that nobody has time to watch, and computer vision turns it into structured events. The models themselves are largely commodity; the hard part is deployment, drift and the law.

• Shelf monitoring. Detect gaps and out-of-stocks from fixed cameras or shelf-scanning robots. Object detection plus product classification — and the classifier needs re-labelling whenever the packaging changes.
• Planogram compliance. Confirm the actual shelf matches the plan: facings, adjacency, share of space. Detection plus spatial matching against a machine-readable planogram.
• Queue and footfall analytics. Count people, measure dwell time and queue length, and trigger till-opening alerts. Anonymous counting, never identification.
• Loss prevention. Detect scan avoidance at self-checkout. The right output is an alert to a colleague, not an automated accusation.

Privacy and consent are the gating constraint

In-store vision processes personal data whether or not you intend to identify anyone. Under GDPR-style regimes and in the EU AI Act era, the obligations are practical: a lawful basis and a documented impact assessment; signage where cameras operate; minimisation by design, meaning on-device processing that emits counts rather than storing faces; retention limits; and a hard line against biometric identification of shoppers, which is prohibited or heavily restricted depending on the jurisdiction.

Design the system so it never needs identity. Teams that build identity in "because we might want it later" end up with a system that legal won't sign off. We cover the operational side in computer vision and MLOps, where the harder problem is keeping hundreds of cameras consistent as lighting and packaging change.

Supply chain, returns, support, and marketing production

The back office is where AI in retail pays back quietly. None of it is visible to the shopper; all of it shows up in cost to serve.

• Supply chain and replenishment. Inventory optimisation, lead-time prediction, supplier risk monitoring and exception triage — the wider pattern is in our guide to AI in supply chain management.
• Returns and reverse logistics. Predict the likelihood of a return at purchase, which in fashion is largely a sizing problem. Then triage the returns: image-based condition grading routes each item to resale, refurbishment or write-off, and that decision is where the margin is.
• Customer support automation. Deflect order-status, returns-policy and delivery queries with grounded retrieval over the actual policy documents. Route the rest by intent, and draft agent replies from account context.
• Fraud and abuse. Payment fraud, promotion abuse and serial-returner detection. Graph features across accounts, devices and payment instruments outperform per-transaction scoring.
• Marketing content production. Generate and localise product copy, ad variants, and category text from verified attributes, with brand-voice constraints and human sign-off.

One caveat on content pipelines: volume is not the objective. Thousands of near-identical generated pages produce thin content that search engines and AI assistants both discount.

Retail AI use cases: technique, data, and time to value

The table below is the short version of this article. Read the data column first.

The AI-search shift: when shoppers arrive through an assistant

The most structurally important change in retail this decade is happening upstream of your site. A growing share of shoppers start their journey with an AI assistant rather than a search-results page: they ask a question in natural language, get a synthesised answer naming a few products, and arrive already decided — or don't arrive at all. There is no page of ten blue links to rank on in that scenario, just a small set of candidates chosen by a model that read your structured data. Either your product was legible to it or it wasn't.

• Structured data becomes demand generation. Schema markup, complete attributes, real availability and specific descriptions are how you become quotable. A thin page with just three bullet points is invisible to a system that compares attributes.
• Feeds are a first-class surface. Merchant and marketplace feeds are the machine-readable version of your catalogue. Errors that used to cost you ad performance now cost you inclusion in an answer.
• Specificity beats keyword density. Assistants match on constraints: fits a 60cm gap, machine washable, suitable for sensitive skin. Marketing adjectives do not get surfaced; constrained facts do.
• Accuracy is enforced harder. An assistant that recommends an out-of-stock or mispriced item creates a bad experience that is attributed to your brand. Data freshness is now customer-facing.
• Attribution gets murkier. Assistant-mediated traffic arrives late in the funnel, with less visible provenance.

The conclusion is unusually clean: the work that improves your position in AI-driven discovery is the same catalogue enrichment that improves on-site search, recommendations and ad feeds. One programme, four returns.

How SoftmaxAI builds retail and e-commerce AI

We build our retail systems the same way we build everything else: a written, client-approved specification before any code, an agent-accelerated build led by senior engineers, and every decision that matters owned by a human.

• Data audit before use-case selection. We establish what data exists and in what state — catalogue completeness, event tracking, stock-out records, image consistency. That audit usually reorders the roadmap, because the exciting use case is often gated by a dataset nobody has.
• Spec first, with acceptance criteria. Before a model is chosen, we define what the system must do, what it must never do, how it escalates and how it is measured — pricing guardrails, personalisation boundaries and camera privacy obligations included.
• Evaluation as a build artefact. Search relevance sets, attribute accuracy against a labelled sample, forecast backtests, and task-completion tests, run on every change.
• Built to run in your cloud. Systems deploy into your own account, with monitoring, drift alerts and retraining paths. As a Google Cloud Partner we build natively on Vertex AI and Gemini where that fits.

You can see how this plays out in our case studies and the sector work on our retail and e-commerce page. If you're still working out which use case deserves the first budget, the AI Opportunity Sprint gives you a ranked shortlist, an honest feasibility assessment and an architecture before you commit to a build.