AI for Retail and E-commerce
High volume, thin margins, and demand that moves weekly. Thirteen verticals, written for the person who has to defend a margin number rather than the person who enjoys a personalisation demo.
The measurement problem is the whole problem.
Retail is the sector where AI is easiest to deploy and hardest to evaluate. Almost every intervention — a new recommender, a price change, a promotion, a personalised email — moves a metric, and almost none of that movement is attributable without a holdout. Retailers have been running uncontrolled changes and reporting the results as effects for a decade, which is why the sector is full of case studies nobody can replicate.
So the discipline that matters here is not modelling. It is insisting on a control group, on incrementality rather than attributed revenue, and on measuring margin rather than conversion. A recommender that lifts conversion by promoting things people would have bought anyway has cannibalised, not created. We will design the experiment before the model, and occasionally that experiment ends the project — which is the cheapest possible outcome.
Where AI actually earns its place in retail
Ranked by evidence rather than by conference-stage frequency. The maturity column is our own read; the catch column is the part the vendor case study omits.
| Where | What it does | Maturity | The catch |
|---|---|---|---|
| Demand forecasting and replenishment | Predicts unit demand by SKU and location to drive ordering and allocation. | Proven | The most reliable value in the sector, and it only pays if the ordering process can actually act on a changed forecast. |
| Fraud and returns abuse detection | Scores orders and returns for payment fraud, promo abuse and serial returning. | Proven | Works well. The cost of a false decline is larger than most teams estimate and almost nobody measures it. |
| Product content and catalogue enrichment | Generates and normalises titles, descriptions, attributes and imagery at catalogue scale. | Proven | Immediately useful and increasingly a distribution requirement, since agent-readable attributes are now how products get found. |
| Search and merchandising | Improves on-site search relevance, ranking and category navigation. | Strong | Frequently the highest-return work in an online business and consistently under-invested relative to recommendations. |
| Recommendations and personalisation | Surfaces products a shopper is more likely to buy. | Strong | Real, and routinely over-credited. Without a holdout you are measuring what people would have bought anyway. |
| Pricing and markdown optimisation | Sets and moves prices against demand, inventory and competition. | Mixed | Effective for markdown and inventory clearance. Personalised pricing is now a live legal exposure rather than a growth tactic. |
| Conversational shopping assistants | Answers product questions and guides selection on-site. | Mixed | Useful for considered purchases with real attribute complexity. Mostly decorative on a simple catalogue. |
| Fully autonomous buying agents | Agents that complete purchases without a person in the loop. | Early | Growing fast from a very small base — AI-driven sessions are still under 0.2 per cent of e-commerce traffic. Worth preparing for, not worth rebuilding around. |
The number that should calibrate your agentic commerce plan
AI-driven visits to US retail sites grew roughly 4,700 per cent year on year in 2025, and AI-driven sessions remain below 0.2 per cent of total e-commerce traffic. Both are true and the second is the one that should size your investment this year. The sensible response is to make your catalogue agent-readable — structured attributes, accurate availability, clean feeds — which also improves conventional search and costs you nothing if agents stay small. Rebuilding your commerce stack around a channel that is currently a rounding error is a different decision, and it should be made deliberately.
No high-risk classification, and more consumer law than you think
Retail is one of the few sectors where almost nothing falls into the EU AI Act's high-risk categories. That is not the same as being unregulated — pricing, reviews, personalisation and product claims are all governed, and the pricing rules changed materially in the last year. This is our reading as at September 2026 and we work alongside your legal team rather than in place of them.
Mostly outside the high-risk categories
Recommendation, pricing, forecasting, merchandising and inventory systems are not Annex III high-risk. What does apply is Article 50 transparency where a customer interacts with an AI system or with AI-generated content, already in force since August 2026. Emotion inference in a retail setting and certain biometric applications in physical stores are a different matter and should be checked before deployment rather than after.
New York changed the rules
New York's Algorithmic Pricing Disclosure Act requires a specific on-screen statement — that the price was set by an algorithm using the shopper's personal data — with civil penalties up to $1,000 per violation, and it survived a First Amendment challenge. The follow-on One Fair Price Act, passed by both chambers in June 2026, would go considerably further and ban surveillance pricing outright, with penalties up to $5,000 for a first violation and $20,000 thereafter, effective 180 days after signature.
The FTC rule has teeth now
The Rule on the Use of Consumer Reviews and Testimonials has been in force since late 2024 and 2026 enforcement has taken shape: warning letters to targeted industries, and rule violations folded into broader deceptive conduct cases. Generated reviews, incentivised positive reviews and employee-written reviews are all in scope. If you are using language models anywhere near review content, this is the rule to read first.
GDPR, ePrivacy and consent
Behavioural personalisation depends on a lawful basis and, in the EU, on consent that survives scrutiny. The practical constraint is that consent rates determine how much of your traffic a personalisation model can even see, which is a modelling problem before it is a legal one — a recommender trained only on consenting users is trained on a biased sample.
The Digital Product Passport is coming
Under the Ecodesign for Sustainable Products Regulation, batteries carry a mandatory digital passport from 18 February 2027, with the textiles delegated act expected in late 2027 and compliance following around late 2028. Footwear is excluded from the first wave. For apparel and consumer goods brands this is a product data problem long before it is a compliance one, and the data work has a lead time measured in years.
Different obligations entirely
If you operate a marketplace rather than sell your own goods, platform regulation applies: trader traceability, recommender system transparency, notice and action, and advertising disclosure. Those are product requirements rather than legal footnotes, and they change what a ranking system is allowed to be.
What this means for a build
Two things are architectural. First, if you price dynamically anywhere in the US, you need to know whether any input is personal data about the shopper — because that single distinction separates ordinary dynamic pricing from a category New York now regulates and may ban. Second, if language models touch review, testimonial or product claim content anywhere in your stack, that is an FTC surface rather than a content operations question. Both are cheap to design for and expensive to retrofit. See AI policy development.
Who we write for
Each page starts from that organisation's own problems, names the regulatory exposure it carries, and routes into the engineering. Depth varies and is stated on each page.
E-commerce & D2C
Search, recommendations, catalogue and fraud — measured on incremental margin rather than attributed revenue.
Read more →Brick-and-mortar retail
Store-level forecasting, allocation, scheduling and shrink — with in-store vision assessed honestly.
Read more →Marketplaces
Ranking, trust and safety, catalogue matching and counterfeit detection — with ranking transparency as a design constraint.
Read more →CPG / FMCG
Demand planning, trade spend effectiveness and retail execution — where you do not own the point of sale.
Read more →Fashion & apparel
Fit and returns, visual search, short-life forecasting and markdown — plus the product data DPP will require.
Read more →Beauty & cosmetics
Shade matching, ingredient data, review integrity and trend-driven forecasting — with claims treated as regulated.
Read more →Grocery
Fresh forecasting, waste reduction, markdown and substitution — where a forecast error becomes food waste.
Read more →Consumer electronics
Attribute search, compatibility, lifecycle pricing and attach rate — a category where specs are the product.
Read more →Home & furniture
Visual and style search, visualisation, long lead-time forecasting and delivery — where returns cost a van.
Read more →Luxury goods
Authentication, clienteling and allocation — with the client relationship left firmly to the people who own it.
Read more →Sporting goods
Fit and spec guidance, seasonal forecasting and technical search — a category where the wrong spec is a return.
Read more →Pet care
Replenishment timing, life-stage matching and consumables forecasting — with pet nutrition claims treated carefully.
Read more →Subscription commerce
Churn causes, cadence, curation and payment recovery — where retention is the entire business model.
Read more →FAQ
Marked up with FAQPage schema so these answers can surface in search results and inside AI assistant responses.
Do you have retail experience?
Yes in demand forecasting, pricing and markdown, search and merchandising, catalogue enrichment and fraud. Less in physical store computer vision, and none in supply chain network design, which is an operations research discipline rather than an AI one and we will say so rather than stretch into it. Each of the thirteen vertical pages states our depth in that area rather than implying uniform expertise.
How do we know a recommender is actually working?
With a holdout, and there is no substitute. A recommender promoting products a customer would have bought anyway shows excellent attributed revenue and creates nothing. Hold out a randomised control group, measure incremental margin rather than attributed conversion, and run it long enough to cover a full purchase cycle. Every retail engagement we run includes that design, and we have recommended stopping projects on the strength of it.
Is dynamic pricing still safe to do?
Ordinary dynamic pricing — moving prices with demand, inventory, competition and time — remains standard practice. Personalised pricing, where the price differs between shoppers based on personal data about them, is now a live legal exposure: New York requires a specific disclosure with penalties per violation, and a bill banning the practice outright passed both chambers in June 2026. The design question is whether any input to your pricing is personal data about the individual shopper, and it is worth answering precisely rather than approximately.
Should we be building for AI shopping agents?
You should make your catalogue agent-readable, which means structured attributes, accurate real-time availability and clean feeds. That work improves conventional search and merchandising too, so it pays regardless of what agents do. Rebuilding your commerce stack around agent traffic is a much larger bet on a channel that grew enormously last year and still represents under 0.2 per cent of e-commerce sessions — worth watching closely, not worth reorganising around yet.
What is the most common way retail AI projects fail?
They optimise conversion and lose margin. A model that increases units sold by discounting, promoting cheaper alternatives or cannibalising full-price demand will show excellent headline numbers and a worse P&L. We insist that the objective is margin or contribution rather than conversion or revenue, and that argument is usually the most valuable thing we contribute in the first month.
Start with the problem, not the technology.
Thirty minutes, no charge, no deck. Tell us what is going wrong in your organisation and we will tell you whether AI is the right instrument — including when it plainly is not.