April 29th, 2026 at 09:38 am
Why AI Is Now the Baseline for Competitive E-Commerce
Amazon’s product recommendation engine generates 35% of the company’s total annual revenue. That is not a side feature. It is a core infrastructure layer that influences what every user sees, what they click, and what they buy — on every visit, across every device, in real time.
Netflix saves an estimated $1 billion per year in subscriber retention through personalised content recommendations. Shopify merchants using AI-powered product suggestions see an average 26% uplift in average order value compared to those using static featured products. ASOS’s visual search feature drove a 48% increase in search-to-purchase conversion among users who engaged with it.
These are not exceptional results from companies with exceptional AI budgets. They are the compounding returns of specific, well-executed AI features applied to the highest-leverage points in the e-commerce funnel. And in 2026, the technology that powers them is accessible to any mobile commerce app with a decent data layer and a development team that knows how to build it.
The question for UK e-commerce app teams is no longer whether to invest in AI. It is which features to prioritise, in what order, and with what technology. This guide answers all three.
The market pressure:
UK online retail reached £117 billion in 2025. Mobile commerce accounts for 63% of all e-commerce transactions. Consumers who experience personalised AI-driven shopping are 76% more likely to make repeat purchases within 30 days than those who see generic product feeds. The apps that do not personalise are not standing still — they are falling behind an expectation baseline that compounds every quarter.
2. AI Features Every E-Commerce App Should Consider
These seven features cover the full e-commerce AI stack — from the moment a user opens the app to the moment they check out. Each is production-ready in 2026 using available APIs and tooling. They are ordered by typical ROI and implementation speed, not complexity.
- Product recommendation engine
| What it does | Surfaces personalised product suggestions — ‘You might also like’, ‘Frequently bought together’, ‘Because you viewed X’ — based on each user’s browsing, purchase, and wishlist history, plus behavioural patterns from similar users | Build Low–Medium |
| Business impact | Amazon attributes 35% of revenue to its recommendation engine. Shopify merchants using AI recommendations report 26% higher average order value. For a UK e-commerce app doing £2M annually, a 15% AOV improvement from recommendations is worth £300,000/year. | |
| How to implement | Start with AWS Personalize or Google Recommendations AI — both are managed ML services that handle model training, updating, and inference without requiring an ML team. Feed them purchase events, view events, and product catalogue. For custom builds: collaborative filtering + content-based hybrid model, retrained weekly. Minimum viable dataset: 10,000 purchase events across 1,000+ users. | |
- AI-powered semantic search
| What it does | Search understands the meaning behind queries, not just the keywords. A user searching ‘something to wear to a summer wedding that’s not too formal’ gets relevant results even if those exact words are not in any product description. Handles synonyms, intent, and natural language queries natively. | Build Medium |
| Business impact | ASOS reports a 48% higher search-to-purchase conversion for semantic search vs keyword search. Shopify data shows semantic search users have 2.3x higher average session value. For high-SKU catalogues (5,000+ products), semantic search typically pays back its build cost within 60–90 days of launch. | |
| How to implement | Embed your product catalogue using OpenAI text-embedding-3-large or an open-source equivalent (all-MiniLM-L6-v2). Store embeddings in a vector database — pgvector for catalogues under 100K products, Pinecone or Weaviate for larger. At query time, embed the user’s search query and retrieve semantically similar products. Re-rank results using user preference history for personalisation. | |
- Visual search
| What it does | Users take or upload a photo and the app finds visually similar products in your catalogue — by colour, shape, style, or pattern. Particularly powerful in fashion, home décor, and beauty, where ‘I want something like this’ is the natural search behaviour. | Build Medium–High |
| Business impact | Pinterest Lens drives 600 million visual searches per month. ASOS visual search users convert at 3.5x the rate of text search users. Visual search is the fastest-growing search modality in fashion e-commerce, with mobile adoption growing 47% year-on-year from 2023 to 2025. | |
| How to implement | Google Vision AI or AWS Rekognition for image feature extraction. Build a visual similarity index using image embeddings stored in a vector database. The complexity is in the image processing pipeline and embedding quality — generic vision models work well for broad categories; fashion-specific models (trained on apparel datasets) perform significantly better for clothing and accessories. | |
- Dynamic pricing and personalised offers
| What it does | Prices, promotions, and offers adapt based on demand signals, inventory levels, competitor pricing, user purchase history, and predicted price sensitivity — without requiring manual markdown decisions. | Build High |
| Business impact | Dynamic pricing typically delivers 5–12% revenue uplift for retailers who implement it well. Personalised offer targeting (showing the right promotion to the right user at the right moment) reduces promotional spend waste by 30–40% compared to blanket promotions. | |
| How to implement | UK Consumer Rights Act constraints apply — prices must not be deceptive or unlawfully discriminatory. Best practice: personalise the offer (bundle, timing, framing) rather than the base price. Demand-based pricing (adjusting prices based on stock levels and competitor prices) is lower regulatory risk than user-profile-based pricing. Requires a real-time pricing API integrated into the checkout flow and a monitoring layer to catch unintended pricing outcomes. | |
- AI customer service
| What it does | LLM-powered assistant handles order tracking, returns, size advice, product queries, stock availability, and gift recommendations in natural language — without human involvement for tier-1 queries. | Build Medium |
| Business impact | Klarna’s AI assistant handles the equivalent of 700 full-time agent workloads monthly. For a UK e-commerce app with 50,000 monthly support contacts at £4/contact average cost, an AI containment rate of 65% saves approximately £130,000 per year in support cost. | |
| How to implement | LLM API (GPT-4o or Claude 3.5) with retrieval-augmented generation — the LLM’s responses are grounded in your product catalogue, order management system, and returns policy via API. Order status queries require a real-time integration with your OMS. Returns initiation requires a backend action trigger. Do not let the LLM generate order details from general knowledge — RAG is not optional for an e-commerce context. | |
- Inventory and demand forecasting
| What it does | ML models predict future demand for each SKU based on historical sales, seasonal patterns, promotional calendars, and external signals (weather, events, trends) — enabling proactive stock management rather than reactive reordering. | Build High |
| Business impact | Accurate demand forecasting reduces stockout events by 30–50% and overstock costs by 20–35%. For a UK retailer with £500K in annual inventory carrying costs, a 25% reduction in overstock represents £125,000 in direct cost savings. | |
| How to implement | Time-series forecasting models (Prophet, ARIMA, or neural network approaches like N-BEATS) trained on historical sales data per SKU. Requires at minimum 12 months of historical sales data with consistent SKU tracking. Can be implemented as a standalone forecasting service that feeds reorder alerts to the operations team — does not require full OMS integration for an MVP. | |
- Personalised push notifications and re-engagement
| What it does | AI determines the right time, right content, and right channel to re-engage each user — based on their session timing, engagement history, and predicted next purchase window. Replaces broadcast promotional messaging with individually timed, content-relevant notifications. | Build Low–Medium |
| Business impact | Personalised push notifications achieve 4–7x higher open rates than broadcast notifications. Apps using AI-timed notifications (sent at the individual user’s historically most responsive time) see 23% higher conversion from notification to purchase compared to fixed-schedule sends. | |
| How to implement | Track notification open rate and resulting session for each user. Train a per-user optimal send-time model (binned hour-of-day weighted by recency). Segment notification content by user purchase category history and browsing recency. Infrastructure: Firebase Cloud Messaging or OneSignal. Content personalisation can be LLM-generated (product-specific copy) or rule-based (category-targeted templates) depending on budget. | |
3. AI Sales Assistants: The Future of In-App Selling
Beyond individual AI features, a new product category is emerging in e-commerce: the AI sales assistant — a conversational agent that guides users through product discovery, selection, and purchase in a way that mirrors the best in-store sales experience.
What conversational commerce looks like in 2026
The AI sales assistant is not a chatbot with a list of product links. It is a conversational interface that understands context, remembers previous interactions, asks clarifying questions, makes confident recommendations, and adapts its suggestions based on the user’s responses — in natural language, on mobile, in real time.
- Zara’s AI styling advisor lets users describe an occasion (‘something for a business dinner in Paris in October’) and generates a curated outfit from the current collection, with the ability to swap individual items and adjust for budget
- ASOS’s AI product finder guides users through a structured but conversational discovery flow — asking about occasion, style preference, and budget before surfacing personalised results, achieving 3.5x higher conversion than standard category browsing
- Shopify’s Sidekick allows merchants to interact with their own store data conversationally (‘what are my best-performing products this month?’ ‘draft a promotional email for my summer collection’) — a model that consumer-facing Shopify stores are beginning to replicate for their own customers
- Amazon’s Rufus AI shopping assistant (launched in the UK in 2024) handles natural language product questions, comparison queries, and gift recommendations directly in the search interface — signalling that conversational product discovery is moving from differentiator to expectation across the sector
Why AI sales assistants convert
The conversion advantage of AI sales assistants comes from two compounding effects. First, they reduce decision paralysis — a catalogue of 50,000 products is overwhelming; an assistant that understands your need and narrows to three relevant options is not. Second, they replicate the dynamic of a trusted product advisor — asking the right questions, giving confident recommendations, and adjusting based on feedback — which research consistently shows produces higher conversion and higher AOV than unguided browsing.
The Nordstone view on AI sales assistants:
The teams that build AI sales assistants well invest as much in conversation design as they do in technology. The quality of the assistant’s questions, the confidence of its recommendations, and the naturalness of its language are product decisions, not engineering decisions. Ship the technology and the conversation design together — one without the other does not convert.
4. What Does It Cost to Add AI to an E-Commerce App?
Costs vary significantly by feature type, implementation approach, and whether you are integrating with an existing platform (Shopify, WooCommerce) or a custom-built commerce stack. Here is a realistic breakdown for 2026.
| AI feature | Platform integration (Shopify / WooCommerce) | Custom app integration | Timeline |
|---|---|---|---|
| Product recommendations | £6,000 – £18,000 (plugin + API config) | £18,000 – £45,000 (custom model) | 3–8 weeks |
| Semantic search | £8,000 – £22,000 (search API integration) | £20,000 – £55,000 (embedding pipeline) | 5–10 weeks |
| Visual search | £12,000 – £28,000 (vision API + UI) | £30,000 – £70,000 (custom similarity index) | 8–14 weeks |
| AI customer service (LLM) | £15,000 – £35,000 (LLM + OMS integration) | £25,000 – £60,000 (full custom build) | 6–12 weeks |
| Dynamic pricing engine | £10,000 – £25,000 (pricing rules + API) | £30,000 – £80,000 (ML pricing model) | 8–16 weeks |
| Demand forecasting | Not typically platform-integrated | £35,000 – £90,000 (full ML pipeline) | 10–20 weeks |
| AI sales assistant (full) | £20,000 – £45,000 (LLM + catalogue RAG) | £45,000 – £120,000 (custom conversational AI) | 10–18 weeks |
| Personalised push notifications | £5,000 – £12,000 (timing model + FCM) | £10,000 – £28,000 (full personalisation layer) | 3–6 weeks |
API-first vs custom model: which is right for e-commerce?
For most e-commerce apps, the API-first approach delivers 80–90% of the value at 20–30% of the cost of a custom model — at least for an initial deployment. AWS Personalize, Google Recommendations AI, and Algolia NeuralSearch are production-grade AI services that can be integrated with an existing Shopify or custom commerce stack without requiring a dedicated ML team.
Custom model development becomes the right choice when you have a large proprietary dataset (500,000+ user interactions), a specialised product domain where general models underperform (highly technical products, niche fashion subcategories, B2B catalogues), or a scale where API costs exceed the amortised cost of running your own model. Plan the migration path from API to custom from day one — the architecture should support it without a rewrite.
5. Getting Started: A Phased Approach to AI in E-Commerce
The most common mistake e-commerce teams make when approaching AI is trying to build everything at once. The right approach is a phased deployment that starts with the highest-ROI, fastest-to-ship feature and uses the revenue impact to fund subsequent phases.
Phase 1: Product recommendations (weeks 1–8)
Start here. Product recommendations have the highest ROI-to-complexity ratio of any AI feature in e-commerce — they are well-understood technically, accessible via managed APIs (AWS Personalize, Google Recommendations AI), and their impact is directly measurable in AOV and repeat purchase rate. Implement ‘You might also like’, ‘Frequently bought together’, and ‘Because you viewed X’ on product pages and in the cart. Measure impact with a holdout group. Use the revenue uplift to fund Phase 2.
Phase 2: Semantic search (weeks 8–18)
Once recommendations are live and measured, move to search. Semantic search delivers the second-highest conversion impact and is particularly high-value for catalogues over 1,000 SKUs where keyword search fails to surface relevant products. Build on the event tracking infrastructure from Phase 1 — search personalisation uses the same user behaviour signals. Measure: search-to-purchase conversion rate and zero-results-found rate (semantic search should drive this to near zero for in-catalogue products).
Phase 3: AI customer service and predictive features (weeks 18–30+)
With recommendations and search generating measurable ROI, Phase 3 adds the higher-complexity features: AI customer service (reduces support cost immediately), personalised push notifications (reactivates lapsed users), and — where the data and business case support it — demand forecasting and dynamic pricing. Phase 3 features have longer payback periods but generate the largest long-term structural advantages.
Phase 4: AI sales assistant (optional, for differentiation)
The AI sales assistant is not a commodity feature — it requires investment in conversation design and LLM integration that goes beyond the managed API approach of Phases 1–3. Build it when you have the data to ground it (purchase history, product catalogue with rich metadata, user preference signals from Phases 1–3) and the product strategy to position it as a meaningful part of the shopping experience. Done well, it creates a defensible differentiation that competitors cannot replicate with a plugin.
Frequently Asked Questions
What are the most impactful AI features for an e-commerce app?
The five AI features with the most consistently measurable commercial impact in e-commerce are: product recommendations (average 15–35% AOV uplift), semantic search (average 30–50% search-to-purchase conversion improvement), AI customer service (average 60–70% query containment, reducing support cost significantly), personalised push notifications (4–7x higher open rate vs broadcast), and visual search (highest conversion uplift in fashion and home décor verticals). For most apps, building in this order — recommendations first, then search, then the others — delivers the fastest measurable ROI.
What is an AI sales assistant for e-commerce?
An AI sales assistant is a conversational interface within an e-commerce app that guides users through product discovery and purchase using natural language — asking about their needs, making confident product recommendations, answering questions, and adapting based on user feedback. It is powered by a large language model (GPT-4o, Claude, or equivalent) with the product catalogue integrated via retrieval-augmented generation so the AI can make accurate, product-specific recommendations. Leading examples include Amazon Rufus (launched UK 2024), ASOS’s product finder, and Zara’s AI styling advisor.
How do I add AI recommendations to an existing Shopify store’s mobile app?
The fastest route is integration with a managed recommendations API. AWS Personalize and Google Recommendations AI both have Shopify connectors that feed on order and browsing event data without requiring custom ML development. A Shopify app with basic AI recommendations can typically be live within 3–6 weeks using this approach, at a cost of £6,000–£18,000 depending on integration complexity. Custom recommendation models — built on your proprietary data using collaborative filtering or a hybrid approach — deliver better performance for large catalogues but require 8–14 weeks and £20,000–£45,000 to build.
Is AI e-commerce personalisation compliant with UK GDPR?
Yes, with the right legal basis and transparency. Personalising product recommendations based on purchase and browsing history is generally permissible under legitimate interests — provided users are informed and given a clear opt-out. For more intrusive personalisation (location-based, cross-device tracking, special category inferences), explicit consent is required. UK GDPR requires a clear privacy notice explaining how behavioural data is used, data retention limits, and a mechanism for users to see and delete their preference data. Build GDPR compliance into the data pipeline from day one — retrofitting it is significantly more expensive.
Ready to add AI to your e-commerce app?
Nordstone builds AI-powered mobile and web applications for UK e-commerce brands, marketplaces, and retail technology companies. We have delivered recommendation engines, semantic search, AI customer service, and conversational commerce features — from Shopify integrations to fully custom mobile app builds. Tell us what you are building and we will scope the right approach.
