RetailTech UK AI Personalisation and Recommendation Engine

Compliant Personalisation Architecture

PECR (Privacy and Electronic Communications Regulations) consent for personalisation:

ClickMasters PECR architecture

• 1consent management platform (OneTrust -- consent captured with timestamp, IP, consent version -- separate consent for personalised recommendations vs marketing emails vs analytics),
• 2personalisation tiers (anonymous browsing -- no personal data used; email-identified -- product history only; logged-in -- full purchase history + browsing),
• 3PECR-first product recommendation (non-personalised recommendations always available -- no dark patterns forcing consent),
• 4consent withdrawal (customer withdraws consent -- recommendations revert to non-personalised immediately, event published to all downstream ML models),
• 5ICO DPIA (personalisation profiling constitutes systematic processing -- DPIA mandatory under GDPR Article 35).

Product recommendation ML

• 1item embeddings (product2vec -- product catalogue vectorised using co-purchase co-occurrence -- similar to word2vec applied to transaction sequences),
• 2user-item collaborative filtering (Alternating Least Squares -- ALS -- matrix factorisation for purchase history -- Spark MLlib),
• 3real-time recommendations API (< 40ms P99 -- AWS ElastiCache Redis + pre-computed recommendation vectors),
• 4context-aware filtering (session context -- category browsing, search term -- reranks recommendations in real-time),
• 5cold start handling (new customer -- category bestsellers until 5 interactions accumulated).

Equality Act 2010 fairness

recommendation engine tested quarterly for demographic bias -- size inclusivity (products not recommended disproportionately by size availability), price segment bias (recommendations should not systematically guide customers to higher price products without product quality justification).

AI search for UK fashion retail

• 1vector search (Weaviate -- product catalogue vectorised -- semantic search -- search for blue summer dress returns navy, cobalt, turquoise dresses),
• 2UK terminology handling (jumper vs sweater, trousers vs pants, trainers vs sneakers -- UK English model),
• 3personalised search ranking (logged-in customer -- search results reranked by purchase history and size -- logged-in personalisation gated by PECR consent),
• 4spell correction (British English -- colour, grey, licence -- standard UK retail vocabulary),
• 5zero results handling (search returns no results -- suggest nearest semantic match, not empty page).

Weaviate

open source vector database -- deployed on ECS Fargate eu-west-2 -- no third-party SaaS data transfer.

Email personalisation with PECR compliance

• 1transactional emails (order confirmation, dispatch notification -- no consent required -- legitimate interest for transaction),
• 2marketing emails (PECR s.22 -- explicit prior consent required -- separate from product recommendation consent),
• 3personalised marketing (product recommendations in marketing emails -- requires both marketing email consent AND recommendation personalisation consent),
• 4email engagement scoring (open rate, click rate by customer -- determines send frequency),
• 5unsubscribe (one-click unsubscribe -- ICO PECR guidance -- unsubscribe must work immediately).

PECR marketing consent rate

ClickMasters A/B tested 3 consent collection UX patterns -- consent captured at account creation (highest rate: 68%), consent captured post-purchase (58%), consent captured at newsletter signup (42%).