eCommerce AI Customer Service Platform — UK D2C Brand

RAG Customer Service AI

Retrieval-Augmented Generation (RAG) architecture: product knowledge base (400+ product pages, FAQ, sizing guides), order data (via Shopify API — order status, tracking, return eligibility), and returns policy documents.

Vector store

pgvector extension on PostgreSQL.

LLM

Claude claude-sonnet-4-20250514 via Anthropic API.

Context injection

customer's order history retrieved from Shopify and injected into context.

Response

AI generates personalised response with specific order details.

AI confidence threshold

• responses with confidence &lt
• 0.7 escalated to human agent queue.

Escalation triggers

• complaint language (keywords: "terrible", "disgusted", "Trading Standards"), order value &gt
• £500, return after 30 days (Consumer Rights Act repair/replacement rights — human judgement required), and any mention of personal injury.

Human handoff

full conversation context passed to agent — no repeated customer explanation required.

ICO AI guidance

decisions affecting customers must be explainable and challengeable.

AI responses

labelled as AI-generated, with option to request human review.

Right to human review

GDPR Article 22 applies if AI makes a decision with significant effect — refusal of return claim escalated to human if customer requests.

Data minimisation

order data fetched per session, not stored in AI conversation logs beyond 30 days.

AI hard rules for Consumer Rights Act compliance

• automatic escalation to human for any request involving: statutory return rights (within 14 days, no reason required), repair/replacement rights (within 6 months — fault must be proven), refund requests for digital content (immediate statutory right), and product safety concerns (OPSS escalation required).
• AI cannot refuse a statutory right — it can only acknowledge and escalate.