Case Study: GenAI product-description engine for a multi-category online retailer
A governed GenAI engine that generates SEO-ready product descriptions for a UK online retailer, grounded only in the retailer’s structured catalog attributes, with brand-tone guardrails, automated claim checks, and human approval before anything publishes.
Project details:
About the Client:
The Client is a UK mid-market online retailer with around 40,000 active SKUs across home, garden, and lifestyle products. Product copy was inconsistent across the catalog: some listings had full descriptions, many had a single supplier sentence, and newly onboarded SKUs often went live with none. Manual copywriting could not keep pace with onboarding, and agency turnaround delayed new-category launches.
Location: United Kingdom
Industry: Retail and e-commerce
Team size:6 specialists (1 AI Architect, 2 AI Engineers, 1 Backend Engineer, 1 Frontend Engineer, 1 Business Analyst)
Project duration: 4+ months
Business challenge:
The Client wanted to generate consistent, on-brand, SEO-ready product descriptions at catalog scale without publishing any claim the product data could not support, and without removing human sign-off before content went live.
Additional requirements:
- Keep generated copy inside the Client’s brand voice across every category
- Block product claims that the structured attributes do not substantiate
- Prove value on a measurable pilot before funding a full rollout
Our solution
SumatoSoft delivered a governed GenAI engine that drafts product descriptions from the Client’s own catalog attributes, checks every draft against brand and factual rules, and routes it to a human reviewer before publishing. We built it under our Agentic Development Lifecycle (ADLC), our delivery model for AI systems that have to run inside enterprise guardrails. The pilot covered three categories so the Client could measure quality and cost before scaling.
Two guardrails on every draft before a human sees it
The brand-tone guardrail checks copy against the Client’s approved style specification and rewrites or rejects off-voice phrasing. The claim guardrail compares every factual statement against the source attributes and withholds anything the data does not support, with a per-batch RAGAS-style score gating low-quality batches out of review.
Additional features:
- Grounded generation from structured attributes only
- SEO structure with title, key features, and long copy
- Human accept, edit, and reject on every draft
- VPC-isolated model with zero data retention
- PIM integration for attribute intake and copy export
Business value
Before:
- Newly onboarded SKUs often went live with no description, and many older listings carried a single supplier sentence.
- Manual and agency copywriting could not keep pace with onboarding, delaying new-category launches by weeks.
- Tone varied across thousands of listings, weakening the brand on the storefront.
- Some supplier-supplied claims could not be substantiated against the Client’s own product data.
- Content production cost scaled directly with catalog growth.
After:
- The engine generates a structured description for every product from its catalog attributes, including newly onboarded SKUs, with 91% approved by the content team with light or no edits.
- Average copy production time per product fell 58%, and new-category launches no longer wait on agency turnaround.
- Generated copy follows one approved brand-voice specification across every category.
- The claim guardrail withholds any statement the attributes do not support, and no flagged unsupported claim reached the live storefront during the pilot.
- Content production scales with the engine instead of with headcount, with cost per description reviewed weekly through the pilot.
