Ecommerce · Utrecht, Netherlands
AI-Assisted Merchandising for a Dutch Retail Group
Glassorchard Retail B.V.
Glassorchard Retail B.V. combined CRO with a production AI assistant grounded in product and policy data — not open-ended chat.
Client: Glassorchard Retail B.V. · Location: Utrecht, Netherlands · Industry: Ecommerce
The problem
Glassorchard’s catalogue was deep, but shoppers bounced when sizing, returns, or compatibility questions weren’t answered quickly. Support answered the same policy questions dozens of times a day. Leadership wanted AI — without hallucinated promises.
- PDP modules not tested systematically
- Support ticket themes mirrored unanswered on-site questions
- Previous chatbot pilot invented return windows
- Merchandising rules hard for humans to maintain at SKU scale
The solution
We ran a disciplined CRO program on PDPs while deploying an assistant that could only answer from approved product specs and policy documents, with human escalation paths.
Our work
Product, CX, and merchandising joined a shared instrumentation plan so experiments and AI answers used the same truth.
Research & instrumentation
Session replays, ticket taxonomy, and event schema for PDP modules and assistant deflections.
CRO test program
Tested fit guides, delivery promises, and proof modules. Launched winners into the design system.
Grounded AI assistant
Indexed approved docs/specs, built refusal behavior for out-of-scope questions, and routed edge cases to humans.
Merchandising assists & handover
Recommended related products with merchant override. Trained CX on transcript review rituals.
Team time invested
Hours include AI engineering, safety review, and CRO engineering — not a thin chatbot install.
| Role | Focus | Hours |
|---|---|---|
| CRO lead | Hypothesis backlog, analysis | 160 |
| UX / UI designer | PDP modules, assistant UI | 140 |
| Frontend developers | Experiments, UI integration | 220 |
| AI / platform engineers | Retrieval, guardrails, logging | 320 |
| Content / knowledge ops | Policy corpus, SKU facts | 120 |
| CX lead (Flexus) | Escalation design, training | 80 |
| Analytics | Dashboards, experiment QA | 40 |
| PM | Risk reviews, timeline | 20 |
Total Flexus effort: ~1,100 hours across 5 months.
Results
Sitewide conversion rate improved 22%. Policy-style support tickets declined as the assistant handled first-line answers with approved language only.
- +22% sitewide conversion rate
- Measurable deflection on repeat policy tickets
- Zero critical hallucination incidents in launch window (monitored)
- Merchants retaining override control on recommendations
“The AI only answers what we approve. That’s why we let it onto the storefront.”
— Digital director, Glassorchard Retail B.V.