Client
SIKO is one of the largest bathroom and kitchen retailers in the Czech Republic and Slovakia. It operates an extensive network of physical stores and the www.siko.cz e-commerce site with tens of thousands of products.
Challenge
Customers shopping for bathroom products online need to navigate a catalog containing tens of thousands of SKUs. Standard search and filters work well when customers know exactly what they need. Many visitors, however, arrive with a vague or complex intent: “I’m renovating a bathroom,” “something in black,” or “what goes with this fixture?” Search and filters alone are insufficient for these types of questions. SIKO therefore wanted to provide contextual guidance in real time, directly on the website, similar to a consultation with a salesperson in a physical store.
Solution
Aspecta designed and built an AI shopping assistant (“Kachnička, AI rádce”), a conversational guide integrated directly into siko.cz. The assistant:
- understands natural language and can handle both vague and complex questions,
- recommends specific products from SIKO’s actual catalog through a connection to the product feed,
- routes service and complaint-related questions to the appropriate channel,
- guides customers through multiple steps of a conversation.
Technology foundation: the solution runs on Amazon Bedrock with Claude (Anthropic) models. It uses a RAG layer over the product catalog and documentation, along with security guardrails to prevent misuse.
Project delivery
| Phase | Scope |
|---|---|
| Proof of Concept | Validation of product recognition and recommendation over the product catalog. |
| MVP development | Conversational widget, catalog connection, RAG, guardrails, and website integration. |
| QA and tuning | Structured testing of response quality, scenarios, and prompt iterations. |
| Production deployment | Launch on www.siko.cz and ongoing optimization based on real-world data. |
| Data analytics | Evaluation of real conversations and recommendations for further development. |
Results
An evaluation of live operation using a cleaned dataset covering 18 days showed:
- 323 unique visitors and 362 conversations in less than three weeks, through a single channel and without significant promotion of the widget.
- More than 1,000 unique products (SKUs) were actively recommended to customers. The recommendations covered a broad part of the catalog, not only bestsellers.
- Approximately 87% of conversations involving multiple questions ended with a specific product recommendation.
- No detected hallucinations. The assistant consistently provides correct contact details and, in approximately 26% of responses, states that it does not know the requested information.
- Resistance to misuse: the assistant resisted an attempted “prompt injection” and remained within its assigned role.
- Response time: the average time to first response was measured in seconds for most conversations.
- Engagement among the most active users: the most active customers had conversations that lasted around 16 minutes on average, comparable to a consultation in a physical store.
Insights from conversations
Real customer conversations provided insights relevant to SIKO’s purchasing and marketing teams. These included a clear trend in demand for black finishes compared with traditional chrome and strong brand recall for SIKO’s private label. A standard e-commerce site does not typically capture this information. Conversations with the assistant provide it as an additional data output.
Scope of the solution
The original brief for a “website chat” expanded to include product discovery, service routing, and data inputs for purchasing and marketing. The project confirmed that a well-designed AI assistant can operate reliably and securely in retail while delivering measurable value.
Technologies used
Amazon Bedrock · Claude (Anthropic) · RAG over the product catalog · conversational web widget · security guardrails · conversation data analytics
