ZapytAI

This project delivers a conversational AI chatbot for consultants working within the Sales-service system, supporting both Shop consultants (TMPL shops & Agents shops) and CC consultants (internal & external). The chatbot is designed to reduce the time consultants spend searching for information and to improve the consistency and accuracy of answers provided to customers, by giving direct, source-linked responses drawn exclusively from a curated internal knowledge base. The solution is built on a Retrieval-Augmented Generation (RAG) architecture using Claude (Anthropic) as the underlying LLM, leveraging enterprise licenses already deployed within TMPL. The chatbot retrieves relevant content exclusively from Clipper, the company's internal knowledge base. Importantly, the chatbot has no access to public internet sources, ensuring that all responses are grounded strictly in vetted, company-approved content. When a question falls within scope, the chatbot identifies the most relevant chunks from the vector database, generates an answer, and provides a direct link to the source article in knowledge base. When a question falls outside its defined knowledge scope, the chatbot responds with a predefined, transparent message rather than attempting to answer — an approach designed to maximize trust and prevent misinformation. Access control is a core design element: users authenticate via specific roles, and visibility into knowledge base content is further restricted by consultant channel (Shops, APS, CC, ECC) and by internal distribution lists, ensuring that each consultant only sees content relevant and authorized for their role. Because articles are updated at irregular intervals, the system includes automated daily synchronization to keep the chatbot's underlying knowledge current, along with an administrative tool for content owners to monitor the state of the knowledge base and track recent updates. To support continuous improvement, every interaction is captured and analyzed through Langfuse: users can rate each chatbot response with a simple like/dislike, and dislikes are further categorized through a predefined list of reasons (5–6 categories plus a free-text "other" option). This feedback, together with usage data (per user_login), response latency, and article-category breakdowns, feeds into a statistics dashboard that helps the business identify heavy users (for cost-efficiency purposes), track quality trends, and prioritize knowledge base updates. From a non-functional standpoint, the system is designed to support up to ~3,000 concurrent users, with a target response time of up to 9 seconds, and horizontal scalability as usage grows. The chatbot maintains a neutral, non-discriminatory tone across religious, political, and gender-related topics, and explicitly avoids answering sensitive, non-public, or unusual questions outside its defined scope. It is available daily from 7 a.m. to 11 p.m., including weekends (reflecting CC and Shops consultants' working hours), with maintenance handled during standard business hours (9 a.m.–5 p.m.). Responses are streamed in real time to improve perceived performance, and the interface — built within the existing Sales-service UI — is designed to be intuitive, minimizing the learning curve for consultants already familiar with the platform. Overall, the project establishes a scalable, compliant, and continuously improvable AI assistant that enhances consultant productivity while maintaining strict control over data sources, user access, and content accuracy — with the pilot serving as the foundation for broader rollout across additional article categories and consultant channels.

Q2 2026 vs Q4 2025: -49 sec average handling time +0.8 pp first contact resolution

Shop consultants (TMPL shops & Agents shops) and CC consultants (internal & external)

Our team combines hands-on AI experience with deep business knowledge. This unique combination allows us to move quickly from concept to prod (3000 users), without the overhead of vendor selection, new integrations, or unfamiliar tooling.