Home/Work/ Statista
Research · Germany · AI

A conversational AI
agent that does the job.

Retrieval-augmented agent for tier-1 support and research intake at Statista, with full audit trails for compliance. 84% containment rate, −71% average handle time, CSAT 4.7/5.

ClientStatista GmbH
SectorResearch · Germany
Duration12 weeks
PracticeAI & Machine Learning
84%
Containment rate
−71%
Average handle time
4.7/5
User CSAT score
The challenge

200+ daily support queries. Each one needing a different data source.

Statista's support team handled 200+ queries daily spanning subscription access, data methodology questions, custom data requests and citation guidance. Each query type required searching a different internal knowledge source — help articles, methodology docs, subscription databases and internal pricing guides.

Tier-1 agents were spending 60–70% of their time on queries that could be answered from documented sources, leaving senior research staff to handle overflows rather than high-value client work.

The additional constraint: Statista operates in three languages (EN, DE, FR) and has strict compliance requirements around what information can be shared with which subscription tier — the agent needed to be aware of customer entitlements before answering.

Our approach

RAG with entitlement-aware retrieval and audit logs.

Phase 01 · Weeks 1–3

Knowledge Base Vectorisation

Ingested 4,000+ help articles, methodology PDFs and internal FAQs. Tagged each chunk with entitlement level (free, basic, enterprise) to enforce access control at retrieval time.

Phase 02 · Weeks 3–7

RAG Agent Architecture

GPT-4o with multilingual instruction tuning, custom retrieval chain that checks customer entitlement before surfacing content, citation tracking so every answer references its source document.

Phase 03 · Weeks 7–10

Escalation & Audit System

Confidence-threshold-based escalation to human agents with full conversation context. Immutable audit log of every query, retrieved chunk and generated answer for compliance review.

Phase 04 · Weeks 10–12

Integration & A/B Launch

Embedded in Statista's Zendesk workflow. A/B rollout: 30% of inbound queries routed to agent for 2 weeks before full launch. Monitored CSAT, containment and escalation rate daily.

Need an AI agent that handles your support load?

Book a discovery call →
Tech used
Python GPT-4o LangChain Pinecone FastAPI Zendesk API PostgreSQL
← All case studies