Customer support AI copilot — FinPálya Kft., −62% AHT

The challenge

FinPálya Kft. (~150-person Hungarian B2C fintech) was handling 8,500 support tickets/month with 11-min average handle time and 25-min peak wait. CSAT was declining and hiring couldn't keep up with traffic.

Targets

Three quantitative goals: (1) 40%+ AHT reduction, (2) >60% self-service on simple questions, (3) zero hallucinations on regulated responses. Timeline: 10 weeks PoC to launch.

Why UseAIEasily?

Budapest-based, knows Hungarian banking support vocabulary, integrates with zero-retention Anthropic Claude in EU region, senior engineering team. Big-4 alternative quoted €400k; we scoped fixed-price at €78k.

Architecture

Hybrid RAG pipeline: multilingual-e5-large embeddings over ~12k Hungarian documents, Qdrant self-hosted EU, BM25 full-text filter, Cohere Rerank, Claude Sonnet 4.6 with citation-bound responses. Human-in-the-loop review on all financial-advice adjacent responses. PII redaction layer before the prompt. Full audit trail in LangSmith.

Delivery

Phase 1 (discovery, 2 weeks), Phase 2 (PoC, 3 weeks), Phase 3 (production, 5 weeks): CRM integration (Freshdesk), admin UI, monitoring, cost caps, A/B test.

Results (6 months post-launch)

• Average AHT: 11 min → 4.2 min (−62%)
• Self-service on simple queries: 12% → 87%
• CSAT: 3.8/5 → 4.6/5
• Escalation to human agent: 88% → 31%
• Team expansion avoided: 2 planned hires cancelled (~€60k/year savings)
• Hallucination incidents: 0 on regulated responses (6 months)

Lessons

The key to RAG success was the 150-question golden eval set (human-reviewed), not the model choice. Cohere Rerank alone boosted accuracy 18% over pure vector search.

Costs

Build fee: €78,000 fixed-scope. Monthly runtime: ~€1,400. Payback: ~7 months via headcount savings and productivity gains.