FinanceAI

RAG-Powered Support Assistant: Transforming Customer Experience

The Problem

FinanceAI was scaling fast—but so were their support costs. With 15,000+ monthly tickets and a team of 20 support agents, they were drowning in repetitive questions. “How do I reconcile this invoice?” “What’s the tax deduction for this expense?” “Why did my transaction fail?”

CEO Jennifer Chen said: “We were adding 5 new customers for every new support hire. Our growth was unsustainable.”

The Solution

Kurai deployed a production RAG system in 6 weeks:

Architecture:

• Embeddings: OpenAI text-embedding-3-small (512 dimensions)
• Vector Database: Pinecone (5M vectors, 95% recall)
• LLM: GPT-5 Turbo with function calling
• Backend: FastAPI with WebSocket support
• Frontend: React widget embedded in help center

Data Pipeline:

1. Ingestion: Scraped documentation, knowledge base, and regulatory docs
2. Chunking: Semantic chunking preserving context (avg. 500 tokens)
3. Embedding: Batch processing with OpenAI API ($0.0001/1K tokens)
4. Indexing: Upserted to Pinecone with metadata filters (category, date, jurisdiction)

Query Pipeline:

1. User question → Embed query
2. Similarity search (k=5) with metadata filter
3. Context + question → GPT-5 Turbo
4. Answer with citations → User

Guardrails:

• Fallback to human if confidence < 80%
• PII redaction before processing
• Audit logging for compliance
• Daily accuracy monitoring

The Results

Immediate Impact (Month 1):

• 9,000 tickets automated (60%)
• Avg. response time: 48 hours → 2 minutes
• CSAT score: 35% → 89%

Cost Savings (Annual):

• Support headcount: -$350K (3 fewer agents needed)
• LLM API costs: +$45K
• Net savings: $450K/year

Scalability:

• System handles 100K queries/month
• P95 latency: 850ms
• 99.94% uptime (4.3 hours downtime/year)

Customer Feedback

“Instead of waiting 2 days for an answer, I get it in seconds with links to the exact regulations. It’s like having a CPA on call 24/7.” — Michael Torres, CFO at TechScale Inc.

What’s Next

Phase 2 roadmap includes:

• Multi-language support (Spanish, French, German)
• Proactive notifications for common issues
• Integration with accounting software (QuickBooks, Xero)
• Voice-based support using Whisper API

Technology Stack

• LLM: OpenAI GPT-5 Turbo
• Embeddings: OpenAI text-embedding-3-small
• Vector DB: Pinecone (Production tier)
• Backend: Python FastAPI + Celery
• Frontend: React + TypeScript
• Monitoring: Prometheus + Grafana
• Logging: ELK Stack
• Infrastructure: AWS (ECS, Fargate, RDS)

Key Metrics

Metric	Before	After	Improvement
Monthly tickets	15,000	6,000	60% reduction
Response time	48 hours	2 minutes	99.3% faster
CSAT score	35%	89%	+154%
Support cost/ticket	$15	$3	80% reduction
Resolution accuracy	N/A	94%	New capability

Timeline

• Week 1-2: Data ingestion and preprocessing
• Week 3-4: RAG pipeline development and testing
• Week 5: Beta testing with 500 users
• Week 6: Full deployment to all 50K users

Lessons Learned

1. Data quality matters: We spent 40% of effort cleaning documentation
2. Citations build trust: Users were 3x more satisfied when sources were shown
3. Human-in-the-loop essential: 20% escalation prevents bad experiences
4. Monitor continuously: Accuracy dropped from 96% → 94% after tax law updates; auto-retraining fixed it

FinanceAI is now scaling to 100K customers without adding headcount. Their support team focuses on complex, high-value inquiries while the RAG assistant handles the rest.