← Work

AI Copilot for Customer Support

Customer Support Copilot

A RAG-based AI copilot for financial advisory support

SCB 10X is the innovation arm of SCBX, the parent group of Siam Commercial Bank, Thailand’s largest bank by market cap at $10B+. Its financial advisory customer service team handled 15,000 tickets a day (~100 per day per agent), with predominantly manual processes.

I delivered a RAG-based AI copilot to production with AI engineers and the client team. It cut information retrieval time for 150+ agents from days to under 5 minutes.

Role
Senior Technical Program Manager · 2023-2024
Industry
Financial Services / Financial Advisory
Tech Stack
LangChain · Pinecone · LangFuse · Salesforce CRM and Knowledge Base
Skills
Product Management · User Research · Evaluation

Problem

One ticket could take hours, or days if it is a cross-department question.

  • Scattered knowledge: Information lived across PDFs, the knowledge base, Salesforce CRM, and shared drives, all with inconsistent formats.
  • Manual ticket summaries: Agents wrote every summary by hand after each chat. Information sometimes got lost, was inaccurate, or incomplete.
  • Strict financial regulations: Every design choice had to comply with SEC Thailand regulations and data privacy policy.

Solution

A RAG-based AI copilot designed for customer service agents to streamline information retrieval and time to resolution with human oversight.

Remark: A mockup, since the real product screens are confidential. Conversation translated into English.
  • Real-time knowledge retrieval: Semantic search across Salesforce knowledge (policies, troubleshooting guides, regulatory docs) and customer data. Sources appear next to each answer so agents can review them.
  • Reply suggestions: The copilot drafts three replies per message, each grounded in retrieved sources. Agents pick one, edit, and send.
  • Auto-generated ticket summaries: The system writes a summary and next steps back to Salesforce after each conversation.

System architecture

CS agent queryCustomer questionPII redactionMask sensitive dataLangChain orchestrationQuery routing + prompt logicLLM generates three answersDomain-specific, Thai languageSalesforce (output)Chat logs, summaries, to-do itemsPineconeVector searchSalesforce KBPolicies, guidesLangFuseMonitoring

PDPA-compliant PII redaction: Sensitive customer data was masked before routing to the LLM, then re-injected back into the final output.

Model evaluation: Automated retrieval tests and LLM-as-judge scoring against a golden dataset built with support leads, plus LangFuse tracing and agent feedback to improve each iteration.

Result: Information retrieval time dropped from days to under 5 minutes for 150+ agents.

← Back to Work