Customer Service Full-Process Agent Loop
The Support Queue That Never Shrinks.
Our client, a leading SaaS platform in Southeast Asia, handles 500,000+ customer service conversations monthly across web chat, mobile app, email, and social messaging channels. Their support operation was caught in a permanent state of catch-up.
First-line agents, many of them newly hired and still learning the product, provided inconsistent answers to the same question depending on who was online. Complex issues required multiple escalation handoffs — L1 to L2 to engineering — each handoff adding hours and losing context. The knowledge base, maintained manually, was perpetually outdated as the product shipped new features faster than the documentation team could update articles.
Why 'Upgrade Your Chatbot' Doesn't Fix Customer Service.
The client had cycled through three generations of chatbot vendors. Each hit the same operational walls.
- 1.The Dead-End Conversation Problem: Chatbots could answer simple FAQs but couldn't do anything. 'How do I reset my password?' worked. 'Reset my password and send me the confirmation' did not. The bot couldn't execute actions — it could only point to help articles.
- 2.The Escalation Black Hole: When the chatbot failed, the escalation to a human agent lost all context. The customer had to re-explain their issue from scratch. The human agent had no visibility into what the bot had already tried.
- 3.The Misclassification Tax: The existing system classified tickets by keyword matching, resulting in a 30%+ misclassification rate. Billing questions landed in technical support. Feature requests were routed to the bug team. Each misroute added hours to resolution time.
- 4.The Flat Cost Problem: Every conversation — whether a simple 'what are your hours?' or a complex multi-step account migration — consumed the same AI resources at the same cost. At 500,000 conversations per month, this was economically broken.
From Chatbot to Closed-Loop Resolution Engine.
Support moves from answer generation to closed-loop resolution, escalation context, and knowledge repair.
I. Intent Routing Layer (3-tier)
L1 Classification (Haiku): Instantly categorizes queries into top-level buckets in under 200ms at minimal cost. L2 Understanding (Sonnet): Parses full intent, extracts entities, determines resolution path. L3 Action Planning (Opus, when needed): For complex multi-step issues, generates a step-by-step resolution plan.
II. Agent Execution Layer (4 Agents)
A Knowledge Retrieval Agent queries the RAG knowledge base for up-to-date documentation. A Tool Invocation Agent executes actions through MCP-connected systems — resets passwords, processes refunds, upgrades subscriptions. A Ticket Management Agent creates and updates tickets with full context. An Escalation Routing Agent transfers complex issues to human agents with a complete context package: transcript, attempted resolutions, system state, and recommended next steps.
III. Self-Healing Knowledge Layer
Bad Case Detection automatically logs failure patterns when conversations result in negative feedback or escalation. Knowledge Gap Identification flags low-confidence retrieval results for the documentation team. Auto-Update Pipeline ingests product changelogs and release notes, transforming them into knowledge base articles — keeping the system current with each product release.
IV. MCP + Token101
All agents connect to the client's existing CRM, ticketing system, payment processor, and product API through a unified MCP layer — no changes to existing systems required. Token101 tiered routing: simple queries to Haiku, standard support to Sonnet, complex decisions to Opus. Result: API costs reduced approximately 45%.
From Responding to Resolving.
"We don't just chat with customers; we close the loop from intent to resolution."
Related implementation briefs.
Claude 3 Opus
Cognitive Recommender Architecture
A Claude cognitive controller improved premium content discovery and recommendation explainability without replacing the existing high-throughput recommender stack.
Multi-Model Cascade + Edge DLP
Full-Process Medical AI
A 7-agent full-lifecycle medical AI system spanning pre-consultation, diagnostic support, drug interaction checking, and follow-up management — with HIPAA architecture and zero severe hallucinations over 6 months.
Edge SFT 7B Models
Autonomous Industrial Control Agent
An edge-deployed autonomous control system replaced hardcoded PLC logic with domain-tuned 7B models and living SOPs, enabling real-time anomaly reasoning without cloud dependency.