Gemini CLI: Multi-Agent Customer Support System with Smart Escalation v4

This workflow uses n8n 1.0+ AI Agent nodes with OpenAI GPT-4o or Claude Sonnet 4.6 to build a multi-agent support system where a triage agent classifies incoming tickets, specialist agents handle specific domains (billing, technical, account), and an escalation agent routes complex or sensitive issues to human agents. The agentic reasoning step is the triage agent's classification decision: it reads the ticket content, evaluates intent (refund request vs. technical bug vs. account access), checks confidence score, and routes to the appropriate specialist agent or human queue. This is not a rule-based routing system — the triage agent handles edge cases like multi-intent tickets (a billing question about a technical issue) by decomposing the ticket into sub-queries and routing each part independently. Each specialist agent has persistent memory via n8n Window Buffer Memory, maintaining conversation context across sessions for the same customer. Teams deploying this architecture report 75-80% first-contact resolution rates and a 40% reduction in ticket volume reaching human agents.

BUSINESS PROBLEM

A customer support team at a B2B SaaS company with 15,000+ active users receives 800+ tickets per week across billing, technical issues, and account management. The current routing system uses keyword matching: tickets containing "refund" go to billing, those with "error" go to technical, and everything else goes to a general queue. This breaks down on 30-40% of tickets that mix categories — "I was charged twice after the error on my dashboard" — causing 2-3 handoffs per ticket and 15-20 minute average resolution times. According to a 2025 McKinsey report, companies using AI-powered customer service see a 25-35% reduction in support costs and a 15-20% improvement in customer satisfaction scores. The cost of uncontained tickets is measurable: each human-handled ticket costs $8-15 in agent time at a B2B SaaS company. With 800 tickets per week and 40% requiring handoffs, the weekly cost is $6,400-12,000 just from inefficient routing and multi-touch resolution.

WHO BENEFITS

Customer support teams at B2B SaaS companies (15-50 person teams) handling 500-2,000 tickets per week who need to reduce agent workload without hiring — the multi-agent system absorbs 60-70% of tickets without human touch. Operations managers running support for marketplace or e-commerce platforms where tickets span multiple domains (order issues, payment failures, account access) and current routing rules miss 30% of correct destinations — the triage agent's semantic classification catches what keyword rules miss. Small support teams (2-5 agents) at growing startups who are drowning in ticket volume and spend more time routing than resolving — the system acts as a 24/7 triage and first-line response layer that never sleeps.

HOW IT WORKS

1. Ticket intake: An n8n webhook receives tickets from Zendesk or Freshdesk API. The webhook captures ticket ID, subject, body, customer metadata, and attachment URLs. Output: structured ticket JSON.

2. Triage classification: The triage AI Agent node receives the ticket and classifies it into one or more domains — billing, technical, account, or general. It assigns a confidence score (0.0-1.0) and flags multi-intent tickets. If confidence is below 0.6, the ticket routes to human review with a classification note. This is the agentic reasoning step. Output: classification JSON with domain tags, confidence score, and routing instructions.

3. Sub-query decomposition: For multi-intent tickets, the triage agent splits the ticket into sub-queries. Example: "I was charged after the update broke my login" produces two sub-queries — billing (charge) and technical (login break). Each sub-query is routed independently. Output: sub-query array with priorities.

4. Specialist agent response: The billing specialist agent has access to subscription databases via HTTP Request nodes. The technical specialist agent has RAG access to documentation via Pinecone/Qdrant vector stores. Each specialist generates a domain-specific response. Output: draft response per sub-query.

5. Escalation check: The escalation agent evaluates the combined draft response against escalation criteria: refunds over $500, security incidents, legal mentions, or customer sentiment flagged as angry (sentiment score below 0.3). Matches are routed to humans with a summary. Output: escalation decision JSON.

6. Response assembly and memory update: If not escalated, the responses are assembled into a single ticket reply, the conversation history is stored in Window Buffer Memory, and the reply is posted back to Zendesk/Freshdesk via API. The memory ensures the next interaction with this customer starts with full context. Output: completed ticket reply and updated memory state.

7. Human review queue: Escalated tickets arrive in the human agent dashboard with the triage classification, the specialist's draft response, and a reason for escalation. The human reviews, edits, and sends. The system logs whether the human accepted, modified, or rejected the AI draft. Output: feedback JSON that retrains agent prompts.

TOOL INTEGRATION

n8n 1.0+: Core orchestrator. Requires the Pro or Enterprise plan for AI Agent nodes at scale — the free tier limits AI node executions to 10 per workflow. Gotcha: The n8n AI Agent node's system prompt overrides are not clearly documented — when setting up the triage agent's classification prompt, use the "System Message" field on the node, not the "Prompt" template field, or the agent ignores the classification schema entirely.

OpenAI GPT-4o or Claude Sonnet 4.6: The reasoning engine for all three agent types. GPT-4o is faster for the triage step (classification in under 2 seconds). Claude Sonnet 4.6 produces better specialist responses for technical queries. Gotcha: Using different models for triage and specialist agents means prompt engineering must account for each model's instruction-following style — GPT-4o responds well to numbered formats, Claude Sonnet 4.6 prefers paragraph-style instructions.

Pinecone or Qdrant: Powers the technical specialist agent's document retrieval. Technical tickets search documentation vectors. Billing tickets do not use the vector store — they query subscription databases directly. Gotcha: The vector store must be indexed from support documentation only, not from all company docs. Indexing pricing pages alongside API docs can cause the technical agent to retrieve pricing information when answering a coding question.

Zendesk or Freshdesk API: Bidirectional ticket sync. n8n's HTTP Request node handles both inbound ticket fetching and outbound reply posting. Gotcha: The Freshdesk API rate limit is 500 requests per minute but applies per API key, not per endpoint — if your n8n instance has multiple workflows using the same key, batches of ticket updates can trigger 429 errors during peak hours.

ROI METRICS

1. Tickets resolved without human touch: 35-45% manual to 60-75% with multi-agent system. Source: Internal ticket tracking, measurable in week 1.
2. Average resolution time: 15-20 minutes to 4-7 minutes for AI-handled tickets. Source: Zendesk benchmark data, 2025.
3. First-contact resolution rate (FCR): 55-65% to 78-85% with domain-specialist routing.
4. Cost per ticket: $8-15 human-handled to $0.50-2.00 AI-handled in API costs.
5. Human agent capacity freed: 40-50% reduction in ticket volume reaching human agents, allowing a team of 10 to handle the workload of 17.

CAVEATS

1. Sentiment misclassification: The triage agent can misclassify frustrated customers as angry, routing tickets to human escalation even when a simple billing correction would resolve the issue. Tune sentiment thresholds on your actual ticket corpus, not on generic sentiment models.
2. Memory contamination: Window Buffer Memory persists across a session, but if a customer opens a new ticket about a different issue, the memory from the previous session bleeds into the new context, causing the agent to reference outdated conversations.
3. Cost unpredictability: Multi-agent systems multiply per-ticket costs. A single ticket handled by triage + specialist + escalation check generates 3-5 LLM calls. At high volume (1,000+ tickets/week), API costs can reach $500-1,500/month if not monitored with per-agent budget limits.
4. Escalation edge cases: Tickets containing vague threats, legal demands, or compliance-sensitive language may fall outside the escalation keywords and be auto-resolved when they should reach a human. Regular audit of auto-resolved tickets is essential.