Designing a Customer Support AI Agent: Architecture and Conversation Flow

Why Customer Support Is Perfect for AI Agents

Customer support follows predictable patterns. Most inquiries fall into a small set of categories — order status, billing questions, technical troubleshooting, returns. Within each category, the resolution path is well-defined. This structure makes support an ideal domain for agentic AI, where an agent can classify intent, retrieve relevant information, execute actions, and escalate only when necessary.

A well-designed support agent reduces average handle time by 60-80% for routine queries while preserving human intervention for complex cases. The key is getting the architecture right from the start.

Core Architecture

A production support agent has four layers: intent classification, knowledge retrieval, action execution, and escalation management. Each layer feeds into the next, and the agent orchestrates them within a conversation loop.

flowchart LR
    USER(["Customer"])
    CHANNEL{"Channel"}
    CHAT["Chat agent"]
    VOICE["Voice agent"]
    EMAIL["Email agent"]
    TRIAGE["Triage and<br/>intent detection"]
    KB[("Knowledge base<br/>RAG")]
    CRM[("CRM context")]
    AUTORES{"Auto resolvable?"}
    RESOLVE(["Resolved with<br/>cited answer"])
    HUMAN(["Tier 2 agent"])
    USER --> CHANNEL --> CHAT --> TRIAGE
    CHANNEL --> VOICE --> TRIAGE
    CHANNEL --> EMAIL --> TRIAGE
    TRIAGE --> KB
    TRIAGE --> CRM
    TRIAGE --> AUTORES
    AUTORES -->|Yes| RESOLVE
    AUTORES -->|No| HUMAN
    style TRIAGE fill:#4f46e5,stroke:#4338ca,color:#fff
    style AUTORES fill:#f59e0b,stroke:#d97706,color:#1f2937
    style RESOLVE fill:#059669,stroke:#047857,color:#fff
    style HUMAN fill:#0ea5e9,stroke:#0369a1,color:#fff

from dataclasses import dataclass, field
from enum import Enum
from typing import Optional

class Intent(Enum):
    ORDER_STATUS = "order_status"
    BILLING = "billing"
    TECHNICAL = "technical"
    RETURNS = "returns"
    GENERAL_FAQ = "general_faq"
    UNKNOWN = "unknown"

class Priority(Enum):
    LOW = "low"
    MEDIUM = "medium"
    HIGH = "high"
    URGENT = "urgent"

@dataclass
class ConversationState:
    session_id: str
    customer_id: Optional[str] = None
    intent: Intent = Intent.UNKNOWN
    priority: Priority = Priority.MEDIUM
    turn_count: int = 0
    resolved: bool = False
    escalated: bool = False
    context: dict = field(default_factory=dict)
    history: list = field(default_factory=list)

    def add_turn(self, role: str, content: str):
        self.history.append({"role": role, "content": content})
        if role == "user":
            self.turn_count += 1

Intent Classification Layer

The first thing the agent does with every user message is classify intent. This determines which tools and knowledge bases to activate. A two-stage approach works well: a fast keyword matcher for obvious cases, and an LLM classifier for ambiguous inputs.

import re
from openai import AsyncOpenAI

KEYWORD_PATTERNS = {
    Intent.ORDER_STATUS: r"(where is my order|track|shipping|delivery|package)",
    Intent.BILLING: r"(charge|invoice|payment|refund amount|bill)",
    Intent.RETURNS: r"(return|exchange|send back|refund|warranty)",
    Intent.TECHNICAL: r"(not working|error|bug|crash|broken|help with)",
}

def classify_by_keywords(message: str) -> Optional[Intent]:
    lower = message.lower()
    for intent, pattern in KEYWORD_PATTERNS.items():
        if re.search(pattern, lower):
            return intent
    return None

async def classify_by_llm(
    client: AsyncOpenAI, message: str, history: list
) -> Intent:
    response = await client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {
                "role": "system",
                "content": (
                    "Classify the customer intent. "
                    "Return exactly one of: order_status, billing, "
                    "technical, returns, general_faq, unknown"
                ),
            },
            *history[-4:],
            {"role": "user", "content": message},
        ],
        max_tokens=20,
    )
    label = response.choices[0].message.content.strip().lower()
    try:
        return Intent(label)
    except ValueError:
        return Intent.UNKNOWN

async def classify_intent(
    client: AsyncOpenAI, message: str, history: list
) -> Intent:
    keyword_result = classify_by_keywords(message)
    if keyword_result:
        return keyword_result
    return await classify_by_llm(client, message, history)

Conversation Flow and Escalation

The main agent loop ties everything together. After classifying intent, the agent retrieves context, attempts resolution, and decides whether to escalate based on confidence, sentiment, and turn count.

ESCALATION_THRESHOLDS = {
    "max_turns": 5,
    "low_confidence": 0.4,
    "negative_sentiment": -0.6,
}

async def should_escalate(state: ConversationState, confidence: float, sentiment: float) -> bool:
    if state.priority == Priority.URGENT:
        return True
    if state.turn_count >= ESCALATION_THRESHOLDS["max_turns"]:
        return True
    if confidence < ESCALATION_THRESHOLDS["low_confidence"]:
        return True
    if sentiment < ESCALATION_THRESHOLDS["negative_sentiment"]:
        return True
    return False

async def run_support_agent(state: ConversationState, user_message: str):
    state.add_turn("user", user_message)
    intent = await classify_intent(client, user_message, state.history)
    state.intent = intent

    # Retrieve relevant knowledge and generate response
    knowledge = await retrieve_knowledge(intent, user_message)
    response, confidence = await generate_response(
        state, knowledge, user_message
    )
    sentiment = await analyze_sentiment(user_message)

    if await should_escalate(state, confidence, sentiment):
        state.escalated = True
        return await transfer_to_human(state)

    state.add_turn("assistant", response)
    return response

This architecture keeps each concern isolated. You can swap out the intent classifier, upgrade the knowledge base, or adjust escalation rules without rewriting the conversation loop.

FAQ

How many intents should a support agent handle?

Start with five to eight broad intents that cover 80% of your ticket volume. You can add sub-intents later as you analyze misclassifications. Trying to cover every edge case from the start leads to fragile classifiers and overlapping categories.

Should I use a fine-tuned model or prompt-based classification for intents?

For most teams, prompt-based classification with a fast model like GPT-4o-mini is the right starting point. It requires no training data and can be updated instantly. Fine-tuning becomes worthwhile once you have 1,000+ labeled examples per intent and need sub-50ms classification latency.

How do I handle multi-intent messages like "Where is my order and I want a refund"?

Detect multi-intent messages by running classification twice — once on each clause after splitting on conjunctions. Process the most urgent intent first, then address the second. Store both intents in conversation state so the agent can circle back naturally.

#CustomerSupport #AIAgents #ConversationDesign #IntentClassification #Escalation #AgenticAI #LearnAI #AIEngineering