Skip to main content

Customer Support Application with Control Zero

This guide shows how to build a customer support chatbot with governance policies that control what the bot can do, which tools it can use, and what data it can access.

What You Will Build​

A customer support agent that:

  • Answers customer questions using an LLM
  • Can look up order status and account information
  • Has policy guardrails preventing unauthorized actions
  • Cannot modify orders, issue refunds, or access sensitive data without policy approval
  • Logs every action for compliance

Architecture​

Customer Message
    |
    v
[Control Zero: enforce session policy]
    |
    v
LLM Processing (understand intent, select tool)
    |
    v
[Control Zero: enforce tool call policy]
    |
    v
Tool Execution (order lookup, account info, etc.)
    |
    v
[Control Zero: enforce response policy]
    |
    v
Response to Customer

Implementation​

Setup​

pip install controlzero openai
import controlzero
import openai

cz = controlzero.ControlZero()
cz.initialize()

openai_client = openai.OpenAI()

Define Support Tools​

# Simulated support tools
def lookup_order(order_id: str) -> dict:
    """Look up order status."""
    return {"order_id": order_id, "status": "shipped", "eta": "2026-03-05"}

def lookup_account(customer_id: str) -> dict:
    """Look up customer account info."""
    return {"customer_id": customer_id, "name": "Jane Doe", "tier": "premium"}

def issue_refund(order_id: str, amount: float) -> dict:
    """Issue a refund for an order."""
    return {"order_id": order_id, "refund_amount": amount, "status": "processed"}

def escalate_to_human(reason: str) -> dict:
    """Escalate the conversation to a human agent."""
    return {"escalated": True, "reason": reason}

TOOLS = {
    "lookup_order": lookup_order,
    "lookup_account": lookup_account,
    "issue_refund": issue_refund,
    "escalate_to_human": escalate_to_human,
}

Enforce Tool Calls​

def execute_tool(tool_name: str, arguments: dict, agent_id: str) -> dict:
    """Execute a support tool with policy enforcement."""

    # Enforce: is this agent allowed to call this tool?
    cz.enforce(
        action="tool.call",
        resource=f"tool/{tool_name}",
        context={
            "agent_id": agent_id,
            "arguments": str(arguments),
        },
    )

    # Tool call allowed -- execute it
    tool_fn = TOOLS.get(tool_name)
    if not tool_fn:
        raise ValueError(f"Unknown tool: {tool_name}")

    return tool_fn(**arguments)

Support Agent Loop​

def handle_customer_message(
    message: str,
    conversation_history: list[dict],
    agent_id: str = "support-bot",
) -> str:
    """Handle a customer message with full policy enforcement."""

    # Enforce: is this agent allowed to generate responses?
    cz.enforce(
        action="llm.generate",
        resource="model/gpt-4",
        context={"agent_id": agent_id},
    )

    # Build messages with conversation history
    messages = [
        {
            "role": "system",
            "content": (
                "You are a customer support assistant. "
                "You can look up orders, check account info, and escalate to humans. "
                "Be helpful and professional. Never share sensitive financial data."
            ),
        },
        *conversation_history,
        {"role": "user", "content": message},
    ]

    # Define available tools for function calling
    tools = [
        {
            "type": "function",
            "function": {
                "name": "lookup_order",
                "description": "Look up order status by order ID",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "order_id": {"type": "string"},
                    },
                    "required": ["order_id"],
                },
            },
        },
        {
            "type": "function",
            "function": {
                "name": "lookup_account",
                "description": "Look up customer account information",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "customer_id": {"type": "string"},
                    },
                    "required": ["customer_id"],
                },
            },
        },
        {
            "type": "function",
            "function": {
                "name": "issue_refund",
                "description": "Issue a refund for an order",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "order_id": {"type": "string"},
                        "amount": {"type": "number"},
                    },
                    "required": ["order_id", "amount"],
                },
            },
        },
        {
            "type": "function",
            "function": {
                "name": "escalate_to_human",
                "description": "Escalate the conversation to a human agent",
                "parameters": {
                    "type": "object",
                    "properties": {
                        "reason": {"type": "string"},
                    },
                    "required": ["reason"],
                },
            },
        },
    ]

    # Call the LLM
    response = openai_client.chat.completions.create(
        model="gpt-4",
        messages=messages,
        tools=tools,
    )

    choice = response.choices[0]

    # Handle tool calls with policy enforcement
    if choice.message.tool_calls:
        for tool_call in choice.message.tool_calls:
            import json
            tool_name = tool_call.function.name
            arguments = json.loads(tool_call.function.arguments)

            try:
                result = execute_tool(tool_name, arguments, agent_id)
                return f"Tool result ({tool_name}): {result}"
            except controlzero.PolicyViolationError as e:
                return (
                    "I am not authorized to perform that action. "
                    "Let me connect you with a human agent who can help."
                )

    return choice.message.content

Example Policy​

Allow read-only operations but block refunds (require human approval):

{
  "name": "support-bot-policy",
  "rules": [
    {
      "effect": "allow",
      "action": "llm.generate",
      "resource": "model/gpt-4"
    },
    {
      "effect": "allow",
      "action": "tool.call",
      "resource": "tool/lookup_order"
    },
    {
      "effect": "allow",
      "action": "tool.call",
      "resource": "tool/lookup_account"
    },
    {
      "effect": "deny",
      "action": "tool.call",
      "resource": "tool/issue_refund"
    },
    {
      "effect": "allow",
      "action": "tool.call",
      "resource": "tool/escalate_to_human"
    }
  ]
}

With this policy, the support bot can look up orders and accounts, and escalate to humans, but cannot issue refunds. When a customer asks for a refund, the bot is automatically blocked and should escalate.

Why This Matters​

Customer support bots handle sensitive operations and interact directly with customers. Without governance:

  • A bot could issue unauthorized refunds, costing the business money
  • A bot could access account data it should not have
  • A bot could use an expensive model when a cheaper one would suffice
  • There is no audit trail of what the bot did and why

Control Zero provides the enforcement layer that ensures the bot operates within its approved boundaries.

Next Steps​