Use Case: AI Refund Agent

End-to-end walkthrough of an AI customer support agent handling refunds with SpendGuard.

The Scenario

You have an AI customer support agent that handles refund requests. Customers chat with the agent, and when a refund is appropriate, the agent issues it automatically.

The problem: Without guardrails, the agent could refund $5,000 on a$20 order, issue refunds on year-old purchases, or process the same refund twice if the customer asks again.

The solution: SpendGuard sits between the agent and the refund action. Before the agent issues any refund, it checks with SpendGuard first.

---

Step 1: Create Your Refund Policy

Start with the AI Support Refund Policy template and customize it:

curl -X POST https://spendguardapi.com/v1/policies \
  -H "Content-Type: application/json" \
  -H "X-API-Key: $SPENDGUARD_API_KEY" \
  -d '{
    "policy_id": "support_refund_policy",
    "name": "AI Support Refund Policy",
    "description": "Controls refunds issued by our AI support agent",
    "rules": [
      {
        "rule_id": "r1",
        "rule_type": "max_amount",
        "description": "Block refunds over $500",
        "parameters": { "limit": 500, "currency": "USD" }
      },
      {
        "rule_id": "r2",
        "rule_type": "refund_age_limit",
        "description": "No refunds after 30 days",
        "parameters": { "max_days": 30 }
      },
      {
        "rule_id": "r3",
        "rule_type": "escalate_if",
        "description": "Escalate refunds over $200",
        "parameters": { "amount_above": 200, "action_types": ["refund"] }
      },
      {
        "rule_id": "r4",
        "rule_type": "duplicate_guard",
        "description": "Block duplicate refunds within 10 minutes",
        "parameters": { "window_minutes": 10 }
      }
    ]
  }'

---

Step 2: Wire the Agent to Check Before Refunding

In your agent’s refund flow, add a SpendGuard check before executing the refund:

import requests

SPENDGUARD_URL = "https://spendguardapi.com"
SPENDGUARD_KEY = "sg_live_your_key_here"

def process_refund(agent_id, customer_id, amount, days_since_purchase):
    """Check with SpendGuard, then process if allowed."""

    # Step 1: Ask SpendGuard
    check = requests.post(
        f"{SPENDGUARD_URL}/v1/checks",
        headers={
            "X-API-Key": SPENDGUARD_KEY,
            "Content-Type": "application/json",
        },
        json={
            "agent_id": agent_id,
            "policy_id": "support_refund_policy",
            "action_type": "refund",
            "amount": amount,
            "currency": "USD",
            "counterparty": customer_id,
            "metadata": {
                "days_since_purchase": days_since_purchase,
            },
        },
    ).json()

    decision = check["decision"]

    # Step 2: Act on the decision
    if decision == "allow":
        execute_refund(customer_id, amount)
        return f"Refund of ${amount} processed for {customer_id}."

    elif decision == "escalate":
        create_escalation_ticket(customer_id, amount, check["message"])
        return f"This refund needs manager approval. I've created a ticket."

    elif decision == "block":
        return f"I can't process this refund: {check['message']}"

---

Step 3: See It In Action

Scenario A: $50 refund, 5 days old → Allow

The agent checks a small, recent refund:

{
  "decision": "allow",
  "confidence": "high",
  "message": "Action is within policy. Proceed."
}

The agent processes the refund automatically.

Scenario B: $300 refund, 10 days old → Escalate

The amount is above the $200 escalation threshold:

{
  "decision": "escalate",
  "confidence": "high",
  "reason_code": "escalation_threshold_exceeded",
  "message": "Refund of $300.00 exceeds the escalation threshold of $200.00.",
  "violated_rule_id": "r3"
}

The agent tells the customer: “This refund needs manager approval. I’ve created a ticket for you.”

Scenario C: $750 refund → Block

The amount exceeds the $500 max:

{
  "decision": "block",
  "confidence": "high",
  "reason_code": "max_amount_exceeded",
  "message": "Amount $750.00 exceeds the policy limit of $500.00.",
  "violated_rule_id": "r1"
}

The agent tells the customer: “I’m unable to process a refund of this size. Please contact our support team directly.”

Scenario D: Same refund submitted twice → Block (Duplicate)

The customer asks for the same refund again within 10 minutes:

{
  "decision": "block",
  "confidence": "high",
  "reason_code": "duplicate_action_detected",
  "message": "This action was already submitted recently."
}

The agent recognizes this is a duplicate and doesn’t process it again.

---

Step 4: Monitor Your Agent

Check the violations log to see what your agent has been blocked or escalated on:

curl "https://spendguardapi.com/v1/violations?agent_id=support-agent-v1&limit=10" \
  -H "X-API-Key: $SPENDGUARD_API_KEY"

This gives you a complete audit trail of every decision that wasn’t a simple “allow.”