E-commerce — Returns Fraud Detection & Processing
Automate return requests — detect fraud patterns, verify purchase history, refund low-risk returns, and send everything else to loss prevention, who approve or decline it before any money moves
Open in the playground and play with it →
halo
// ============================================================
// E-commerce — Returns Fraud Detection & Processing
// PURPOSE: Automate return requests — detect fraud patterns,
// verify purchase history, refund low-risk returns,
// and send everything else to loss prevention, who
// approve or decline it before any money moves
// ============================================================
// --- Policies ---
policy PII : Redact(CreditCard) & Redact(SSN)
// Refund abuse and wardrobing are judged by FraudDetector's score, which routes the return
// below; the policy makes sure that score is a real number from 0 to 100.
policy FraudPolicy : Enforce(JSON) & Validate("fraud_score", ">=", "0") & Validate("fraud_score", "<=", "100")
policy CostGuard : MaxSpend(0.50)
// --- Tools ---
tool OrdersAPI : REST("api.shopify.com/admin/orders", apiKey: Env.SHOPIFY_KEY)
tool PaymentGW : POST("api.stripe.com/v1/refunds", apiKey: Env.STRIPE_KEY)
tool CXSlack : POST("hooks.slack.com/services/cx-team")
// --- Store ---
tool ReturnHistory : VectorStore(Env.PINECONE_KEY, namespace: "return-patterns")
tool FraudSignals : VectorStore(Env.PINECONE_KEY, namespace: "fraud-indicators")
// --- Struct ---
schema ReturnDecision {
approved : Boolean
refund_amount : Number
fraud_score : Number
action : String
}
// --- Agents ---
agent ReturnBot [openai/gpt-4o-mini, temp: 0.0]
equips OrdersAPI, PaymentGW
requires PII, CostGuard
agent FraudDetector [anthropic/claude-sonnet-4, temp: 0.0]
system """
Score how likely a return is abuse (0-100): serial returns, wardrobing
(worn and returned), item-not-received claims against delivered orders.
Answer JSON: {"fraud_score": <0-100>, "signals": [...]}.
"""
equips ReturnHistory, FraudSignals
requires PII, FraudPolicy, CostGuard
// --- The Pipeline ---
@timeout(2m)
@retry(2, 5s)
pipeline ProcessReturn(request: JSON) -> ReturnDecision:
request
// Step 1: Fan-out — verify order and check fraud signals in parallel
|* {
ReturnBot.verify_order(
tool: OrdersAPI,
order_id: context.data.request.order_id,
customer_id: context.data.request.customer_id
),
FraudDetector.score_risk(
customer_id: context.data.request.customer_id,
return_reason: context.data.request.reason,
item_value: context.data.request.item_price,
days_since_purchase: context.data.request.days_owned
)
}
// Step 2: Route on the fraud score: a low-risk return is refunded,
// anything else waits for loss prevention
|? {
(context.data.score_risk.fraud_score < 30) => Refund(context.data.request),
else => LossPrevention(context.data.request, context.data.score_risk, context.data.verify_order),
} as outcome
// Step 3: Record the decision in the declared shape
|> ReturnBot.decide(
fraud: context.data.score_risk,
outcome: context.data.outcome,
request: context.data.request
) as decision
// Step 4: Log the return for future pattern detection
|> ReturnHistory.log(
customer: context.data.request.customer_id,
amount: context.data.request.item_price,
reason: context.data.request.reason,
fraud_score: context.data.score_risk.fraud_score
)
|> System.Value(context.data.decision)
// Issue the refund to the original payment method.
pipeline Refund(request: JSON) -> Result:
request
|> ReturnBot.issue_refund(
tool: PaymentGW,
order_id: context.data.request.order_id,
amount: context.data.request.item_price,
method: "original_payment"
)
// A risky return: loss prevention approves or declines it before any money moves.
pipeline LossPrevention(request: JSON, risk: JSON, order: JSON) -> Result:
risk
|> CXSlack.post(
message: "FRAUD ALERT — manual review required",
order: context.data.order,
fraud: context.data.risk,
channel: "loss-prevention"
)
|! yield(CXSlack, context.data.risk, [context.data.order], 4h)
~> listen(CXSlack.Reaction) as review
|? {
(context.data.review.approved == true) => Refund(context.data.request),
else => System.Value("declined by loss prevention"),
}