Insurance — Claims Processing & Fraud Detection
Automate the claims lifecycle — from initial filing through fraud scoring, adjuster review, payout calc, and disbursement. Replaces BPM tools (Camunda, Pega) that cost $500K+/year to license and maintain.
Open in the playground and play with it →
halo
// ============================================================
// Insurance — Claims Processing & Fraud Detection
// PURPOSE: Automate the claims lifecycle — from initial filing
// through fraud scoring, adjuster review, payout calc,
// and disbursement. Replaces BPM tools (Camunda, Pega)
// that cost $500K+/year to license and maintain.
// ============================================================
// --- Policies ---
// Medical details have no detector: keep medical records out of the claim
// payload. Identifiers and contact details are masked before any model sees them.
policy ClaimsPrivacy : Redact(SSN, CreditCard, Email, Phone)
policy FraudPolicy : Block(DestructiveOps, PromptInjection)
// The payout is a JSON record, and never more than the policy covers.
policy PayoutLimit : Require(JSON) & Validate("payout_amount", "<=", "context.data.verify_coverage.max_payout")
policy CostGuard : MaxSpend(1.50)
// --- Tools ---
tool PolicyAPI : REST("api.guidewire.com/v1/policies", apiKey: Env.GUIDEWIRE_KEY)
tool ClaimsDB : REST("api.guidewire.com/v1/claims", apiKey: Env.GUIDEWIRE_KEY)
tool PaymentGW : POST("api.stripe.com/v1/payouts", apiKey: Env.STRIPE_KEY)
tool AdjusterSlack : POST("hooks.slack.com/services/claims-team")
// --- Store ---
tool FraudPatterns : VectorStore(Env.PINECONE_KEY, namespace: "fraud-indicators")
tool ClaimHistory : VectorStore(Env.PINECONE_KEY, namespace: "historical-claims")
// --- Struct ---
schema ClaimFiling {
claim_id : String
policy_number : String
incident_date : String
type : String
claimant_name : String
amount : Number
description : String
metadata : Object
payment_method : String
}
schema ClaimDecision {
claim_id : String
fraud_score : Number
coverage_valid : Boolean
payout_amount : Number
decision : String
adjuster_notes : String
}
// --- Agents ---
agent ClaimsBot [openai/gpt-4o-mini, temp: 0.0]
equips PolicyAPI, PaymentGW
requires ClaimsPrivacy, CostGuard
agent FraudAnalyst [anthropic/claude-sonnet-4, temp: 0.0]
system "Score how likely the claim is fraudulent. Answer with JSON: {\"fraud_score\": 0-100, \"reasons\": [...]}"
equips FraudPatterns, ClaimHistory
requires ClaimsPrivacy, FraudPolicy, CostGuard
@correct(3)
agent AdjusterAssistant [openai/gpt-4o, temp: 0.2]
system """
Calculate the payout from the coverage limits and the damage.
Answer with JSON: {"claim_id": "...", "fraud_score": 0, "coverage_valid": true,
"payout_amount": 0, "decision": "approve" | "deny", "adjuster_notes": "..."}
"""
equips ClaimHistory
requires ClaimsPrivacy, CostGuard, PayoutLimit
// --- The Pipeline ---
@timeout(5m)
@retry(2, 10s)
pipeline ProcessClaim(claim: ClaimFiling) -> ClaimDecision:
claim
// Step 1: Parallel verification — policy lookup + fraud scoring
|* {
ClaimsBot.verify_coverage(
tool: PolicyAPI,
policy_number: context.data.claim.policy_number,
incident_date: context.data.claim.incident_date,
claim_type: context.data.claim.type
),
FraudAnalyst.score_fraud(
claimant: context.data.claim.claimant_name,
amount_requested: context.data.claim.amount,
incident_description: context.data.claim.description,
submission_metadata: context.data.claim.metadata
)
}
// Step 2: A high fraud score goes to the investigators, and the claim stops here
|? {
(context.data.score_fraud.fraud_score > 80) => AdjusterSlack.post(
message: "HIGH FRAUD RISK — manual review required immediately",
claim_id: context.data.claim.claim_id,
fraud_score: context.data.score_fraud.fraud_score,
channel: "fraud-investigations"
),
}
|? (context.data.score_fraud.fraud_score <= 80)
// Step 3: Gate — stop unless the policy covers this claim type
|? (context.data.verify_coverage.covered == true)
// Step 4: Calculate payout based on coverage limits + damage assessment
|> AdjusterAssistant.calculate_payout(
coverage: context.data.verify_coverage,
claim: context.data.claim,
fraud_assessment: context.data.score_fraud,
comparable_claims: ClaimHistory
)
|? (context.data.calculate_payout.decision == "approve")
// Step 5: Small claims pay out automatically; an adjuster approves larger ones
|? {
(context.data.calculate_payout.payout_amount > 5000) => AdjusterReview(context.data.calculate_payout, context.data.verify_coverage, context.data.score_fraud),
else => System.Value(context.data.calculate_payout),
} as review
|? (context.data.calculate_payout.payout_amount <= 5000 || context.data.review.approved == true)
// Step 6: Disburse payment
|> ClaimsBot.disburse(
tool: PaymentGW,
claim_id: context.data.claim.claim_id,
amount: context.data.calculate_payout.payout_amount,
method: context.data.claim.payment_method
)
// Step 7: Log to claims system, and return the decision
|> ClaimsDB.post(
event: "CLAIM_RESOLVED",
claim_id: context.data.claim.claim_id,
payout: context.data.calculate_payout.payout_amount,
fraud_score: context.data.score_fraud.fraud_score,
decision: "approved"
)
|> System.Value(context.data.calculate_payout)
// An adjuster approves a payout over $5,000.
pipeline AdjusterReview(payout: JSON, coverage: JSON, fraud: JSON) -> Result:
payout
|! yield(AdjusterSlack, payout, [coverage, fraud], 24h)
~> listen(AdjusterSlack.Reaction) as adjuster
|> System.Value(context.data.adjuster)