Claims Processing Pipelines

New domain. No handbook prerequisite — this chapter is self-contained.

Your auto insurance platform processes 4,000 claims per day. A policyholder reports a fender-bender at 8:47 AM, uploads three photos, and expects a response. Forty-eight hours later they call back — nobody has looked at it. The claim sat in a queue behind 1,200 others because your intake pipeline classified it as “standard” and routed it to a team already at capacity. The regulatory clock started the moment FNOL arrived. In California, you had 15 days to acknowledge. In Florida, 14. You are now late on both, and the state insurance commissioner tracks late-acknowledgment rates by carrier.

Claims processing is not a customer service problem. It is a pipeline engineering problem with regulatory SLAs, fraud exposure, and financial reserve implications at every stage.

The Claim Lifecycle State Machine

A claim moves through a fixed set of states. Transitions are explicit, auditable, and — in regulated insurance — legally meaningful. The state machine is the spine of the pipeline.

FNOL → ACKNOWLEDGED → INVESTIGATING → ADJUDICATING → APPROVED/DENIED → PAYMENT → CLOSED
                                          ↓
                                      SUBROGATION

Each state has entry conditions, exit conditions, and a regulatory clock.

State	Entry condition	SLA	Who acts
FNOL	Loss event reported	—	Policyholder
ACKNOWLEDGED	System confirms receipt	15 days (CA), 14 days (FL), 10 days (NY)	Automated
INVESTIGATING	Adjuster assigned	30 days to complete (most states)	Adjuster
ADJUDICATING	Investigation complete, coverage confirmed	15 days to decide after proof of loss	Adjuster + rules engine
APPROVED	Liability confirmed, amount determined	30 days to pay after agreement	System
DENIED	Coverage exclusion or fraud confirmed	Must issue written explanation	Adjuster
PAYMENT	Check issued or EFT initiated	5 days after approval (some states)	Payment system
CLOSED	Payment cleared and accepted	—	System
SUBROGATION	Third-party liability identified	Statute of limitations (2-6 years)	Recovery team

The state machine enforces two invariants. First, no claim skips a state — you cannot go from FNOL to PAYMENT without passing through investigation and adjudication. Second, every transition logs the actor, timestamp, and reason. An auditor or regulator will reconstruct the full timeline of a disputed claim.

VALID_TRANSITIONS = {
    "FNOL":          ["ACKNOWLEDGED"],
    "ACKNOWLEDGED":  ["INVESTIGATING"],
    "INVESTIGATING": ["ADJUDICATING", "DENIED"],
    "ADJUDICATING":  ["APPROVED", "DENIED"],
    "APPROVED":      ["PAYMENT", "SUBROGATION"],
    "DENIED":        ["INVESTIGATING"],  # reopened on appeal
    "PAYMENT":       ["CLOSED", "SUBROGATION"],
    "CLOSED":        ["INVESTIGATING"],  # reopened
    "SUBROGATION":   ["CLOSED"],
}

State machines for claims differ from payment state machines in one critical way: claims can reopen. A closed auto claim reopens when a claimant discovers additional damage six months later. Your state machine must handle CLOSED → INVESTIGATING without corrupting the original claim record. Reopening creates a supplemental claim linked to the original, not a mutation of the original’s history.

First Notice of Loss

FNOL is the intake stage. The policyholder reports a loss through an app, phone call, web form, or agent. Your system must capture enough information to route the claim without requiring the full story upfront.

The minimum FNOL payload:

Field	Purpose	Routing impact
Policy number	Links to coverage, limits, deductible	Determines coverage type
Date of loss	Starts the regulatory clock	Flags late-reported claims
Loss type	Auto, property, liability, workers’ comp	Routes to specialized queue
Loss description	Free text or structured input	Feeds triage classifier
Contact information	Where to reach the claimant	Adjuster assignment

ACORD forms standardize the data model. ACORD 1 (property), ACORD 2 (auto), and ACORD 3 (workers’ comp) define 200+ fields each. Most digital FNOL implementations capture 15-20 fields at intake and backfill the rest during investigation.

Lemonade’s AI Jim processes FNOL through a conversational interface that extracts structured data from natural language. Their published metric: 3 seconds from submission to first payment on simple claims. That speed is achievable only because the triage classifier immediately separates simple claims (below deductible, clear liability, no injury) from complex ones that require human investigation.

The triage classifier routes claims into three lanes:

Lane	Criteria	Handling	Volume share
Straight-through	Below deductible, single vehicle, no injury, low fraud score	Automated adjudication	15-25%
Standard	Moderate complexity, clear coverage	Human adjuster, normal queue	60-70%
Complex	Multi-party, disputed liability, injury, high fraud score, large reserve	Senior adjuster, priority queue	10-20%

The routing decision must happen within minutes of FNOL, not hours. A complex claim misrouted to the standard queue sits for days before an adjuster realizes it needs escalation. A simple claim routed to the complex queue wastes adjuster time.

Document Processing

Claims generate documents — damage photos, police reports, medical records, repair estimates, witness statements. A single auto claim with injuries averages 12-15 documents across 4-5 document types. A property claim from a natural disaster can generate 40+.

The document pipeline has three stages: ingestion, classification, and extraction.

Ingestion accepts documents from multiple channels — app uploads, email attachments, fax (still common in medical records), and adjuster field uploads. Each document gets a unique ID, a SHA-256 hash for deduplication, and an association to the claim.

Classification determines the document type. A photo might be vehicle damage, property damage, or a receipt. An uploaded PDF might be a police report, a medical bill, or a repair estimate. Classification accuracy directly affects downstream extraction — a medical bill parser applied to a police report produces garbage.

Document type	Source	Extraction targets
Damage photos	Policyholder, adjuster	Damage severity score, affected components
Police report	Law enforcement	Parties involved, citations, fault determination
Medical records	Healthcare provider	ICD-10 codes, treatment timeline, prognosis
Repair estimate	Body shop, contractor	Line items, labor hours, parts cost, total
Invoice/receipt	Service provider	Amount, date, service description

Extraction pulls structured data from classified documents. OCR handles scanned PDFs. Computer vision models score damage severity from photos — Tractable’s AI estimates auto damage from photos with accuracy comparable to human adjusters, reducing the need for in-person inspections on claims under $10,000. Medical record extraction maps free-text clinical notes to ICD-10 codes and treatment timelines.

The engineering constraint: medical records are PHI (Protected Health Information) under HIPAA. Your document pipeline must encrypt at rest, restrict access by role, and audit every access. A claims adjuster can see the medical records for their assigned claim. A data engineer debugging the extraction pipeline cannot.

Fraud Detection

Insurance fraud costs the US industry $80 billion per year. Between 5% and 10% of all claims contain some element of fraud — ranging from exaggerated damage to staged accidents.

Fraud detection operates at two points in the pipeline: at FNOL (pre-investigation) and during adjudication (post-investigation).

FNOL-stage fraud scoring uses features available at intake:

fraud_features = {
    "policy_age_days": 47,          # new policy — higher risk
    "prior_claims_12mo": 3,         # high frequency
    "loss_reported_delay_days": 14, # late reporting
    "time_of_loss": "02:30",        # unusual hour
    "loss_type_match": False,       # loss type doesn't match coverage emphasis
    "geo_risk_score": 0.78,         # high-fraud-rate ZIP code
}

A fraud score above the threshold (typically 0.7-0.8) routes the claim to the Special Investigations Unit (SIU). Below the threshold, the claim proceeds through normal channels with a fraud flag attached for the adjuster’s awareness.

Adjudication-stage fraud detection uses the full investigation record — repair estimates compared against photo damage, medical treatment patterns (unusual frequency, out-of-network providers, excessive duration), and network analysis (shared addresses, phone numbers, or providers across multiple claims).

Signal	Detection method	False positive rate
Inflated repair estimate	Compare estimate to photo-based damage model	8-12%
Staged accident	Network analysis — linked claimants across claims	3-5%
Exaggerated injury	Treatment pattern analysis against ICD-10 baselines	15-20%
Premium fraud	Application data vs external data sources	5-8%

The false positive rate matters. A 15% false positive rate on exaggerated injury means 1 in 7 legitimate injury claimants gets flagged for SIU investigation. That investigation delays their payment, degrades the customer experience, and — in bad faith litigation — creates liability for the carrier. Tuning fraud models is a precision-recall tradeoff where false negatives cost money and false positives cost reputation and legal exposure.

Fraud detection in claims shares the same champion/challenger deployment pattern as fraud detection in payments. See Fraud Model Deployment for the model lifecycle — shadow scoring, gradual rollout, and rollback triggers apply identically.

Adjudication and Reserve Management

Adjudication determines two things: does the policy cover this loss, and if so, how much does the carrier owe?

Coverage determination walks the policy’s insuring agreements, exclusions, conditions, and endorsements. This is a rule engine problem. The policy language is structured — “we will pay for direct physical loss to covered property caused by a covered peril, subject to the exclusions listed in Section II” — but the interpretation is not always mechanical. A rule engine handles the 80% of claims where coverage is clear. The remaining 20% require human judgment.

coverage_result = evaluate_coverage(
    policy=policy,
    loss_type="collision",
    loss_date=datetime(2026, 9, 15),
    deductible=500,
    limit=50000,
)
# Returns: COVERED, EXCLUDED, or REVIEW_REQUIRED

Loss reserving happens at FNOL and is revised at each pipeline stage. The initial reserve is an estimate of the ultimate payout — set too low, the carrier underreports liabilities; set too high, it ties up capital unnecessarily.

Reserve stage	Basis	Accuracy
FNOL (initial)	Loss type average × severity estimate	±40-60%
Post-investigation	Repair estimate + medical projection	±15-25%
Post-adjudication	Agreed settlement or judgment	±5-10%

Reserve accuracy has financial reporting implications. IBNR (Incurred But Not Reported) reserves appear on the carrier’s balance sheet. An underestimate triggers reserve strengthening, which hits the income statement. An overestimate inflates reserves, reducing reported profitability and misleading investors. Actuarial models (loss development triangles, Bornhuetter-Ferguson) project ultimate losses from current data, but the claims pipeline feeds those models — garbage reserve estimates in, garbage projections out.

Subrogation

Subrogation recovers money from a third party who caused the loss. Your insured rear-ends another driver — your carrier pays the other driver’s claim, then subrogate against your insured (or their carrier) to recover the payout.

The subrogation pipeline runs parallel to the main claim after payment.

Step	Action	Timeline
Identification	Determine third-party liability exists	During adjudication
Demand	Send demand letter with supporting documentation	30-60 days post-payment
Negotiation	Carrier-to-carrier negotiation or arbitration	60-180 days
Recovery	Payment received, offset against claim cost	6-18 months total
Deductible refund	Return policyholder’s deductible from recovery	After recovery

Intercompany arbitration follows ACORD’s Arbitration Forums rules. Most auto subrogation disputes between US carriers resolve through this process rather than litigation.

Recovery rates vary by line of business. Auto collision subrogation recovers 40-60% of paid losses. Property subrogation from negligent third parties recovers 20-30%. The recovery rate is a key financial metric — a 5% improvement in subrogation recovery on a $500M paid loss book returns $25M.

SLA Management

Every state regulates claim handling timelines. Missing them triggers regulatory action — fines, market conduct examinations, and in severe cases, restrictions on writing new business.

Your pipeline must track regulatory deadlines per claim, per state.

SELECT
    c.claim_id,
    c.state,
    c.current_state,
    r.sla_days,
    c.state_entered_at + (r.sla_days || ' days')::interval AS deadline,
    NOW() > c.state_entered_at + (r.sla_days || ' days')::interval AS overdue
FROM claims c
JOIN regulatory_slas r
    ON r.state = c.state
    AND r.claim_state = c.current_state
WHERE c.current_state NOT IN ('CLOSED', 'SUBROGATION');

The SLA dashboard is not optional. State insurance departments conduct market conduct examinations that sample 50-200 claims and measure compliance with handling timelines. A carrier with a 15% late-acknowledgment rate faces fines and increased examination frequency. A carrier at 2% gets lighter scrutiny.

Regulatory requirement	Typical SLA	Consequence of violation
Acknowledge FNOL	10-15 days	Fine per violation ($100-$1,000)
Accept or deny	15-30 days after proof of loss	Fine + bad faith exposure
Pay approved claim	5-30 days after agreement	Interest penalties + bad faith
Respond to inquiries	15 days	Fine per violation

The pipeline must also generate regulatory reports — claim counts by state, average handling time, late-response rates, complaint ratios. These reports are not analytics. They are compliance deliverables with legal weight, and the data lineage from individual claim events to aggregated report numbers must be auditable.