AI in Underwriting, Claims, and Fraud

Lesson 1 established the business: predict expected loss, price it, pay claims, and live on a thin combined ratio. This lesson covers the machinery insurers actually build, concentrated where the money is.

Four areas carry nearly all the value:

• Underwriting and pricing: getting the expected loss right, which drives the loss ratio.
• New data for pricing: telematics and behavioral signals that measure risk instead of proxying it.
• Claims: intake, triage, damage assessment, and settlement, the largest cash outflow and the biggest expense pool.
• Fraud and special investigation: finding the claims that should not be paid as presented.

A framing to carry throughout: in insurance, AI is not replacing a manual, intuition-driven process with a mathematical one. Actuaries have modeled statistically for over a century. What AI changes is narrower and more specific: which model classes are usable, and which data can be fed to them. Decades of rating practice ran on generalized linear models over structured fields. The shift is toward more flexible models and toward data that was previously unreadable, photographs, free text, sensor streams.

That distinction matters, because it sets realistic expectations for what any of this buys.

Underwriting answers two questions: do we accept this risk, and at what price? Traditionally a human underwriter reviewed an application against guidelines, with a rating model supplying the price. AI reshapes both halves.

Automated underwriting. For high-volume, lower-complexity lines (personal auto, homeowners, simple term life), models can decide most cases outright, accepting a risk and issuing a price in seconds, and routing only ambiguous or high-value cases to a human. The economic logic is straightforward: underwriter time is expensive, and most applications are unremarkable. Concentrating scarce human judgment on the cases that need it is where the gain comes from.

Better rating models. The industry standard for decades has been the generalized linear model (GLM), transparent, well understood by regulators, and genuinely effective. Newer practice adds gradient-boosted trees and similar methods, which capture interactions and nonlinearities a GLM must be told about in advance. The gain is real but incremental, a few points of predictive lift, and Lesson 1 explains why that matters anyway: on a combined ratio hovering near 100, a few points of loss ratio is the whole margin.

Reading the unreadable. Underwriting evidence is often unstructured: medical records, property inspection reports, financial statements, satellite or aerial imagery of a roof. Models that extract structured facts from these documents remove a large manual bottleneck.

The most conceptually interesting pricing development is telematics, or usage-based insurance (UBI), most established in personal auto. Instead of inferring how you drive from proxies, a device or phone app measures it: mileage, time of day, hard braking, rapid acceleration, cornering, phone handling.

Why this is a genuine shift rather than more data: traditional auto rating leans on proxies for driving behavior, age, vehicle type, garaging location, credit-based scores in some jurisdictions. These correlate with risk but are not risk. A careful young driver pays for the average of young drivers. Telematics substitutes direct measurement of the insured behavior for a correlate of it.

The implications run in several directions at once:

• Fairer in a specific sense. Pricing on what you actually do is more defensible than pricing on what people like you tend to do, and it lets a low-risk member of a high-risk-looking group be priced correctly.
• It changes behavior. Unlike a static rating factor, a driving score is feedback, and it attacks moral hazard from Lesson 1: a premium responsive to your driving restores an incentive that insurance otherwise dulls.
• It raises privacy questions that a rating table never did. Continuous location and behavior monitoring is an intrusion, and consent is doing real work.

Telematics is the clearest case of the industry's direction: from classifying people toward measuring behavior.

Claims is where the money leaves and where the customer finally experiences what they bought. It is also enormously labor-intensive, which makes it the biggest operational target for AI. The pipeline runs in stages, and models attach to each.

• First notice of loss (FNOL). The claim is reported, by phone, app, or web. Language models handle intake conversationally, extract structured details from an unstructured account, and start the file.
• Triage and routing. Predict the claim's likely complexity and severity, then route accordingly: simple claims to fast automated handling, complex or large ones to experienced adjusters. Getting this right is high leverage, since misrouting is expensive in both directions.
• Damage assessment. Computer vision estimates repair cost from photographs of a damaged vehicle or property. This is a genuinely new capability: an adjuster's physical inspection replaced, for suitable claims, by a model reading images the claimant uploads.
• Straight-through processing (STP). For small, clear-cut claims, the whole path from report to payment runs without human touch.
• Reserving support. Predicting a claim's ultimate cost early helps set reserves accurately, which matters because reserves are a balance-sheet estimate of money not yet paid.

The attraction is that claims automation improves both sides of the ledger at once: it cuts expense (fewer touches per claim) and improves customer experience (a claim paid today instead of in three weeks), which is unusual, because those normally trade off.

Fraud is not a marginal nuisance in insurance; it is a structural cost. The Coalition Against Insurance Fraud estimated in its 2022 study that insurance fraud costs the United States 308.6 billion dollars annually, its first update to the figure in 27 years. The study broke it down across lines: life insurance about 74.7 billion, Medicare and Medicaid about 68.7 billion, and property and casualty about 45 billion, and it put the burden at roughly 933 dollars per person per year. Fraud losses are paid by honest policyholders through higher premiums.

Insurance fraud detection differs from payment-fraud detection in an important way. The task is not to block a transaction in milliseconds; it is to decide whether a claim narrative, and the evidence around it, hangs together. Signals include:

• Claim-level anomalies: the loss reported days after the policy incepted, inconsistencies between the story and the damage, unusual timing.
• Network and link analysis. This is distinctive and powerful. Organized fraud is relational: the same clinic, body shop, attorney, or set of "unrelated" claimants recurring across supposedly independent claims. Graph analysis finds rings that no single claim looks suspicious enough to reveal.

The governing discipline: the model refers, the human decides. Output goes to the special investigation unit (SIU) as a prioritized lead. A fraud score is a reason to investigate, never a finding of fraud. Accusing a policyholder is a serious act with legal consequence, and the base rate is low, so most flagged claims are legitimate. Treating a score as a verdict would deny valid claims at scale.

Trace one realistic auto claim through the machinery to see how the pieces compose.

1. FNOL. A policyholder reports a rear-end collision through the app at 8pm. A language model runs the intake conversation, extracts the structured facts (date, location, vehicles, injuries reported: none), and opens the file. No human involved yet.
2. Triage. A severity model reads the file: no injuries, single vehicle damaged, modest estimated cost. It classifies the claim as low-complexity and routes it to the fast track rather than to a senior adjuster.
3. Damage assessment. The claimant uploads eight photos. A computer-vision model identifies the damaged panels and estimates repair cost at roughly 3,200 dollars, consistent with the described impact.
4. Fraud screen. A model scores the claim. Nothing fires: the policy is two years old, the story matches the damage pattern, and no network links to known rings. Score is low, so no SIU referral.
5. Decision. All checks agree and the amount sits under the straight-through threshold, so the claim is approved and paid the same evening.
6. The other path. Change one fact, the policy incepted nine days ago, damage inconsistent with the described impact, and the body shop appears in a cluster of prior claims. Now the fraud model scores high, straight-through processing is blocked, and the claim goes to the SIU as a prioritized lead for a human investigator. It is not denied by the model. It is queued for a person.

That branch is the whole design: automate the unambiguous, escalate the ambiguous.

Assemble the machinery, with an honest reading of what each delivers and where it can go wrong.

Two patterns run down that table. First, the wins are concentrated in volume: automating the routine so people handle the exceptional. Almost nothing here is a model doing what an expert does; it is a model doing what an expert should not have to.

Second, every failure mode is a decision made without recourse. The declined applicant who never learns why, the claim denied on a score, the payout automated past a mistake. That is not a coincidence. It is why insurance regulation focuses so precisely on explanation, testing, and human oversight, which is Lesson 3.

A reported claim is intake-processed and triaged by severity, assessed from photos, and fraud-screened; clean low-value claims settle straight through, while high scores or complexity route to human adjusters or the special investigation unit rather than being decided by the model.