How Insurance Actually Works, and Where AI Fits

Insurance is unusual among businesses: you collect money today for a promise to pay an uncertain amount at an unknown future date, or possibly never. The product is a contingent promise, and the entire economics depend on predicting how often and how expensively that promise will be called in.

That single fact explains why insurance is so tightly coupled to data. A manufacturer knows its costs when it makes a product. An insurer does not: it sets a price before knowing what the thing will cost, because the cost is a future claim that has not happened yet. The industry's whole apparatus, actuaries, underwriters, reserves, exists to manage that inversion.

And it explains why AI landed here so naturally. Insurance is a forecasting business wearing a financial costume. Machine learning is a forecasting technology. The fit is structural, not fashionable.

This cursus works through that fit in three parts:

• This lesson: how insurance works, and where AI attaches to it.
• Lesson 2: the operational core, AI in underwriting, pricing, claims, and fraud.
• Lesson 3: the regulation, fairness duties, and honest limits that constrain all of it.

We start with the mechanism, because you cannot judge where AI helps without knowing what the business is actually doing.

The foundation is risk pooling. Any individual's risk is wildly unpredictable: whether your house burns down this year is essentially unknowable, and catastrophic if it happens. But across a large pool of similar houses, the proportion that burn is remarkably stable year to year.

This is the law of large numbers doing the work. Individual outcomes are volatile; averages over many independent outcomes converge. So an insurer that cannot predict any single policyholder can predict a portfolio quite well. Each person swaps an unbearable, unpredictable risk for a small, certain premium, and the insurer absorbs the volatility by aggregating it away.

Two consequences follow, and both matter for AI:

• Insurers need scale and independence. Pooling only smooths risk if the risks are not all correlated. This is why a single hurricane or pandemic is dangerous in a way that a million unrelated fender-benders is not: correlated risk defeats the averaging.
• Accuracy is about groups, not individuals. The business does not need to know if you will crash. It needs the expected claim cost of people like you to be right. That is a statistical estimation problem, precisely the shape of problem machine learning attacks.

Pooling is why insurance exists at all. Prediction is how it is priced.

How is a premium actually built? Conceptually it is a sum:

Premium  =  Expected loss  +  Expenses  +  Risk margin / profit
              (the pure          (running     (capital cost and
               premium)           the firm)    uncertainty buffer)

The expected loss, often called the pure premium, is the forecast: how much this risk will cost in claims, on average. Everything else is loaded on top. Get the expected loss wrong and no amount of operational excellence saves you.

Insurers measure whether they got it right with two standard ratios:

• Loss ratio = claims paid / premiums earned. How much of the premium went out as claims.
• Expense ratio = operating expenses / premiums.
• Combined ratio = loss ratio + expense ratio.

The combined ratio is the industry's scoreboard, and the reading is sharp: below 100 percent means an underwriting profit (you took in more than you paid out and spent), above 100 percent means an underwriting loss. A combined ratio of 97 percent means 3 cents of underwriting profit per premium dollar.

Notice how thin that is. Insurance margins are typically narrow, which means small improvements in prediction accuracy or expense are disproportionately valuable. A few points of loss ratio is the difference between a good year and a bad one. That economics is exactly why insurers invest heavily in modeling, and it frames every AI use in the next lesson.

Two classic problems shape everything insurers do, and both are information problems, which is why data is the industry's battleground.

Adverse selection. The customer knows more about their own risk than the insurer does. People who expect to claim are more eager to buy, and more willing to pay. If an insurer prices everyone at the pool average, the low-risk customers find it a bad deal and leave, which raises the average risk of who remains, which forces a price rise, which drives out the next tier. This unravelling spiral is why insurers must distinguish risks: pricing accuracy is not greed, it is survival. An insurer that prices worse than its competitors gets adversely selected against, it systematically wins the customers its rivals correctly declined.

Moral hazard. Being insured changes behavior. Someone fully covered for theft may be less careful about locking up. This is why policies carry deductibles and coverage limits: leaving the policyholder with some exposure preserves their incentive to avoid the loss.

Both explain why insurers hunger for information that predicts risk, and why the industry's history is a long march toward finer risk classification. It also previews the tension of Lesson 3: the drive to distinguish risks ever more finely runs directly into rules about which distinctions society permits.

An insurer's work runs along a chain, and knowing the stages is how you locate any AI claim you hear.

• Distribution and marketing: finding and acquiring customers, direct, through brokers, or via agents.
• Underwriting: deciding whether to insure a given risk and on what terms. Risk selection.
• Pricing / rating: deciding what to charge, applying the rating model to the risk's characteristics.
• Policy administration: issuing, servicing, renewing, endorsements.
• Claims: the moment of truth. Intake (first notice of loss), investigation, valuation, settlement, payment.
• Reserving: setting aside money for claims that have happened but are not yet fully paid, including claims incurred but not reported. An accounting and actuarial estimate of future liability.
• Reinsurance and capital: insurers insure themselves, ceding some risk to reinsurers, especially for correlated catastrophe exposure.

Two stages dominate the economics. Underwriting and pricing determine the loss ratio, whether you priced the risk correctly. Claims is where the money actually leaves, and where both the customer experience and a large share of expense live.

So it is no accident that those are exactly where AI has concentrated, which is the subject of Lesson 2. The chain also frames a point worth carrying: the promise you sold is kept, or broken, at the claims stage.

Put the pieces together and the affinity is structural. Insurance has every property that makes machine learning pay.

• The core task is prediction. Estimating expected loss is literally a supervised regression problem: given features of a risk, predict a future cost.
• The labels arrive by themselves. Claims are the ground truth. Unlike most industries, which must pay to label data, insurers automatically accumulate labeled outcomes as a by-product of operating: every policy year produces the answer to "did this risk cost what we thought?"
• The data is deep and long-running. Insurers have decades of structured policy and claims records, plus mountains of unstructured material, adjuster notes, photos, medical reports, police reports, that older methods could not use.
• Margins are thin and volumes are huge. With combined ratios near 100 percent, small accuracy or expense gains scale into real money across millions of policies.
• There is a long modeling tradition. Actuaries have used statistical models for over a century, so AI is a continuation, not an invasion. Generalized linear models have been standard in rating for decades; modern methods extend that lineage.

That last point deserves emphasis, because it corrects a common misconception. Insurance did not discover statistical modeling with AI. The industry was quantitative before computing existed. What changed is the class of models and the kinds of data they can consume, which is precisely what Lesson 2 examines.

Mapping AI onto the value chain gives the shape of the rest of the cursus.

That bottom row is not a footnote. Insurance is a regulated industry in which pricing decisions are legally constrained, and the same drive toward finer risk classification that adverse selection demands is the thing regulators scrutinize hardest. An insurer cannot simply deploy the most accurate model it can build.

So the honest frame for this cursus is a business whose core activity, predicting and pricing risk, is a natural fit for machine learning, operating under rules about what predictions it is permitted to act on. Lesson 2 covers the machinery. Lesson 3 covers the rules, which are not an afterthought but a design constraint on every model an insurer ships.

Premiums flow in from distribution through underwriting and pricing, which set the loss ratio; claims flow out and, with expenses, determine the combined ratio, while reserving and reinsurance absorb uncertainty and correlated risk.