The Perfectly Optimized Insurance Company

By James W. Moore

Key Takeaways

• Faster settlements, lower costs and greater claimant satisfaction can be genuine improvements. None independently establishes that claimants received fair compensation.
• AI makes it possible to move from estimating what a claim is worth toward predicting what a particular claimant might accept. Those are different objectives.
• Human approvals and regulatory compliance matter, but management must decide what the system should never be rewarded for optimizing.

The best year the claims department ever had

Imagine a third-party bodily injury claimant dealing with an auto insurer after an accident. Treatment is complete, the documentation is in, and the carrier offers a prompt settlement. The adjuster explains the offer, answers questions and doesn’t pressure the claimant. The claimant accepts, glad to avoid a long negotiation and the expense of hiring an attorney.

The claim closes. The claimant is satisfied. Everything appears to have worked as intended.

But suppose an independent evaluation of that same claim would have supported a meaningfully higher settlement. No new medical crisis needs to emerge. No one needs to have lied. The claimant may never learn that the accepted amount was lower than a reasonable independent valuation.

Now imagine that same insurer across thousands of claims. Its senior vice president of claims has set a straightforward goal: reduce bodily injury severity without increasing litigation or compromising the claimant experience. An AI-assisted program helps settle claims faster, lower expenses and improve satisfaction. Complaints decline. Adjusters review the recommendations. Compliance monitors the process.

It’s the best year the department has had in a long time.

Those results matter. The question is what, precisely, they prove.

Faster can genuinely be better

There is a sound argument for making bodily injury settlements faster and less expensive.

The Insurance Research Council published a study in July 2026 using more than 7.4 million auto injury claims closed with payment, from nine insurers representing roughly 43% of the private passenger auto market. The claims covered mid-2017 through mid-2022. In the bodily injury portion of the data, claimants with attorneys waited a median of nearly 440 days for closure, more than twice as long as unrepresented claimants.

IRC also found that represented bodily injury claimants received $1.40 in net payment for every dollar of medical expense, compared with $1.80 for unrepresented claimants, after medical costs and estimated legal fees. For that calculation, attorney fees were assumed to equal one-third of the settlement. The study did not measure each claimant’s actual attorney bill.

For an injured person, time and legal expense are not trivial. Faster payment can be valuable. A claim settled without unnecessary dispute can leave everyone better off.

But the findings need context. The data are observational. Injuries involving attorneys may be more severe, more complicated or more contested from the beginning. IRC is supported by property-casualty insurers and industry associations. Its figures describe differences between groups, not what would have happened had the same claimant made the opposite decision about representation.

Our fictional executive can nevertheless take the findings seriously. Reducing avoidable expense and delay is a legitimate management objective, and often a benefit to claimants.

The risk arises when a finding about averages becomes advice to one injured person deciding whether to sign a release.

From valuing the claim to valuing the claimant

That problem predates today’s AI.

In Jacobsen v. Allstate, the Montana Supreme Court described the insurer’s Claims Core Process Redesign, developed in the 1990s with research and recommendations from McKinsey. The court characterized the strategy as seeking lower claims payouts through earlier settlements with unrepresented claimants and reduced attorney involvement. It recounted internal research suggesting that represented claims settled for two to three times more.

The court also described Robert Jacobsen’s own claim. Following a 2001 accident caused by an Allstate insured, he settled without counsel six days later for $3,500 plus 45 days of open medical payments, and signed a release. About three weeks later he began experiencing significant pain. After retaining an attorney, he secured the release’s rescission, and Allstate ultimately settled the underlying claim for $200,000.

The 2013 Supreme Court decision concerned class certification, not a final judgment that the entire claims program was unlawful. The plaintiffs’ allegations about the program are not the same thing as a merits finding. But the court’s account of the documents and procedures reveals an important tension.

One of those procedures was an attorney-economics script. According to the opinion, adjusters were instructed to tell claimants that attorneys often charged 25 to 40 percent of a settlement and that claimants were free to retain counsel. The script also cautioned against improperly discouraging representation. What it did not disclose, the court noted, was Allstate’s own research that represented claims generally settled for more.

That is where the aggregate becomes personal. A person could hear true information about legal fees and still receive an incomplete picture of the tradeoff. Similarly, an insurer today could accurately cite IRC’s finding that unrepresented claimants, on average, netted more per dollar of medical expense. It would not follow that remaining unrepresented was the best decision for the person on the phone.

Nor do the two studies measure the same thing. The historical internal research concerned gross settlement amounts. IRC’s modern comparison is net payment relative to medical expense, using an assumed attorney fee. Both can be true without resolving the fairness of an individual settlement.

For decades, computerized valuation tools have been used to help answer a familiar question: What is this claim worth?

More capable AI could help answer a different one: What is this claimant likely to accept, and when?

Those are not necessarily the same number.

A model that assesses injuries, liability, treatment and reasonable compensation starts with claim value. One that learns which individual characteristics predict acceptance starts somewhere else. That distinction is a potential use of AI, not an accusation that a particular insurer currently deploys such a system.

Management should settle that distinction before the program starts producing results.

The perfectly green dashboard

Back at our fictional insurer, the results look excellent. Average paid severity is lower. Claims close faster. Satisfaction improves. Fewer people complain or hire attorneys. Adjusters and compliance staff are following established procedures.

But acceptance is not an independent measure of fair value. Satisfaction reflects the experience the claimant believes they had, not necessarily the outcome an informed, independent review would support. And a claimant who doesn’t suspect a shortfall is unlikely to file a complaint about it.

That leaves a blind spot. Overall results can improve while a smaller group receives less than its claims reasonably warrant. The dashboard might never show that group if nobody measures it separately.

This is not evidence that fast settlements are improper, or that lower severity equals underpayment. Reducing unnecessary expense while paying claims fairly is a genuine achievement. The issue is whether the carrier can tell those savings apart from reductions that come at claimants’ expense.

The NAIC’s Unfair Claims Settlement Practices Model Act addresses prompt, fair, equitable and good-faith practices and certain patterns of misconduct. States adopt and enforce their own laws, not a single national version of the model. But the practical question exists even where a company satisfies its reporting and compliance obligations: how do you detect harm the affected people have no reason to recognize?

An insurer can have excellent evidence that its system is efficient without having equally good evidence that every aspect of its objective is appropriate.

The humans approved everything

Perhaps the most reassuring fact about our fictional insurer is that the humans never left the process.

Adjusters review recommendations. Compliance checks the rules. Model-risk teams document performance and test whether the system behaves as designed. Management receives reports, and exceptions have an escalation path.

Those are valuable controls. None should be discarded.

But consider the chain of decisions. Management decides what success means. Technologists build a system to improve that measure. Adjusters and reviewers check recommendations against the rules management approved. Senior executives then evaluate the result against the same performance targets.

Who is independently testing whether the target itself creates an unacceptable incentive?

The NAIC’s 2023 AI Model Bulletin, which takes effect only where a state adopts it, emphasizes insurer accountability, governance, testing and documentation for AI-supported consumer decisions. That is important. It does not relieve executives of deciding which objectives are appropriate in the first place.

In The Carrier Owns the Decision Architecture, I argued that responsibility remains with the insurer regardless of which decisions technology helps execute. That remains true here, but ownership is only the beginning.

A human can approve every individual recommendation without independently challenging the business objective behind all of them.

The AI worked perfectly. The humans approved every decision. The dashboard was green.

Values that cost something

This is not an argument against efficiency. Unnecessary delays, inconsistent valuations and avoidable litigation are costly for claimants as well as insurers. Better processes are worth building.

The question is what the carrier is optimizing for.

If the company says fairness matters, that value has to be visible in decisions that may cost it money. An independent review of accepted settlements, including those involving unrepresented claimants, could compare outcomes against claim value rather than acceptance alone. Management could examine whether particular groups are consistently receiving lower relative settlements. It could explicitly prohibit using a claimant’s predicted financial vulnerability or willingness to accept less as a substitute for evaluating injury and liability.

Those are not merely technical settings. They’re management choices about what counts as success.

They may also mean that the insurer declines a profitable opportunity. That’s what makes them evidence of institutional values, rather than another statement on a corporate website.

A company’s real values are revealed by what it is willing to sacrifice to uphold them.

When is enough, enough?

Imagine the senior vice president of claims reviewing another strong quarter. The program has reduced delays and eliminated waste, and it continues to find ways to reduce severity a little further.

Some of those improvements may be entirely legitimate. Others may simply identify people more likely to accept less.

Somewhere in those numbers is the claimant from the opening: paid promptly, satisfied, and perhaps never aware that an independent review might have supported a different amount.

The executive has a choice that no efficiency metric can make. Not just how much further the system can optimize, but where it should stop.

Because a company that doesn’t know when to stop optimizing probably shouldn’t build a system that never gets tired of trying.

Sources

• Insurance Research Council, Auto Injury Insurance Claims: A Study of Increasing Claim Severity, 2026 summary
• Insurance Research Council, July 2026 study summary PDF
• Montana Supreme Court, Jacobsen v. Allstate Insurance Company, 2013 MT 244, court opinion PDF reproduced by Justia
• NAIC, Unfair Claims Settlement Practices Model Act, Model #900
• NAIC, Members Approve Model Bulletin on Use of AI by Insurers, December 2023