An AI Insurance Model Undervaluing Injury Claims: How Do Lawyers Evaluate Settlement Value and Trial Risk?

When an insurer's software prices an injury claim below what the case is worth, lawyers do not argue with the software — they rebuild the number from the evidence and then price the risk of walking away from it. Settlement value comes from documented damages, clear liability and the insurer's own exposure at trial; trial risk comes from win rates, median awards, delay and cost, which together usually argue for settling rather than trying the case.

Claims software is not new and not secret. Allstate's Colossus, the program that guides bodily-injury offers, was the subject of an 18-month multistate regulatory examination that ended in a $10 million agreement in 2010, and the newer wave of AI-driven claim tools now sits under a National Association of Insurance Commissioners model bulletin adopted in December 2023. Both matter to an individual claimant less as a scandal than as leverage.

Official resources:

Read the official notice from Oregon — Use this primary source to verify the official announcement.
Explore the official data from Ojp — Use this primary source to review the underlying data.

What a claims valuation model actually does
Why the model's number is an input, not a ceiling
The standards a lawyer can actually invoke
How trial risk gets quantified
What representation changes, and what it costs in time
The benchmarks are moving upward
Frequently Asked Questions

What a claims valuation model actually does

Colossus and its successors do not read a file the way an adjuster does. The program converts an adjuster's answers about diagnosis, treatment type and treatment duration into "severity points," then maps those points to a dollar range. According to Nolo's explainer on the program, it codes roughly 600 to 720 injury types against diagnostic codes, and each insurer sets its own points-to-dollars conversion by region. That conversion setting — "tuning" — is the lever.

The same injury with the same treatment can produce different offers depending on how a carrier has calibrated its regions. In the multistate exam, regulators found Allstate had not tuned the program uniformly across claims regions, and the company agreed to a $10 million payment with 41 additional states later signing on. One limitation deserves to be stated plainly, because it cuts against the popular version of this story. As Insurance Journal reported on the settlement, the examination found no systemic underpayment of bodily injury claims, and the $10 million funded a regulator training fund for reviewing claims software — not restitution to policyholders. A low offer is evidence about your file, not proof of a scheme.

Why the model's number is an input, not a ceiling

A valuation model prices what the adjuster typed in. That makes the inputs the first place a lawyer looks: missed treatment dates, an injury coded at a lower severity than the records support, a gap in care read as recovery, or a diagnosis recorded without the imaging that confirms it. The practical response is documentation, not argument about algorithms.

Complete records, a clear causation opinion from the treating physician, wage-loss proof, and a written account of daily limitations all change the coded inputs the software depends on. So does eliminating the ambiguity that drives conservative coding — an unexplained six-week gap between visits is worth more to the insurer than it is to you. Liability is scored separately from damages. Where fault is contested or comparative negligence is arguable, the model discounts the range, so police reports, scene photographs, witness statements and any admission in the recorded statement matter as much as the medical file.

The standards a lawyer can actually invoke

Two documents do most of the work. The baseline is NAIC Model Act 900, the unfair claims settlement practices model, which makes it an unfair practice to fail to effectuate a prompt, fair and equitable settlement of claims in which liability has become reasonably clear. Its limits are real: the conduct generally must be flagrant or frequent enough to constitute a general business practice, and the model itself creates no private cause of action. The newer standard is aimed directly at automated valuation.

The NAIC model bulletin on the use of artificial intelligence systems by insurers, adopted December 4, 2023, requires insurers to maintain a written AI systems program, expressly covers claim administration and payment, and states that AI use must not produce unfair claims settlement practices. Section 4 lists the documents a state insurance department may demand in a market conduct action. Adoption is uneven. The NAIC tracks state uptake on a live implementation map, with roughly two dozen jurisdictions having adopted the bulletin; some non-adopting states have applied bulletin-aligned expectations through market conduct examinations instead. Before you frame a complaint around it, check whether your state has adopted it — and remember that the enforcement channel is the insurance department, not your settlement negotiation.

How trial risk gets quantified

Trial risk is not a feeling. It is a set of numbers, and the numbers are more sobering than headline verdicts suggest. Bureau of Justice Statistics data on civil bench and jury trials in state courts found plaintiffs won 54% of jury trials and 68% of bench trials, with a median tort award of about $24,000 and a median motor-vehicle award for winning plaintiffs of $15,000.

Trial is also the rare path. BJS estimated that bench and jury trials accounted for about 4% of all tort dispositions, which means settlement valuation — not verdict value — governs nearly every case. A lawyer pricing your case is pricing a settlement, with trial as the alternative that gives the settlement its floor. The arithmetic a lawyer runs looks like this:.

Expected value: likely verdict range multiplied by the probability of winning, then discounted for comparative fault.
Costs subtracted: expert fees, deposition and court costs, and the fee structure at trial versus settlement.
Collectability: policy limits and the defendant's assets cap what a verdict is actually worth.
Time: months or years of delay, and what that delay costs you now.
Downside: a defense verdict pays nothing and leaves costs behind.

What representation changes, and what it costs in time

Representation moves the number and slows the clock, and the trade is measurable. The Insurance Research Council's study of attorney involvement in auto injury claims drew on 7.4 million closed claims from nine insurers: attorney involvement among claimants rose from 40% in 2017 to nearly 50% in 2022, average bodily injury payments rose from about $14,000 to over $20,000, and represented BI claims took a median of roughly 440 days to close — more than double the time for unrepresented ones. Read both halves.

Higher average payments do not translate mechanically into a better net outcome on a small, clearly documented claim with a fast offer. On a serious injury with contested liability, disputed causation or a layered treatment history, the delay is usually the price of getting the file properly valued in the first place. The claims worth the wait share features: significant or permanent injury, ongoing treatment, disputed fault, a low offer arriving before treatment has ended, or an adjuster's number that ignores wage loss entirely.

The benchmarks are moving upward

Insurer valuation models are recalibrating against a rising market, which is useful context when an offer feels anchored to an older number. According to CCC Intelligent Solutions' Crash Course 2026 report, bodily injury passed auto physical damage for the first time in 2025 at 52.3% of combined dollars paid, up from 44.4% in 2022, with bodily injury claims now appearing in one of every four physical-damage claims and verdicts of $10 million or more up 52% year over year. That trend does not set the value of any single file.

It does mean that an offer justified as "what these claims are worth" is being measured against benchmarks that have shifted, and that asking the adjuster to state the basis for the figure is a fair question rather than an aggressive one. If you want the reasoning behind a number, ask in writing for the specific factors used to evaluate the claim, the injuries and treatment the carrier recorded, and whether any of it was excluded. Keep the response. If the file shows records the carrier never reviewed, a treatment period cut short in its coding, or an injury coded below the diagnosis, that written answer is the document a lawyer or a state insurance department examiner will work from — and in states that have adopted the NAIC bulletin, the carrier is already required to keep the documentation behind it.

Frequently Asked Questions

Can I demand to see the software's output on my claim?

There is no general consumer right to the raw output. The NAIC model bulletin requires insurers to document their AI systems and lets a state department demand those records in a market conduct action, so the route runs through your insurance commissioner, not the adjuster.

Does a low offer mean the insurer broke the law?

No. Model Act 900 targets conduct flagrant or frequent enough to be a general business practice, and the 2010 Allstate examination found no systemic underpayment of bodily injury claims. A single low offer is a negotiating position to answer with evidence.

Is the median motor-vehicle award really only $15,000?

That BJS figure is the median for plaintiffs who won at trial, so half of winning awards were higher. It also excludes the 46% of jury plaintiffs who lost, which is why expected value, not the top of the range, drives settlement decisions.