Running the Case 10,000 Times

A malpractice file lands on a defense lawyer's desk: a missed diagnosis, a permanent injury, a grieving family, and a demand letter with a number on it. For most of the modern era, the response to that number was a ritual of experience, pull a few comparable verdicts, consult a colleague who tried something similar, and arrive at a single figure for the reserve and the settlement authority. The trouble is that litigation is not a single number. It is a probability cloud, shaped by which judge draws the case, how a jury reads an injury, how an expert performs on cross, and which of a dozen contingent events actually happens. Outcome-simulation tools exist to model that cloud directly, and in healthcare, where claims take years to resolve and the tail risk is measured in tens of millions, the shift is reshaping legal strategy from the ground up.

The numbers behind the discipline explain the urgency. A long-run analysis of every malpractice payment reported to the U.S. National Practitioner Data Bank counted roughly 530,000 payments totaling about $136.4 billion between 2000 and 2025, even as annual case volume fell about 40% over the same period. The cost is not only in the payouts. Researchers writing in Health Affairs estimated the total annual cost of the U.S. medical-liability system, including defensive medicine, at $55.6 billion in 2008 dollars, about 2.4% of national health spending. Every dollar of that system rides on someone's estimate of how a future case will turn out.

The Old Way: One Number, A Lot of Faith

The legacy approach to valuing a healthcare claim was a single point estimate dressed up as judgment. A claim would be assigned a reserve, the insurer's best guess of its ultimate settlement value, and that figure flowed straight into financial statements and settlement negotiations. The estimate was rarely a distribution; it was a number arrived at by analogy and seasoned instinct, then revised reactively as discovery unfolded.

This mattered enormously because of how malpractice claims behave over time. Industry accounting bodies note that the average gap between an incident of medical negligence and the payment of a claim is four to five years, with many claims taking longer. A reserve set on instinct in year one had to survive years of contingent events. Actuarial standards openly concede the fragility: estimates are "inherently uncertain because they are dependent on future contingent events," and the amount actually needed to settle "can be significantly different from the stated reserve amount." When juries in a jurisdiction shift toward larger awards, every open claim has to be re-priced at once.

Litigation outcomes themselves are deeply non-linear, which point estimates conceal. An analysis of malpractice litigation against U.S. physicians found that about 55% of claims involved litigation, that 54% of litigated claims were dismissed by the court, and that just 4.5% reached a trial verdict, where defendants prevailed in nearly 80% of cases. The mean time to close a claim was 19 months overall but stretched to 39 months for cases that went to a defense verdict at trial. A single average buried all of that structure.

The Shift: From Point Estimates to Distributions

The intellectual core of outcome simulation is old; the scale is new. Decision-tree analysis, mapping each contingent event in a case as a branching node, assigning probabilities and dollar values, and rolling the tree back to an expected value, has been a staple of sophisticated litigation risk work for decades. Forensic and valuation practitioners describe how a decision tree produces a "risk-adjusted damages value," and how overlaying a simulation generates a full distribution of outcomes rather than one figure, giving counsel both a central tendency and the shape of the tail (Baker Tilly; Cogent Valuation).

Monte Carlo simulation is what scales the decision tree from a handful of branches to thousands of runs. Instead of one liability probability and one damages number, each uncertain input is expressed as a range, and the model samples those ranges repeatedly, "as if the case were tried 100 times," in the language of one widely used modeling workflow, to build an empirical distribution of net outcomes. The output is not a verdict prediction but a risk profile: the probability the exposure exceeds the settlement demand, the chance of a low-probability, high-severity result, and the zone of possible agreement that should anchor negotiation (Cornerstone Research; decision-analysis literature).

A point estimate tells you the case is worth $400,000. A simulation tells you there is a one-in-eight chance it costs you $5 million, and that is the number that changes how you negotiate.

Layered on top of the simulation machinery is a second wave: data-driven prediction trained on the historical record. Researchers studying the European Court of Human Rights showed that a model using only the textual content of cases could predict outcomes with 79% accuracy on average, with the factual narrative the strongest predictor. A separate study of nearly 8,000 U.S. Supreme Court decisions across two centuries reported a model that correctly forecast about 70.2% of case outcomes using only information available before each decision. These are not healthcare cases, but they establish the principle that structured features of a dispute carry real predictive signal that can feed probability estimates inside a simulation.

That divergence, falling case counts, rising average payments, is exactly the environment that rewards probabilistic modeling. When the distribution of outcomes is skewed by a small number of very large payouts (only about 2.7% of paid claims in the NPDB dataset exceeded $1 million, yet they dominate total dollars per Settlement Insight), the mean is a poor guide to risk. Simulation surfaces the tail that an average hides.

What It Looks Like Now

In a present-day healthcare legal team, outcome simulation shows up at four decision points, and the workflow is increasingly generic across the platforms emerging in this space.

Reserve setting and case triage

When a claim is opened, structured features, injury severity, error type, venue, specialty, and the strength of the medical record, feed a model that returns a probability distribution of ultimate cost rather than a single reserve. This matters because the loss categories that dominate are well-mapped: a ten-year NPDB analysis found that treatment-related and diagnosis-related errors accounted for the largest share of payouts, with diagnostic errors the single most costly category (Gill Ports Hoste; Munley Law on NPDB data). Modeling teams use those base rates as priors, then adjust per case.

Settlement modeling and scenario comparison

The decision-analysis question, settle now or litigate, becomes a side-by-side comparison of distributions. The simulation produces an expected value for each path plus the variance, letting counsel weigh not just the mean but the risk of an outcome far worse than expected. This is precisely the structure healthcare defense work needs, given that trial-decided claims historically produced median payouts at least 2.5 times larger than settled claims in several states (U.S. Bureau of Justice Statistics).

Reserve adequacy and portfolio risk

At the book level, insurers and self-insured health systems run simulations across an entire portfolio of open claims to test reserve adequacy. The discipline is non-trivial: industry analysis of "social inflation" found that, for one claims-made medical-malpractice book, booked ultimate losses ran $3.99 billion, about 11.2%, above what later experience justified, a reminder that adverse development cuts both ways and that probabilistic reserving is a guard against both over- and under-reserving.

Adoption is real, and so is the caution

The appetite for these tools tracks the broader, rapid embrace of analytics in legal work. Survey data from the Thomson Reuters Institute found that legal professionals overwhelmingly expect AI-driven tools to become central to daily workflow, with 95% anticipating that within five years. But the same body of research shows the brakes are on: respondents named inaccuracy (70%) and data security (68%) as top concerns in the 2024 generative-AI survey, and a separate report noted that only about 10% of law firms had a formal AI policy in late 2024.

The Next Few Years

Three shifts look likely over the next three to seven years. First, simulation will move upstream from litigation into clinical-risk prevention. Because the loss pool is concentrated in identifiable error categories, and because researchers estimate that fewer than 1% of medical errors ever become a claim while a Harvard-led study found nearly three-quarters of physicians practice defensive medicine, health systems have strong incentive to use predicted-exposure models to target the events most likely to generate catastrophic claims, rather than ordering tests reactively.

Second, expect convergence between the two analytic traditions described above: human-estimated decision trees and data-trained prediction. Today, decision-analysis practitioners are explicit that probabilities and damages are estimated by counsel, "not AI." That boundary will blur as models trained on verdict and settlement histories supply calibrated priors that experts then adjust, a hybrid in which the machine proposes a distribution and the lawyer interrogates it.

Third, governance will harden. Courts and bench guidance already flag that outcome-prediction tools raise concerns "about how such predictions might impact the delivery of justice," and prediction research itself shows the limits, models stumble precisely on the close, similar cases where two near-identical fact patterns diverge (BBC on the ECHR study; judicial bench guidance).

Conclusion

The transformation underway in healthcare litigation is less about replacing lawyers than about replacing a single number with a picture of risk. A demand letter still arrives with one figure on it, but the response is increasingly a distribution, a map of how a case could break across thousands of simulated runs, anchored where possible in the historical record and always interrogated by human judgment. In a system that moves $136 billion through its claims over a quarter-century and takes half a decade to close a single file, the move from intuition to distribution is not a luxury. It is how a defensible strategy gets built. The open question is not whether the tools work, but whether the people using them remember that a confident-looking curve is still, at bottom, a set of assumptions waiting to be tested.

Sources

1. Settlement Insight, "We Analyzed 529,804 Medical Malpractice Payments" (NPDB, 2000 to 2025). https://settlementinsight.com/research/medical-malpractice-settlements
2. Mello, Chandra, Gawande & Studdert, "National Costs of the Medical Liability System," Health Affairs, 2010 (PMC). https://pmc.ncbi.nlm.nih.gov/articles/PMC3048809/
3. NAIC, "Medical Malpractice Insurance Report: A Study of Market Conditions." https://content.naic.org/sites/default/files/inline-files/topics_Med_Mal_Rpt_Final.pdf
4. Jena, Chandra, Lakdawalla & Seabury, "Outcomes of Medical Malpractice Litigation Against US Physicians," JAMA Network, 2012. https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/1151587
5. Baker Tilly, "Exposure analysis: expanding the use of modeling and decision trees." https://www.bakertilly.com/insights/exposure-analysis-expanding-the-use-of-modeling
6. Cogent Valuation, "Valuation of Litigation Claims" (decision tree & simulation). https://www.cogentvaluation.com/wp-content/uploads/2014/08/ValuationOfLitigationReprint.pdf
7. Cornerstone Research, "Applying Monte Carlo Simulations in Litigation." https://www.cornerstone.com/insights/articles/applying-monte-carlo-simulations-in-litigation/
8. Eperoto, "An Introduction to Decision Analysis for Corporate Counsel." https://eperoto.com/articles/an-introduction-to-decision-analysis-for-corporate-counsel/
9. Aletras, Tsarapatsanis, Preoţiuc-Pietro & Lampos, "Predicting judicial decisions of the European Court of Human Rights," PeerJ Computer Science, 2016. https://eprints.whiterose.ac.uk/id/eprint/144020/1/peerj_cs_93.pdf
10. Katz, Bommarito & Blackman, "A general approach for predicting the behavior of the Supreme Court of the United States," PLOS ONE, 2017 (PubMed). https://pubmed.ncbi.nlm.nih.gov/28403140/
11. BBC News, "AI predicts outcome of human rights cases," 2016. https://www.bbc.com/news/technology-37727387
12. Gill Ports Hoste, "Medical Malpractice in America: A 10-Year Analysis" (NPDB). https://www.gphlaw.com/medical-malpractice-in-america-a-10-year-analysis/
13. Munley Law, "Medical Malpractice Statistics" (NPDB summary). https://munley.com/medical-malpractice-lawyers/statistics/
14. U.S. Bureau of Justice Statistics, "Medical Malpractice Insurance Claims in Seven States, 2000 to 2004." https://bjs.ojp.gov/content/pub/pdf/mmicss04.pdf
15. The Doctors Company, "Social Inflation and Loss Development Report." https://www.thedoctors.com/contentassets/db04a8d14aac449d9363e19cce85b237/13739_tdc_socinflation_hex_withlinks_fr.pdf
16. Thomson Reuters, "Generative AI Adoption Nearly Doubles" (2025 report). https://www.thomsonreuters.com/en/press-releases/2025/april/from-incubation-to-integration-generative-ai-adoption-nearly-doubles-as-professional-services-reach-crossroads
17. Thomson Reuters, "2024 Generative AI in Professional Services" report. https://www.thomsonreuters.com/content/dam/ewp-m/documents/thomsonreuters/en/pdf/reports/tr4322226_rgb.pdf
18. Legal IT Insider, "Just 10% of law firms have a GenAI policy," 2024. https://legaltechnology.com/just-10-of-law-firms-have-a-genai-policy-new-thomson-reuters-report-shows/
19. Harvard Medical School, "Does Defensive Medicine 'Work'?" 2015. https://hms.harvard.edu/news/does-defensive-medicine-work-0
20. AIJA Benchbook, "Prediction of Litigation Outcomes" (judicial guidance). https://www.dfvbenchbook.aija.org.au/article/102971/3+Areas+of+AI+Use+in+Courts/3.3+Prediction+of+Litigation+Outcomes