TL;DR — To accurately forecast litigation escalation, systems must triangulate between sparse settlement data and verdict-heavy public records without blending them. This separation prevents outliers from inflating day-one reserves and allows for honest uncertainty bands.
Public dockets are cemeteries of worst-case scenarios. If an engineering team scrapes state court records to build a predictive model, that model will inevitably conclude that every slip-and-fall ends in a multi-million dollar jury verdict. This is the reality of selection bias. The vast majority of insurance claims settle quietly, leaving little to no public trace. When we set out to build a forecasting platform for claims and litigation, we had to confront a fundamental data engineering problem. We needed to calculate an accurate trial escalation probability when the accessible comparable dataset is heavily skewed toward cases that went the distance.
The legal system filters cases in a highly predictable way. As articulated by the Priest-Klein hypothesis, the disputes that actually reach a verdict are those where the plaintiff and defendant have mutually incompatible estimates of the outcome. The obvious cases settle early. The public data we see is the residual fraction of disputes where negotiations failed. If a machine learning pipeline treats public verdicts as a representative sample of all claims, its predictions will be structurally flawed. The model will recommend reserves that are orders of magnitude too high, unnecessarily tying up capital and distorting the carrier's financial posture.
Solving this requires a strict architectural separation of concerns within the platform. In our system, generative AI has exactly one job: reading and structuring the unstructured case file. It ingests thousands of pages of pleadings, medical records, and correspondence, extracting the underlying facts of the claim. We built the ingestion pipeline to handle massive document loads with low latency, ensuring that highly sensitive medical records are processed securely and transiently. The generative model does the reading. It does not predict the outcome. Prediction is handled by a separate set of geometric machine learning models trained on large numbers of resolved cases with known outcomes. This separation is critical for system reliability and security. If we allowed an LLM to predict case value, it would be easily manipulated by the dramatic, trial-oriented language often found in a plaintiff's initial demand letter. By restricting the LLM to factual extraction and passing those structured facts to a deterministic mathematical model, we maintain traceability by design. Every variable used in the prediction maps directly back to a specific sentence in the source documents.
Triangulating Across Asymmetric Datasets
To generate a reliable escalation probability, the predictive models must account for the silent majority of settlements. We cannot simply take a verdict and apply a flat percentage reduction to estimate a settlement value. That approach ignores the underlying dynamics of liability, damages, and venue. Instead, the system maintains distinct mathematical spaces for settlement-anchored data and verdict-heavy comparable cases. Keeping verdict-heavy comps in a separate module from settlement-anchored data prevents the extreme outliers from dragging the median settlement prediction upward.
When the system evaluates a new claim via our API, it maps the extracted facts against these distinct distributions. If the facts align closely with a cluster of cases that all proceeded to trial, the escalation probability rises. But this calculation is constrained by field-specific priors. The model incorporates baseline settlement rates for the specific jurisdiction and injury type. It triangulates between the sparse, highly verified settlement data provided by closed claim files and the voluminous, heavily biased public docket data. This triangulation requires honest uncertainty bands. If the comparable set for a traumatic brain injury in a specific county consists entirely of nuclear verdicts, the system will not invent a tight settlement range. It outputs a calibrated settlement range and explicitly flags the variance. The platform is designed to tell the claims professional exactly why the uncertainty exists. The drivers behind the output, including the specific comparable resolved cases, are exposed directly in the user interface. We do not hide behind a black box single-point guess.
Integrating Liability and Limits
Escalation probability is deeply tied to liability probability and collectability. A plaintiff might demand a figure based on a massive verdict they found in a similar case. If the defendant's policy limit is a fraction of that demand and there are no aggregate corporate assets to pursue, that verdict comparable is mathematically irrelevant to the settlement value. The system factors these constraints into the prediction, ensuring the output reflects the actual financial mechanics of the claim rather than a theoretical maximum exposure.
By modeling the probability of establishing liability alongside the damages, the system identifies cases where the plaintiff has a strong incentive to push for trial. Third-party litigation funding often distorts these incentives, providing the financial runway for plaintiffs to reject reasonable settlement offers. Social inflation further complicates this by raising the perceived value of a trial outcome in certain venues. The platform detects these structural anomalies early in the lifecycle of the claim, providing a reserve delta versus the current reserve. When the system outputs a high escalation probability, it is signaling a fundamental misalignment between the objective facts of the case and the likely demands of the plaintiff. This data dictates the operational response. A high escalation probability justifies a larger allocation of defense spend and immediate preparation for litigation. A low probability signals an opportunity for early mediation.
Our onboarding process focuses heavily on aligning these outputs with existing claims workflows. System integration means routing the escalation probability and settlement range directly into the carrier's core platform via API on day one. The objective is to set realistic reserves immediately and negotiate from a position grounded in calibrated data rather than gut instinct. You cannot build a reliable forecasting system by pretending the public record is the whole truth. The cases that go to trial are the exceptions, and a useful model must understand exactly why they failed to settle.
Related articles.
Want to talk to an executive?
Press, partners, investors, candidates — the inbox is monitored. Tell us who you are and we'll route it to the right person within two business days.