AI helps with claims triage and severity prediction by scoring a claim's likely severity, litigation risk, and expected time to close at the moment it is reported, using features such as the coverage line, injury description, vehicle or property type, location, and even weather data for the loss date, so the claim can be routed to the right adjuster tier immediately instead of after a human reviews it manually. High predicted severity or litigation propensity routes the claim to a senior or specialized adjuster from day one, which tends to produce better outcomes than escalating a claim only after it has already grown more complex on a junior adjuster's desk. Low predicted severity, well documented claims can move onto a fast, low touch handling path, sometimes including straight through payment for small, clear cut losses. These models also improve reserve accuracy at first notice of loss, since an early, data driven severity estimate is generally more consistent than an adjuster's initial gut assessment before the file is fully developed. Triage models need periodic retraining as claim mix, costs, and litigation environment shift over time. Nanobase AI, an NVIDIA Inception Program member, builds severity and triage models trained on an insurer's own historical claims outcomes.
The feature set that actually drives a severity score
A severity model is only as good as the signals it can see at first notice of loss, before an adjuster has added anything to the file. The strongest inputs are usually structural rather than exotic: coverage line and limits, injury or damage description text, vehicle or property characteristics, geography, and time-of-loss context such as weather conditions on the date of the event. Claimant representation status, meaning whether an attorney is already involved, is one of the single strongest litigation-risk signals available and should be captured the moment it appears in any communication.
Most production severity models are gradient-boosted trees (XGBoost or LightGBM) trained on structured fields, with a large language model layered on top purely for text: pulling injury severity indicators, treatment mentions, and liability language out of the free-text loss description that a tree-based model cannot read directly. Combining both outperforms either alone, since the structured model handles the bulk of the signal cheaply while the language layer catches nuance in the narrative.
A severity model's accuracy depends more on capturing attorney involvement and narrative text early than on which algorithm scores the structured fields.
Routing tiers: turning a score into an action
| Severity / litigation signal | Routing outcome | Adjuster assigned |
|---|---|---|
| Low severity, no attorney, clean documentation | Fast-track / straight-through for small claims | Junior or automated payment path |
| Moderate severity, standard documentation | Standard queue | Generalist adjuster |
| High severity or attorney involved from day one | Priority queue, same-day assignment | Senior or specialty adjuster |
| High severity plus complexity flags (multi-party, coverage dispute) | Escalation queue | Specialty unit with technical claims lead |
The tiers themselves matter less than getting the file to the right desk before it has had time to develop past a junior adjuster's comfort zone. A claim that starts on the wrong tier and gets escalated later almost always costs more, in both dollars and litigation risk, than one routed correctly from day one.
The business value of triage comes from routing speed at first notice of loss, not from the marginal accuracy gained by a more sophisticated model.
Building the pipeline: a practical sequence
- Assemble a historical dataset of closed claims with known final severity, litigation outcome, and time to close, and confirm the fields available at first notice of loss actually existed at that point in time, not fields added later in the file's life (a common source of leakage that inflates test accuracy).
- Train the structured model first, validate against a holdout period rather than a random split, since claim mix shifts over time.
- Add the text-extraction layer for narrative fields and measure its lift over the structured model alone before committing to the added complexity.
- Define routing tiers with claims leadership, not just data science, since the tiers need to map to actual desk capacity and adjuster skill levels.
- Run the model in shadow mode, scoring live claims without changing routing, and compare its tier assignment to what an experienced adjuster would have chosen.
- Go live on a subset of claim types first, expand once the shadow comparison holds up.
Running a triage model in shadow mode against experienced adjuster judgment before go-live catches disagreements a training-time accuracy metric alone would never surface.
Monitoring after launch
Claim mix, litigation environment, and cost trends shift continuously, and a triage model trained on last year's data quietly degrades if nobody is watching. The two metrics worth tracking on an ongoing basis are routing agreement, meaning how often the model's tier assignment matches what a senior adjuster would independently choose on a sampled review, and reserve development, meaning how far the model's early severity estimate ends up from the claim's actual final cost. Both should be reviewed on a fixed cadence, with retraining triggered by drift rather than a fixed calendar, since claim patterns do not degrade on a schedule. This ties into the broader question of how AI handles first notice of loss intake, since triage quality depends on how complete the FNOL data capture already was.
Retraining triggered by monitored drift, rather than a fixed calendar, is what keeps a triage model accurate as claim mix and litigation patterns shift over time.
Frequently asked questions
Does a triage model replace adjuster judgment on a claim?
No. It decides which desk and what priority a claim gets, not how the claim is ultimately handled. The assigned adjuster still makes every substantive decision about coverage, liability, and settlement; the model only shortens the time it takes for the right person to start working the file.
How much historical data is needed to build a severity model?
There is no fixed threshold, but a model needs enough closed claims across each coverage line and severity band to learn reliable patterns, typically several years of history to cover seasonal and litigation-environment variation. An insurer with thin data on rare, high-severity claims should expect the model to be less confident there and route more of those cases to a human by default.
Should triage models be built per line of business?
Generally yes. Auto bodily injury, property, and workers' compensation each have different severity drivers and litigation dynamics, so a single generic model tends to underperform line-specific models trained on that line's own claim patterns and outcomes.
How Nanobase AI helps
Nanobase AI, an NVIDIA Inception Program member, builds claims triage and severity models trained on an insurer's own historical outcomes rather than a generic industry model, including the shadow-mode validation step against experienced adjuster judgment before go-live. The team also sets up the ongoing monitoring pipeline so drift is caught before it affects routing decisions. See the broader claims automation approach this fits into.
Ready to discuss your project? Contact Nanobase AI or email hello@bumu.tech.