Compare · Predict vs Carrier-side AI

The carriers built theirs in 2014. Now we have ours.

Allstate ran claims-side ML on its internal settlement data starting 2014. By 2019, every major US property-casualty insurer was setting reserves with a model. The plaintiff side has been outgunned on data for a decade. Predict closes the asymmetry.

Predict
Carrier-side claims AI
Who sees it
Plaintiff attorneys only — contractually
Internal to the carrier; adjusters and reserve teams
Training data
312K verdicts and reported settlements — plaintiff outcomes
The carrier's own claims history — paid losses, reserve setting
What it optimizes
Estimated fair value to the plaintiff, with confidence bands
Reserve adequacy and combined-ratio targets — the carrier's economics
Methodology disclosure
Public — gradient-boosted regression, stratified jurisdiction folds
"Proprietary model" — opaque to the claimant and to plaintiff counsel
Confidence band
Always shown — ± dollar range, 90% CI, sample size
Internal to the model; rarely surfaced in settlement memos
Recalibration
Quarterly model-refresh changelog, public
Internal cadence; not disclosed to the other side
Side of the table
Plaintiff-only by design — we will never sell to carriers
Carrier-only by design — the asymmetry is the product
What you do with it
Defensible counter-anchor, demand-letter valuation block
Sets the reserve and the opening offer — the floor of the negotiation

312K verdicts and reported settlements across MVA and premises liability. On a held-out test set, predicted values are 90 to 92% accurate against the realized settlement (median absolute percentage error), each shown with a 90% confidence band. How we measure this →

Same kind of instrument. Opposite side of the table.

In plain terms: the carrier shows up to negotiation with a number from a model. Until now you showed up with comps and instinct. Predict gives you a number from a model built on plaintiff outcomes, with the same statistical rigor, so the table is even.

The architecture pattern is the same one the carrier-side claims models use, gradient-boosted regression with stratified jurisdiction folds. The deliberate differences are upstream of the model: the training data is plaintiff outcomes, not carrier paid losses; the methodology is public, not proprietary; the recalibration policy is contractual, not internal; the side of the table is fixed by the brand commitment, not by the buyer.

For 12 years the plaintiff bar has been pricing cases with Verdict Search and institutional memory while the carrier opened with a modeled number. The asymmetry defines a generation of practice. Predict ends it.

Read the full methodology →
$268,000
± $36,000 · 90% CI
IL · COOK · MVA
Carriers have priced cases like this against an internal model since the 2010s. The negotiation used to start with their number setting the anchor. Now both sides bring an instrument to the table.