Claims: assessing everything at once instead of in sequence

The shape of the problem

A carrier takes first notice of loss by phone, app and broker feed, then hands the claim along a chain: intake sets it up, an adjuster assesses severity, a special investigations referral happens if something looks odd, and a subrogation review happens at the end if anyone remembers.

Straightforward claims — a windscreen, a simple property leak, a low-value collision with clear liability — take as long as complex ones, because they travel the same chain.

Subrogation is the quiet loss. Recovery opportunities are most visible at first notice, when the claimant is describing what happened and who else was involved. By the time a review runs weeks later, the detail has gone cold.

Why the obvious fix does not work

Carriers optimise each handoff and never question the chain. Every stage is reasonably efficient. The sequence is the problem.

Severity, fraud indicators and recovery potential are not sequential questions. They are three readings of the same evidence, and they are all most accurate at the moment the evidence is freshest.

The design

One pass at first notice

Every claim opens all three assessments at once against the same assembled evidence: the policy record, claimant history, the loss description, photos and documents, telematics where the policy carries it, bureau and rating data, and the fraud consortium lists.

Deterministic models handle coverage verification, policy limits and constraint checks — is this loss in scope, is the policy in force, does the described damage fit the vehicle. Retrieval assembles comparable prior claims and their eventual development. The reasoning lane weighs it and writes the assessment a handler reads.

An eligibility gate, not a confidence threshold

Straight-through settlement applies only to claims passing an explicit eligibility test written by the claims and compliance functions: value band, coverage clarity, no injury, no third-party dispute, no fraud indicator above threshold, claimant not flagged vulnerable.

This is deliberately not "settle it if the model is confident". Confidence is one input to eligibility, not a substitute for it — a model can be extremely confident about a claim that policy says a human must look at.

Subrogation identified where it is visible

Recovery potential is assessed at notice, from the loss narrative and the parties involved, and written onto the claim as a structured flag with the evidence attached. It stops being a review that happens later and becomes a field that exists from the start.

What to measure afterwards

• Median notice-to-reserve, and the same-day settlement rate among eligible claims.
• Reserve accuracy against 90-day development, which is the only honest way to measure it.
• Recovery identified at first notice versus at later review.
• Leakage, computed by the actuarial function and treated with appropriate scepticism.

The governance position

Every automated settlement carries the eligibility test result, the evidence set, the model versions that ran and the computed confidence per lane.

The control to insist on: a monthly sample of automated settlements re-worked by a human handler blind to the automated outcome, with disagreements going to a governance group. Never reduce that sample rate because agreement is high — that is exactly when the measurement is cheap.