An AI Mission for Insurance: Subrogation Review

Executive Summary

Subrogation is money your carrier has already spent and has a legal right to recover — and most of it never comes back. After you pay a claim caused by a third party, you can pursue that party or their insurer to recoup the outlay. The trouble is that identifying recovery potential requires reading the full claim file, applying liability and comparative-negligence rules, checking statutes of limitation by state, and doing it on every closed file before the window shuts. Human referral catches a fraction. In this article I describe how we build subrogation review as an AI Mission on the StudioX Enterprise AI Platform — a stateful, observable Worker that reads each file, scores recovery potential, drafts the demand package, and routes recommendations into a Decision Queue for a recovery specialist to approve. I'm Trevor Solis, Lead AI Engineer on Missions, and this is one of the highest-ROI Missions I've shipped, because it turns unrecovered dollars directly into recovered ones.

The Problem

Subrogation potential is buried in unstructured data. Whether a paid claim is recoverable depends on facts scattered across the loss description, the police report, the adjuster's notes, photos, repair estimates, and the coverage that responded. A rear-end collision where your insured was stopped is textbook recoverable; a comparative-negligence intersection case might net 60 cents on the dollar; a single-vehicle loss is a dead end. Nobody has time to read every file to sort these out, so referral leans on adjuster instinct at close-out — and adjusters are measured on cycle time, not recovery. The result is systematic under-referral, with recoverable dollars quietly written off and statutes of limitation lapsing unnoticed.

The Traditional Approach

Most carriers run subrogation as a downstream, mostly-manual function. The claims system (Guidewire ClaimCenter, Duck Creek, or Sapiens) has a "subrogation" flag an adjuster can set, and there may be a business rule that auto-flags certain loss types — rear-end auto, water damage from an upstairs unit. Flagged files drop into a recovery unit's worklist, where a specialist reads the file, decides whether to pursue, identifies the adverse carrier, drafts a demand letter, and tracks the statute. Some carriers outsource the whole function to a third-party recovery vendor on contingency, ceding 20–30% of every recovered dollar in exchange for capacity they can't staff internally.

Why It Fails

The auto-flag rules are coarse — they catch the obvious rear-end but miss the product-defect fire, the contractor-caused water loss, and the negligent-security liability case, because those turn on facts a rule can't read. Manual referral depends on the one person under the most cycle-time pressure to volunteer extra work. Files that are referred sit in a queue long enough that the statute of limitations — as short as one year in some states, two to six in others — burns off before the demand goes out. And the outsourced model, while it recovers something, is expensive and gives you no visibility into why a file was pursued or dropped. In every version, the core failure is the same: no one reads every file, and the reading is where the money is.

How StudioX Solves It

On StudioX we make an Autonomous AI Worker read every closed-with-payment file as an AI Mission. The Worker pulls the full claim record through a Model Context Protocol (MCP) connector to the claims system, ingests the unstructured content — adjuster notes, police report, estimates, photos — and reasons over liability using the carrier's own Enterprise Knowledge: subrogation guidelines, comparative-negligence rules by state, and the statute-of-limitation matrix. It produces a recovery-potential score, the adverse party and their carrier, an estimated net-of-comparative recovery figure, and the days remaining on the statute.

Then it drafts the recovery package — the demand letter, the itemized damages, the liability rationale — and proposes referral. That proposal lands in the Decision Queue, where a recovery specialist reviews the reasoning, adjusts the liability assessment, and approves or declines. Every inference the Worker made streams to the Explain rail as Observations, so the specialist sees exactly why the file scored the way it did before approving a dollar of pursuit.

Benefits

The headline benefit is recovery yield: when a Worker reads every paid file instead of the fraction adjusters remember to flag, referral volume rises and previously-missed categories — product defect, contractor negligence, premises liability — start converting. The second benefit is timeliness: because the Mission runs on close-out, statute-of-limitation clocks are surfaced with days-remaining while there's still runway to act, ending silent lapses. The third is margin: work that would have gone to a contingency vendor at 20–30% now runs in-house at platform cost, and where you still use a vendor, you hand them pre-scored, pre-packaged files. And because the reasoning is observable, your recovery leadership can audit decline decisions and defend pursued files to reinsurers.

Example Workflow

A concrete auto-subrogation Mission:

Trigger — A collision claim closes with a $14,200 payment. The close-out event triggers the subrogation Mission.
Pull the file — Via MCP, the Worker retrieves the full record from ClaimCenter, including adjuster notes, the police report PDF, and repair estimates.
Establish liability — Reading the police report, it determines the insured was stopped at a signal and struck from behind; liability on the adverse driver is clear.
Apply state rules — Against Enterprise Knowledge, it applies the loss-state comparative-negligence rule (pure comparative, 0% insured fault here) and finds the statute of limitation is two years with 21 months remaining.
Score and quantify — It scores recovery potential high and estimates net recovery at the full $14,200 plus deductible reimbursement for the insured.
Identify the adverse carrier — It extracts the at-fault party's insurer and policy number from the report.
Draft the package — It generates a demand letter, itemized damages, and a liability rationale, and proposes referral to the recovery unit.
Human approval — The proposal enters the Decision Queue. The specialist reviews the Explain-rail Observations, confirms liability, and approves. The Mission logs the referral and returns its verdict: recoverable, high, demand ready.

Related StudioX Capabilities

Subrogation pairs naturally with other Missions: an FNOL triage Mission can flag subrogation potential at intake, and a litigation-management Mission can pick up files where the demand is contested. Portals give recovery specialists and adverse carriers a branded surface for demand correspondence and status. Enterprise Integrations via MCP tie the claims system, document management, and your recovery-tracking tool into one Mission. And private or VPC Enterprise Deployment with LLM Independence keeps claimant and third-party data inside your boundary — essential when files contain medical records and police reports.

Frequently Asked Questions

Does the Worker send demand letters on its own? No. It drafts the demand package and proposes referral; a recovery specialist approves in the Decision Queue before anything leaves the building. State-changing actions are always human-approved.

How does it handle comparative-negligence differences between states? The comparative-negligence rules and statute-of-limitation matrix live in your Enterprise Knowledge. The Mission applies the correct rule for the loss state on every file and shows its work on the Explain rail.

Will it replace our contingency recovery vendor? It can move volume in-house, or make your vendor more efficient by handing them pre-scored, pre-packaged files. Either way you gain visibility into why each file was pursued or declined.

Can we trust the liability assessment? You review it. Every liability inference is an Observation on the Explain rail, so the specialist validates the reasoning against the source documents before approving pursuit — see the Enterprise AI Platform for how observability works.

Call to Action

If unrecovered subrogation dollars and lapsing statutes are costing you real money, a subrogation-review Mission recovers both. Contact my team for a proof of value on the StudioX Enterprise AI Platform against a sample of your own closed files — the recovery gap usually pays for the project several times over.