Subrogation Recovery
“Which claims have recovery potential?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
Which claims have recovery potential?
US P&C insurers recover $40-60B annually through subrogation, but industry analysis suggests $15-25B more is left on the table (Verisk). The challenge: subrogation opportunities must be identified early, before evidence is lost and statutes of limitations expire. Most insurers rely on adjusters manually flagging subrogation potential, but 40% of recoverable claims are never flagged because the third-party liability is not obvious at FNOL. A rear-end collision where the other driver was clearly at fault is easy to flag; a water-damage claim where the manufacturer's defective pipe caused the loss requires connecting claims data with product recall databases and building-code records.
Quick answer
AI identifies subrogation recovery opportunities by connecting claims data with product recall databases, building records, third-party liability signals, and historical recovery outcomes. Traditional approaches rely on adjusters manually flagging recovery potential, missing 40% of recoverable claims because the third-party liability is not obvious at first glance. Graph-based models flag opportunities within 24 hours of FNOL, increasing recovery rates by 20-35%.
Approaches compared
4 ways to solve this problem
1. Manual Adjuster Flagging
Adjusters review claims and flag subrogation potential based on experience and obvious third-party liability (e.g., the other driver was at fault). The default process at most insurers.
Best for
Clear-cut liability situations like rear-end auto collisions where fault is unambiguous.
Watch out for
Adjusters miss 40% of recoverable claims because the liability is not obvious at FNOL. Product defects, manufacturer recalls, and construction defects require cross-referencing external databases that adjusters do not have time to check on every claim.
2. Rules-Based Screening
Automated rules that flag claims matching known recovery patterns: auto claims with police reports indicating third-party fault, property claims with specific cause codes (pipe burst, product failure), or claims above a dollar threshold.
Best for
Catching the easy wins: auto subro with clear liability, product liability with active recalls already mapped.
Watch out for
Rules need constant updating as new recall notices and liability patterns emerge. They catch known patterns but miss novel recovery opportunities like building-code violations or contractor negligence.
3. Text Mining on Claim Notes
NLP models that scan adjuster notes and FNOL descriptions for keywords and phrases indicating third-party involvement (e.g., 'other driver', 'defective', 'manufacturer'). Adds coverage beyond manual flagging.
Best for
Surfacing hidden subro signals buried in unstructured claim notes that adjusters may not have flagged.
Watch out for
Limited to what is mentioned in the notes. If the adjuster did not document the manufacturer or building year, the model has nothing to work with. Cannot cross-reference external databases like product recalls.
4. Relational Deep Learning (Kumo's Approach)
Connects claims to product databases, recall notices, building records, historical recovery outcomes, and third-party information in a relational graph. Identifies recovery opportunities by matching claim characteristics with known liability patterns across the full data network.
Best for
Discovering non-obvious subrogation opportunities: product defect recalls, building-code violations, contractor negligence, and manufacturer warranty claims that adjusters would never flag manually.
Watch out for
Recovery probability predictions are only as good as the external data they connect to. Maintaining up-to-date product recall and building-record databases is an ongoing operational requirement.
Key metric: 40% of recoverable subrogation claims are never flagged by adjusters (Verisk). Graph-based detection at FNOL recovers an additional $50-100M annually for a top-20 P&C insurer.
Why relational data changes the answer
Flat subrogation models see each claim as an isolated row: peril type, amount paid, third-party information (if any). They can flag an auto collision where the police report names a third-party at fault. But they cannot see that a water-damage claim involves a pipe fitting manufactured by a company with an active recall, installed during a building period with known plumbing defects, in a region where similar claims have achieved 65% recovery rates from the same manufacturer. These connections span claim records, product recall databases, building permits, and historical recovery outcomes. A flat model would need a human to manually check each claim against these external databases.
Relational learning automates this cross-referencing at scale. The model walks from the claim to the incident details (cause code, product involved), to the manufacturer database (recall status, solvency), to historical recoveries (success rate, average amount, timeline), and to building records (construction year, known defect windows). It learns that water-damage claims involving PEX-200 fittings installed between 2006 and 2012, where the manufacturer is solvent and paying claims, have a 65% recovery probability with an average recovery of $7,800. This pattern would take an adjuster hours to research manually for a single claim. The relational model applies it to every incoming claim automatically within 24 hours of FNOL, before evidence deteriorates and statutes of limitations become a concern.
Identifying subrogation from a flat claims table is like finding product recall candidates by reading individual customer complaints one at a time. Each complaint looks isolated. But connect the complaints to the product database, cross-reference with the manufacturer's batch records, and map the installation dates, and a clear recall pattern emerges. Graph-based subrogation detection is the insurance equivalent of a Consumer Product Safety Commission investigation, except it runs on every single claim automatically.
How KumoRFM solves this
Relational intelligence built for insurance data
Kumo connects claims, police reports, policy details, third-party information, product databases, building records, and historical recovery outcomes into a relational graph. The model identifies that Claim CLM-9220 (pipe burst) involves a pipe manufactured by a company with an active recall notice, the building was constructed during a period of known plumbing defects, and similar claims in the region have achieved 65% recovery rates. These cross-table signals surface subrogation opportunities that adjusters would otherwise miss, flagging them within 24 hours of FNOL.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
CLAIMS
| claim_id | peril | amount_paid | third_party_info | status |
|---|---|---|---|---|
| CLM-9220 | Water | $11,200 | Unknown at FNOL | Open |
| CLM-9225 | Auto Collision | $18,400 | Other driver: at fault | Open |
| CLM-9230 | Property | $8,600 | Unknown at FNOL | Open |
INCIDENT_DETAILS
| claim_id | cause_code | product_involved | manufacturer | building_year |
|---|---|---|---|---|
| CLM-9220 | Pipe Burst | PEX-200 Fitting | AquaFlow Inc | 2008 |
| CLM-9225 | Rear-End | N/A | N/A | N/A |
| CLM-9230 | Roof Leak | Asphalt Shingles | RoofTech | 2012 |
PRODUCT_RECALLS
| manufacturer | product | recall_date | defect_type | claims_filed |
|---|---|---|---|---|
| AquaFlow Inc | PEX-200 Fitting | 2024-06-15 | Premature Failure | 2,400 |
| RoofTech | 30-Year Shingles (2010-2013) | 2023-11-01 | Premature Wear | 1,800 |
RECOVERY_HISTORY
| peril | cause_code | manufacturer | recovery_rate | avg_recovery | avg_days_to_collect |
|---|---|---|---|---|---|
| Water | Pipe Burst | AquaFlow Inc | 65% | $7,800 | 120 |
| Auto Collision | Rear-End | N/A | 82% | $15,200 | 90 |
| Property | Roof Leak | RoofTech | 45% | $3,900 | 180 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT BOOL(CLAIMS.SUBROGATION_RECOVERY > 0, 0, 180, days) FOR EACH CLAIMS.CLAIM_ID WHERE CLAIMS.STATUS = 'open'
Prediction output
Every entity gets a score, updated continuously
| CLAIM_ID | PERIL | RECOVERY_PROB | EST_RECOVERY | PRIORITY | RECOVERY_PATH |
|---|---|---|---|---|---|
| CLM-9225 | Auto Collision | 0.88 | $15,200 | High | Third-Party Liability |
| CLM-9220 | Water | 0.72 | $7,300 | High | Product Defect Recall |
| CLM-9230 | Property | 0.48 | $3,900 | Medium | Manufacturer Warranty |
Understand why
Every prediction includes feature attributions — no black boxes
Claim CLM-9220 (Water Damage, pipe burst)
Predicted: 72% recovery probability, est. $7,300
Top contributing features
Active product recall match
AquaFlow PEX-200
30% attribution
Building year in defect window
2008 (2006-2012)
25% attribution
Historical recovery rate for cause
65%
20% attribution
Manufacturer solvency status
Solvent, paying claims
14% attribution
Statute of limitations remaining
18 months
11% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about subrogation recovery
What is insurance subrogation and how does AI improve it?
Subrogation is the process of recovering claim payments from the responsible third party (another driver, a product manufacturer, a negligent contractor). AI improves subrogation by automatically identifying recovery opportunities at FNOL, cross-referencing claims with product recall databases, building records, and historical recovery patterns. This catches 20-35% more recoverable claims than manual adjuster flagging.
How much money do insurers leave on the table in subrogation?
US P&C insurers recover $40-60B annually through subrogation, but industry analysis by Verisk suggests $15-25B more is recoverable. The gap exists because 40% of recoverable claims are never flagged: the third-party liability involves product defects, building-code violations, or manufacturer warranties that are not obvious from the claim alone.
How quickly can AI identify subrogation opportunities?
Graph-based models flag subrogation opportunities within 24 hours of FNOL by automatically cross-referencing claim details with product recall databases, building records, and third-party liability patterns. This is critical because evidence degrades over time and statutes of limitations vary by state, making early identification essential for successful recovery.
What types of claims have the highest subrogation recovery potential?
Auto collision claims with clear third-party fault have the highest recovery rates (80-85%). Product liability claims involving active manufacturer recalls recover at 55-65%. Property claims tied to construction defects or building-code violations recover at 40-50% but are the most commonly missed because the liability requires cross-referencing external databases.
Bottom line: Identify 20-35% more subrogation opportunities at FNOL and accelerate recovery timelines by 30%, recovering an additional $50-100M annually for a top-20 P&C insurer.
Related use cases
Explore more insurance use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
