Subrogation Scoring Model: Expansion Under a Fairness Cloud
Northline Mutual writes 410,000 personal auto policies across six states. In March it deployed a vendor's machine-learning model that scores every new auto injury claim for subrogation potential on a 1-100 scale. Claims scoring above 70 route directly to a three-person recovery unit and skip the normal desk-adjuster review. After 90 days: the recovery unit's referral volume is 2.4 times forecast; average days-to-contact for low-score claims has risen from 1.8 to 4.6; and an internal audit found that scores correlate strongly with claimant ZIP code and with whether the claimant called the Spanish-language intake line. The vendor contract calls the model proprietary and supplies only a global feature-importance chart. State regulators have begun asking carriers to document how automated decision tools are validated. The chief claims officer wants to expand the model to property claims next quarter.
1. Identify the distinct risks this deployment creates and specify the evidence you would require before accepting that the model is performing as intended.
Read the worked answer
Three risk categories are being conflated and need separate evidence.
Model performance. A 2.4x referral volume means the score-70 threshold was never calibrated on production data. Require the score distribution against the validation sample, precision and recall measured against closed claims with known recovery outcomes, calibration by month, and drift statistics. Days-to-contact rising to 4.6 on low-score claims suggests the rerouting starved the desk units, not that the model is accurate.
Fairness and legal exposure. A score correlated with ZIP code and language line is a proxy problem. ZIP code carries race and national-origin information; language preference may correlate with claim complexity and attorney representation. Test outcomes by group: referral rate, recovery per dollar of subrogable loss, time to first contact, and denial rate. If similarly situated claimants get materially different recovery outcomes, the carrier faces a disparate-impact question under state unfair-discrimination law and NAIC AI guidance regardless of whether ZIP code is an approved rating factor, because claims handling is not rating.
Consumer harm and explainability. Skipping desk review for high scores has an unexplained opportunity cost for everyone else, and there is no appeal path for a claimant harmed by a low score. Require override logging and a documented threshold rationale.
Do not accept a feature-importance chart as validation; insist on outcome-based, segmented testing.
2. Recommend the governance controls and a specific go/no-go decision on the property claims expansion.
Read the worked answer
Treat this as a tier-1 model under the carrier's model risk framework.
Before any expansion: (1) Freeze the threshold and run the model in shadow mode alongside full desk review for 60-90 days so no claimant is routed solely by score. (2) Commission independent validation by someone who does not report to the claims officer and who can obtain vendor documentation under the existing contract or an amendment. (3) Institute documented human review, override codes, and an adverse-outcome appeal route. (4) Add monitoring dashboards with pre-agreed stop triggers: referral rate outside a band, group-level outcome gaps over a set threshold, or contact-time slippage.
Contractual work: audit rights, model-change notification, validation access, and service-level commitments on score latency and refresh.
Recommendation: no-go on property now. The property model would inherit the same vendor opacity and the same unproven threshold logic, against a claims population with different loss drivers, while the auto deployment already shows an unexplained fairness signal and an operations failure. Approve expansion only after the auto model passes independent validation, group-outcome testing shows no material disparity or the disparity is explained and mitigated, and the claims organization demonstrates it can sustain contact-time standards. Document this as a risk-acceptance decision with a named owner and a revisit date, not as an informal deferral.
3. A state regulator asks how the carrier ensures the model does not produce unfair outcomes, and the vendor refuses to disclose feature-level logic. What do you tell the regulator, and what do you do about the vendor?
Read the worked answer
The carrier owns the answer; it cannot outsource responsibility to a vendor. A defensible response to the regulator has four parts.
Governance: describe the model's purpose, the business justification for the score threshold, who approved it, and how it is monitored.
Validation evidence: present outcome-based testing by segment (referral rate, recovery dollars per dollar of subrogable loss, time to contact) rather than internal mechanics. Outcome testing is exactly how you demonstrate non-discrimination when you cannot see the weights.
Compensating controls: human decision authority, override logs, appeal path, and the fact that no claimant is denied a benefit by score alone.
Honest gap disclosure: state that the carrier does not have feature-level visibility and describe the contractual and independent-validation steps underway. Do not claim explainability you do not have.
On the vendor: a global feature-importance chart is not an acceptable basis for a regulated decision tool. Seek an amendment giving audit rights, access to validation documentation, and change notice. If refused, the options are third-party validation under NDA, replacement of the model component with an interpretable rules-plus-logistic approach for the subrogation score, or termination. Give the vendor a defined cure period and document the analysis. Regulators increasingly expect explainability to the consumer and to the supervisor; a black-box defense is not a strategy, it is a deferral. In the interim, hold the threshold and keep humans in the loop for every claim.