Claim leakage is the difference between what a claim should have cost and what it actually cost. It is not fraud. It is the cumulative effect of settlement decisions that are individually defensible but collectively inconsistent — claims settled too high because the handler was risk-averse, claims that settled too early before all facts were established, and claims where the reserve was set below what the eventual outcome required. Industry estimates put leakage at 5 to 10 percent of total indemnity spend. On a $500 million indemnity book that is between $25 and $50 million annually, embedded in thousands of individual file decisions, none of which looks obviously wrong at the time it was made.
The source of that leakage is outcome variance. Two handlers working identical claim profiles will reach different settlement amounts. The difference is not incompetence — it reflects the genuine difficulty of estimating the right settlement value for a claim with incomplete information, competing pressures, and a counterparty who is actively negotiating. But at portfolio scale, that variance is the leakage. Reducing it does not require removing handler judgment. It requires anchoring that judgment to what the portfolio’s full experience on comparable claims actually says the right range is.
Where the settlement decision breaks down
The handler making a settlement decision has access to the facts of the file in front of them and their own experience of prior settlements. What they rarely have is systematic visibility into how the full portfolio of comparable claims has settled — the range of outcomes on similar injury profiles, comparable liability scenarios, and equivalent jurisdictions. That portfolio experience exists in the claims system. It is not routinely surfaced at the point of decision.
Without that anchor, handlers calibrate to their own experience base, which is limited and unevenly distributed. A handler who has settled fifty soft tissue injury claims in a specific jurisdiction has a different mental model of the right range than one who has settled five. The difference in their settlement outcomes is predictable and measurable. It is also avoidable.
A model trained on the full historical settlement record — injury profile, liability facts, jurisdiction, claimant representation status, medical spend — produces a settlement range that reflects the portfolio’s collective experience rather than the individual handler’s. That range does not replace the handler’s judgment on case-specific factors. It provides the baseline from which case-specific adjustments are made, consistently, across every file.
The cost runs in both directions
Settlements above the modelled range represent direct leakage — overpayment relative to what comparable claims have historically required. That is the visible component of the problem and the one most commonly cited in leakage analyses.
Settlements below the modelled range create a different cost. Early low settlements on claims that subsequently deteriorate generate reopened files, supplemental payments, and litigation that is more expensive than the original settlement would have been. The handler who settles quickly and cheaply on a bodily injury claim with emerging medical treatment is not saving money. They are deferring a larger cost and adding the administrative expense of a reopened file.
Model-guided ranges reduce both errors simultaneously. The handler is anchored to a range that reflects both the upper bound of appropriate settlement and the lower bound that is likely to hold.
The litigation dimension
Settlement decisions that fall significantly outside the modelled range for the claim profile are the most common driver of unnecessary litigation. A claimant or their attorney who receives an offer they perceive as unreasonably low compared to the settlement value a reasonable assessment would support is more likely to litigate. A model that identifies which claims are being offered below the range where early settlement is achievable enables proactive escalation — more senior handler involvement, earlier authorisation to offer within range — before the claim enters litigation and the cost trajectory changes permanently.
The technology dimension
Claims settlement models require access to the full historical claims record — injury profiles, settlement amounts, litigation outcomes, jurisdiction data, and medical spend — held in the claims management system. For carriers running their claims infrastructure on IBM Z, deploying the settlement guidance model via IBM Machine Learning for z/OS makes the modelled range available to the handler within the claims workflow, drawn from the full data estate on the same platform. The model output is a settlement range and the key factors that drove it, surfaced at the point of decision rather than as a separate analytical exercise.
What success looks like
The primary metrics are settlement leakage rate measured against modelled expectation, variance in settlement outcomes across handlers on comparable claim profiles, litigation rate, and reopened claim rate. The programme baseline should establish the current distribution of settlement outcomes relative to the modelled range before deployment. Narrowing that distribution — reducing the proportion of settlements outside the range in both directions — is the direct measure of whether the model is working. The combined ratio improvement follows from that narrowing.
