Billions Left Behind: How AI Rewrites the Economics of Subrogation

May 13, 2026

insuranceindustry.ai article hero image

By James W. Moore

Key Takeaways

  • An estimated $15-20 billion in recoverable subrogation dollars goes uncollected annually — not from incompetence, but from economics that made most cases not worth pursuing.
  • AI shifts subrogation from selective, episodic recovery to continuous, portfolio-level recovery management — changing which cases are worth pursuing, not just how fast they are processed.
  • Top performers recover 80% of eligible subrogation opportunities; laggards recover 30-40%. AI will widen that gap.
  • When recovery becomes predictable at first notice of loss, subrogation stops being a post-mortem and starts being a reserving input — with direct implications for loss reserves and combined ratio.

Somewhere in your claims system right now, there is a file stamped “Closed.” The settlement was paid, the adjuster moved on, and the case was archived — buried in a PDF note on page 47 — without anyone noticing that a third party was partially at fault. The recovery opportunity, perhaps $8,000 or $25,000, will never be pursued. Not because your team missed it out of carelessness. Because under your current cost structure, finding it wasn’t worth the effort.

Multiply that file by tens of thousands, and you have the subrogation problem in American insurance.

In 2024 alone, Arbitration Forums members filed nearly 1.2 million arbitration disputes and 2.4 million subrogation demands collectively valued at more than $26.3 billion filed through AF’s platform. Travelers’ dedicated subrogation program, QuantumSubro, has recovered more than $6.9 billion across business lines over the past ten years ending 2024. And yet, industry estimates consistently place missed subrogation opportunities at $15-20 billion annually in recoverable dollars that never make it back to carriers’ balance sheets. Subrogation is constrained by economics — and those economics are changing.

Carriers Aren’t Failing at Subrogation. They’re Behaving Rationally.

Before diagnosing the problem, it’s worth defending the people who live with it. Most carriers are not bad at subrogation. They are making economically rational decisions inside a cost structure that makes most recovery opportunities not worth pursuing.

The underlying equation is simple: carriers pursue subrogation when the expected recovery, adjusted for probability of success, exceeds the cost to identify and pursue the claim. Most cases fall just below that threshold — not because they lack merit, but because they are small, uncertain, or expensive to validate. A $2,500 recovery that requires four hours of adjuster time, two rounds of document gathering, and a demand letter review simply does not pencil out under a manual process.

The result is striking. A 2025 industry survey found that 91% of insurers pursue fewer than 30% of the subrogation cases they identify — and they know it. That is not negligence. It is math.

What makes this a strategic problem rather than an operational nuisance is the performance spread between carriers who have found ways to beat the economics and those who haven’t. Top-performing carriers recover approximately 80% of eligible subrogation opportunities. Laggards recover 30-40%. That 40-50 point gap represents a significant and compounding competitive disadvantage — one that shows up directly in combined ratios, loss performance, and ultimately in pricing power.

Your current economics are hiding recoverable money — and you have accepted that as normal.

Why the Problem Has Persisted

Understanding why subrogation has remained structurally under-optimized for decades requires looking honestly at the forces that created the status quo.

The data problem. Approximately 80% of the information relevant to a subrogation claim is unstructured — buried in adjuster notes, police reports, witness statements, photographs, and repair estimates. Humans are not equipped to systematically mine that volume of text across thousands of open files. And by the time someone tries, the file has already aged. The average subrogation recovery cycle runs approximately six months from identification to payment. Subrogation success rates decay with time; every day between first notice of loss and identification narrows the recovery window and reduces win rates.

The talent cliff. Subrogation is artisanal knowledge. Identifying a product liability angle in a fire claim, or spotting a contractor negligence theory in a commercial property loss, requires the kind of judgment that accumulates over years of exposure. According to CCC Intelligent Solutions’ 2025 research, more than 50% of first-party adjusters at some carriers have less than two years of experience. Subrogation referrals have declined by up to 16% as adjusters with heavy caseloads and limited experience simply do not recognize the opportunities in front of them. The senior professionals who carry that institutional knowledge are retiring, and the traditional apprenticeship model — shadowing a tenured adjuster, learning by osmosis — cannot scale fast enough to replace them. For the first time, subrogation expertise may be transferable at scale instead of apprentice-based. But that transfer requires infrastructure that most carriers don’t yet have.

The cost floor. Manual pursuit of small-dollar claims costs more than it recovers. Outsourcing to third-party subrogation vendors does not solve this — contingency fees typically run 15-30% of recovery, and those vendors face the same talent shortages internally. The cost floor exists whether the work is done in-house or farmed out, and limited carrier visibility into how outsourced demands are handled compounds the risk further.

The process gap. The downstream consequences of manual, inconsistent subrogation processes are measurable. According to CCC’s research, one in four subrogation demands contains a data error — missing documentation, mismatched line items, or unsupported charges. Incomplete demands get denied, disputes go to arbitration, and cycle times extend. Each error is both a direct cost and a symptom of a process that was never designed to operate at scale.

AI Changes the Underlying Economics

The case for AI in subrogation is not that it makes adjusters faster at doing what they already do. It is that it changes the cost structure of the entire function — and in doing so, changes which cases are worth pursuing.

Detection cost drops toward zero. Large language models and natural language processing can scan every claim at first notice of loss for liability signals — adjuster notes, police reports, photographs, telematics data — in seconds, without depending on adjuster memory or available bandwidth. Cases that would have required hours of manual review to identify now surface automatically, before the file ages and before the recovery window narrows.

Probability estimation improves early. AI scores recovery likelihood before the file goes cold, moving confidence from late-cycle guesswork to early-stage triage. This matters because it is not just detection cost that constrains subrogation pursuit — it is also confidence. Adjusters historically erred toward inaction when liability was uncertain because the cost of being wrong was high and the tools to assess it quickly were limited. AI changes both variables simultaneously: cost and confidence move in the right direction at the same time.

The small-dollar flip. When the marginal cost of evaluating a subrogation opportunity approaches zero, the economic threshold moves dramatically. A $2,500 recovery that could never justify manual pursuit becomes viable. Cases that were structurally ignored become routinely pursued. AI does not just find more recoveries — it expands the universe of recoveries that are economically rational to chase. That is a fundamentally different operating model, not an incremental improvement.

The continuous learning loop. Every AI-flagged case that a senior adjuster reviews, adjusts, or overrides becomes a training signal. The system improves with every claim processed. Subrogation expertise does not have to retire when experienced professionals do — it can be systematized, retained, and compounded over time. This is not theoretical: it mirrors the feedback infrastructure already operating in fraud detection and underwriting risk scoring.

That is the theory. Here is what is already happening.

Early results from carriers deploying AI in subrogation are concrete. Shift Technology reports that carriers using AI-driven subrogation detection have documented up to a 30% increase in identified opportunities, 20% faster cycle times, and combined ratio improvements of up to 4% — which, for many carriers, represents the difference between operating in the red and the black. FM Global is actively piloting AI tools for subrogation identification, with its SVP of Claims noting the goal of eliminating unnecessary consultant costs and improving recovery targeting. Virginia Farm Bureau piloted CCC’s inbound subrogation solution in early 2025 and reported early improvements in subrogation variance, cycle times, and arbitration outcomes — while simultaneously building estimating knowledge in adjusters who had previously lacked it.

That 25% error rate is not inevitable. AI-generated demand packages start with structured, validated data, drawing from claims records, repair estimates, and documentation databases automatically. Clean demands do not get denied for missing VINs or illegible estimates — they get paid.

The ecosystem multiplier. The economics improve further when AI adoption extends beyond the carrier. Law firms specializing in subrogation are increasingly deploying AI for demand preparation, negotiation support, and arbitration analysis. The dynamic this creates is not zero-sum. When a carrier uses AI to identify and triage a higher volume of viable cases — including small-dollar claims that would previously have been abandoned — the volume of work flowing to subrogation counsel increases. The law firm, operating with AI-assisted demand preparation at lower cost per file, can handle that increased volume profitably. Both parties capture efficiency gains. The viable recovery threshold drops further than either could achieve independently, and the total recoveries flowing back to the carrier expand accordingly. This is early, but claims professionals at forward-looking carriers are already asking the right questions about it.

The Strategic Reframe

Subrogation is not a back-office claims function. It is a financial performance lever that most carriers have left in manual mode while the economics of running it effectively quietly shifted around them.

Historically, carriers optimized subrogation around averages — directing attention to the obvious, high-confidence, high-value cases and accepting everything else as unrecoverable leakage. AI makes something different possible: high-volume probabilistic recovery management across the entire claims portfolio. Subrogation becomes actuarial rather than artisanal. That shift is operationally similar to what has already happened in fraud analytics, catastrophe modeling, and portfolio underwriting — disciplines that carrier executives already fund and respect precisely because they turn individual uncertainty into portfolio-level predictability.

The reserve implication deserves explicit attention. When recovery becomes predictable at first notice of loss rather than uncertain at file closure, subrogation stops being a post-mortem and starts being a reserving input. Actuaries can begin crediting expected recoveries earlier in the claims lifecycle — a direct effect on loss reserve adequacy and reported combined ratio. For CFOs and chief actuaries, that is not an operational story. It is a financial reporting story.

Upendra Belhe and colleagues at EXL, writing in Carrier Management, frame the shift precisely: subrogation is no longer a claims cleanup task but a strategic avenue for value creation that drives underwriting performance and combined ratio management. That framing is right, and it implies organizational accountability that most carriers have not yet assigned.

The competitive divide will not narrow on its own. Carriers deploying AI in subrogation now are extending their recovery performance advantage year over year. The 80% vs. 30-40% recovery spread documented today will widen as AI adopters compound their gains and laggards continue operating below the economic threshold. In subrogation, your biggest competitor is not the other carrier. It is your own cost structure.

One caveat: over-automation of demand generation carries real regulatory and relationship risk. Human judgment remains essential for legal strategy, negotiation, and ethics; AI handles triage, not decisions.

What to Do With This

The executives who will move first on AI-enabled subrogation are not waiting for a perfect business case. They are asking three questions now:

  • What is our current recovery rate as a percentage of identified subrogation opportunities — and how does that compare to top-quartile performance?
  • At what point in our claims lifecycle does subrogation identification happen — at first notice of loss, or after settlement?
  • What would a 5-10 point improvement in our recovery rate mean to our combined ratio?

The math is carrier-specific, but the direction is not. A carrier writing 500,000 claims annually that recovers an additional $400 on 5% of newly identified files adds $10 million in pure margin. A carrier lifting its overall recovery rate from 35% to 45% on a $2 billion paid-loss book with 20% subrogation potential adds $40 million. Those are illustrative numbers. Run your own.

If AI can evaluate every claim for subrogation potential at near-zero cost, at what point does choosing not to do so become an active financial decision rather than an operational limitation?

Sources

Subscribe to AI Insights Newsletter

AI Disclaimer: This content was created with assistance from artificial intelligence technology. While content is based on factual information from the source material, readers should verify all details directly with the respective sources before making business decisions.

Related Posts

468