The Recovery Insurers Systematically Miss
Subrogation is the insurer's right, after paying a claim, to step into the policyholder's shoes and recover from the third party responsible for the loss. When an insurer pays a motor claim caused by another driver, a property claim caused by a contractor's negligence, or a cargo claim caused by a carrier's fault, it can pursue the at-fault party to recover what it paid. Done well, subrogation returns a meaningful share of paid claims to the insurer and ultimately to the pricing pool that benefits policyholders.
In practice, Indian insurers recover far less than they could. The reason is that subrogation is labour-intensive and depends on three things being done consistently across a high volume of claims: identifying which paid claims have a recovery opportunity, gathering the evidence to support the recovery, and pursuing the at-fault party or its insurer to settlement. Each step is manual in most operations, and at the volumes Indian insurers process, recovery opportunities are routinely missed because no one identified them, evidence was not preserved, or the pursuit was not worth a claims handler's time relative to its expected value.
The missed recovery is a direct loss to the insurer's results and an indirect cost to policyholders, because uncollected recoveries raise the net cost of claims that pricing must cover. Industry experience suggests that a material share of recoverable amounts is never pursued, and that even pursued recoveries settle for less than they should because the evidence and the prioritisation were weak. Subrogation is therefore one of the larger pools of recoverable value sitting unaddressed in Indian claims operations.
Where AI Changes the Subrogation Economics
AI changes subrogation economics primarily by solving the identification and prioritisation problems that cause recoveries to be missed, and secondarily by accelerating the evidence and pursuit steps.
The identification problem is that recovery opportunities are buried in claims data and free-text claim notes that no one has time to read systematically. A model trained on historical claims can read the structured data and the unstructured notes, surveyor reports, and FNOL descriptions to flag claims that carry a recovery opportunity, the third-party-at-fault motor claim, the contractor-caused property loss, the carrier-caused cargo damage, that a manual triage misses. The model surfaces these opportunities at the point of claim handling rather than leaving them to be found by chance or not at all.
The prioritisation problem is that handlers cannot pursue every recovery and must focus on those worth pursuing, but the manual assessment of recovery value is rough. A model can estimate the recoverable amount, the probability of recovery, and the expected net value after pursuit cost, ranking opportunities so that handler effort goes to the recoveries with the highest expected return. This prioritisation is where much of the value lies, because it directs limited recovery resource to where it pays and clears the low-value cases that consume effort for little return.
The evidence and pursuit steps benefit from automation that assembles the recovery file, drafts the recovery correspondence, and tracks the pursuit through to settlement. Generative tools can compile the evidence the recovery requires from the claim file and produce the demand and follow-up correspondence, while workflow automation tracks the recovery through its stages and escalates stalled pursuits. The combination shifts subrogation from a manual, opportunistic activity to a systematic process that identifies, prioritises, and pursues recoveries across the whole book.
Line-by-Line Application
Subrogation recovery and its automation differ by line, and the application of AI reflects those differences.
Motor is the largest subrogation pool by volume because third-party-at-fault accidents are common and the recovery is against the at-fault driver's insurer. The identification challenge is determining fault from the claim record and matching the recovery to the at-fault party and its insurer, which AI supports by reading the claim narrative, the surveyor assessment, and the accident circumstances to flag recoverable claims and estimate the recoverable share given the fault allocation. The high volume and relative standardisation of motor recoveries make them well suited to automated identification and prioritisation.
Property recoveries arise where a third party caused the insured loss, a contractor, a supplier, a neighbouring occupier, or a product that failed, and the recovery is against that party or its liability insurer. The identification challenge is recognising the third-party cause in the loss circumstances, which is often buried in the surveyor and investigation reports, and AI that reads those reports can surface property recovery opportunities that a handler focused on settling the policyholder's claim overlooks. Property recoveries are individually larger and more complex than motor, so the prioritisation and evidence-assembly support matters more per case.
Commercial and cargo recoveries involve carriers, contractors, and other commercial counterparties, and the recovery often turns on contractual liability and the terms of carriage or contract. The identification challenge includes recognising the contractual recovery route, and the evidence assembly is more document-intensive, drawing on contracts, bills of lading, and survey reports. AI that reads these documents to identify the liable party and the contractual basis for recovery supports a class where recoveries are valuable but the manual effort to pursue them is high.
Across all lines, the common pattern is that AI surfaces opportunities and prioritises them, while the legal and negotiation work of actually recovering remains a human-led process supported by automated evidence assembly and correspondence. The model finds and ranks; the handler and the recovery specialist pursue.
The India-Specific Mechanics That Shape Recovery
Subrogation in India runs through a set of market structures and forums that any AI recovery system has to understand, because they decide whether a flagged opportunity is genuinely recoverable and through which route.
The single most important structure in motor is the Knock for Knock agreement administered by the General Insurance Council. Under it, signatory insurers agree to bear their own policyholder's own-damage loss and not to pursue subrogation against each other where both vehicles are insured and the losses fall within agreed parameters, which avoids two insurers litigating a recovery that nets out across the market. The practical consequence for an AI system is critical: a model that flags every third-party-at-fault motor claim as a recovery opportunity will generate large volumes of false positives, because a substantial share of those claims fall under Knock for Knock and are not pursued. A model built for the Indian market has to encode the Knock for Knock parameters and the insurer's own participation in the agreement, so that it surfaces only the recoveries that are actually pursuable, typically those involving uninsured at-fault parties, commercial-vehicle counterparties outside the agreement, or losses above the agreed thresholds.
Third-party bodily-injury recoveries run through a different forum entirely, the Motor Accident Claims Tribunals (MACT) established under the Motor Vehicles Act. These are long-tail, document-heavy, and adversarial, and the recovery economics depend on fault determination, police records, and the tribunal's award practice. AI helps less with the litigation itself and more with triaging which files are worth the tribunal route and assembling the evidence (FIR, charge sheet, surveyor and investigation reports) that the recovery turns on.
Property and commercial recoveries depend on liability evidence that Indian claims files often capture poorly. A contractor-caused fire or a carrier-caused cargo loss is recoverable only if the third-party cause and the contractual liability are documented at the time of loss, and in practice the surveyor report is the central document. Because the licensed surveyor's report drives both the policyholder settlement and the recovery, a model that reads surveyor reports to extract the cause-of-loss and any third-party contribution is reading the one document most likely to contain the recovery signal. Cargo recoveries add the carriage documents, the bill of lading, the carrier's limitation of liability under the relevant carriage law, and the time bars that can extinguish a recovery if missed, which a workflow system should track as hard deadlines.
A realistic recovery model for India therefore does three India-specific things: it filters motor opportunities through the Knock for Knock logic, it routes bodily-injury and contested matters to the right forum with the right evidence pack, and it treats the surveyor report and the carriage documents as the primary text to mine for liability and time-bar information.
Building the Capability: Data, Models and Workflow
Deploying AI subrogation recovery requires bringing together the claims data, the models that read it, and the workflow that acts on the output. Each element has practical requirements that determine whether the deployment delivers.
The data foundation is the insurer's claims data, including the structured claim records and the unstructured notes, surveyor reports, and investigation documents that hold the fault and causation information. The model's ability to identify recovery opportunities depends on the quality and accessibility of this data, and insurers with fragmented claims systems or poorly digitised documents face a data-preparation effort before the models can perform. The historical recovery data, which claims were pursued, which recovered, and for how much, is the training signal that lets the model learn what a good recovery opportunity looks like.
The models combine classification to identify recovery opportunities, estimation to value them and predict recovery probability, and language models to read the unstructured documents. The estimation models learn from the historical recovery outcomes to predict the recoverable amount and the probability of recovery, which drives the prioritisation. The language models read the free text that holds much of the causation and fault information that structured data omits. The combination is what surfaces and ranks the opportunities.
The workflow integration determines whether the model's output translates into recoveries. The identified and prioritised opportunities must flow into the recovery team's workflow at the right point, with the supporting evidence assembled and the pursuit tracked, so that handlers act on the model's output rather than leaving it in a report. An insurer that builds excellent identification models but does not integrate them into the recovery workflow captures little of the value, because the opportunities are surfaced but not pursued. The workflow integration, not the model sophistication, is often the binding constraint on realised recovery.
Measuring the Return and Avoiding the Common Pitfalls
Because AI subrogation is an economic investment rather than a compliance exercise, it should be measured against realised recovery outcomes, and the measurement reveals whether the deployment is working and where it is leaking value.
The headline measures are the recovery rate and the recovery value: the share of paid claims with a recovery opportunity that is actually pursued, the share of pursued recoveries that settle, and the rupee value recovered against the value identified. An insurer that tracks these before and after deployment can see whether the model is surfacing opportunities that convert into collected recoveries rather than into reports that go unactioned. The most useful comparison is the change in net recovery, the value recovered less the cost of pursuit, since the model's value is in collecting more while spending less effort on low-value cases.
Three pitfalls recur in deployments. The first is chasing identification accuracy while neglecting conversion: a model that flags more opportunities looks impressive, but if those opportunities do not convert to settlements the recovery rate does not move, and the effort is wasted. The measure that matters is converted recoveries, not flagged opportunities. The second is ignoring evidence preservation at first notification: many recoveries fail not because they were not identified but because the evidence needed to pursue them was not captured when the claim was first handled, so the model should feed back into the claims-intake process to preserve recovery evidence early. The third is stale models: recovery patterns shift as fault rules, counterparty behaviour, and the claims mix change, and a model trained once and left static degrades, so the model needs periodic retraining on fresh recovery outcomes.
A simple worked example shows why prioritisation, not raw identification, drives the return. Suppose a motor book pays 100 claims in a month that a model flags as carrying some third-party fault. After filtering out the claims that fall under Knock for Knock and the claims where the at-fault party is untraceable or uninsured, perhaps 30 are genuinely pursuable. Of those, a handful of larger commercial-vehicle and uninsured-driver matters account for most of the recoverable rupee value, while the long tail of small recoveries would cost more in handler time than they return. A model that ranks the 30 by expected net value lets the recovery team take the high-value cases first and batch or drop the uneconomic tail, which lifts net recovery even if the total number of files pursued falls. The figures here are illustrative, and each insurer should calibrate them against its own recovery history rather than assume a market average.
The insurers that capture the most value treat AI subrogation as an operating loop: identify, prioritise, pursue, measure the realised recovery, and feed the outcomes back into both the model and the claims-intake process. The loop, rather than any single model, is what turns surfaced opportunities into a durable lift in collected recoveries.
Governance, Accuracy and the Recovery Decision
AI subrogation operates within the governance expectations that apply to analytical systems in Indian insurance, and the recovery context shapes how those expectations apply.
The recovery decision is lower-risk than many AI claims applications because it does not produce an adverse outcome for the policyholder. Subrogation pursues a third party after the policyholder's own claim has been paid, so a model error that flags a non-recoverable claim wastes recovery effort but does not harm the policyholder, and a model that misses a recovery costs the insurer but again does not harm the policyholder. This lower-stakes profile means the governance burden is lighter than for models that decide policyholder claims or pricing, though the insurer still needs accuracy and oversight to ensure the recovery effort is well directed.
Accuracy matters for the economics rather than for fairness. A model that over-identifies recovery opportunities sends handlers after recoveries that are not there, wasting effort; a model that under-identifies leaves recoveries uncollected. The accuracy that matters most is in the prioritisation, because directing effort to the highest-value recoverable claims is where the economic return is made. Insurers should measure the model against realised recovery outcomes, the share of flagged opportunities that recover and the value recovered, and refine it on that feedback.
The recovery against another insurer raises a market dynamic worth noting. Much motor and liability subrogation is recovery from another insurer, and as more insurers deploy AI recovery, the pursuit becomes a contest between increasingly capable recovery operations on both sides. The insurer with better identification, evidence, and prioritisation recovers more and defends recoveries against it more effectively, so the capability becomes a competitive necessity rather than only an opportunity. Insurers that lag in recovery capability will find themselves paying recoveries to better-equipped counterparties while failing to collect their own.
For brokers and their commercial clients, the insurer's recovery capability is part of the claims service that affects the client's experience and ultimately the cost of cover. Platforms such as Sarvada are emerging in the Indian commercial broking market to connect insurer recovery and claims capability with broker claims advocacy, helping clients understand how their insurer handles the full claims lifecycle. Request Access to evaluate platform options.