An AI Mission for Insurance: Subrogation Review

Insurers treat subrogation as a judgment problem — whether a file is worth pursuing. It is really a reading problem, and the money is lost long before anyone gets to exercise judgment.

Somewhere in the middle of a routine auto claim, an adjuster types a sentence that will decide whether the carrier ever sees its money again. The insured mentions, almost in passing, that the other vehicle had a company logo on the door and the driver said he was on his way to a job site. The adjuster records it faithfully in the notes, because adjusters are conscientious people and that is what the notes are for, and then moves on to the next of the several dozen files that need attention that week. The claim is investigated, the damage is estimated, the payment goes out, the file closes. The sentence sits in the file, perfectly preserved and completely unread, until the recovery window quietly expires and the matter is archived as settled. Nobody made a bad decision. Nobody made any decision at all.

This is what makes subrogation such a strange line on an insurer's books. Unlike most of the loss column, it is not money spent on risk that materialized as expected. It is money the carrier has already paid out and is, in principle, entitled to get back from whoever actually caused the loss. It is the closest thing in the business to a recoverable asset sitting in plain sight. And the industry's shared experience is that a meaningful portion of it is never recovered — not because the claims were weak, not because the recovery effort failed, but because the file was never identified as a recovery candidate in the first place. The loss is silent. There is no adverse outcome, no denied demand, no unfavorable settlement to review. There is only a closed file that looks exactly like every other closed file, and a sentence in the notes that nobody re-read.

The identification step is a reading problem, not a judgment problem

Carriers invest enormously in the part of subrogation that comes after identification. There are dedicated recovery units, escalation criteria, negotiation playbooks, vendors, and metrics. Every one of those investments improves what happens to a file once someone has decided it might have recovery potential. Almost none of them affect whether that decision gets made, and the decision is the entire constraint. A recovery function operating at world-class effectiveness on the files it receives is still capped by the referral rate, and the referral rate is set upstream by a frontline adjuster who is managing a caseload, working under cycle-time pressure, and being measured primarily on resolving the claim in front of them rather than on spotting a third party who might owe money on a file they are about to close.

That is not a criticism of adjusters, and it is important to be precise about why. The signals that indicate recovery potential are genuinely hard to catch in the flow of handling a claim. They are rarely in a structured field. They live in the narrative — a phrase in a recorded-statement summary, an offhand attribution of fault in a first notice of loss, a repair invoice that names a part that failed rather than a part that was damaged, a police report annotation, a property adjuster's description of a water loss that mentions an appliance and a model number, a contractor's name that appears in a photograph caption. Each of these is a thread that a trained recovery specialist would pull. Each of them appears in exactly one place, in unstructured text, inside one file among thousands, and stays visible for as long as one person happens to be reading that page. The signal is not hidden. It is merely diluted past the point where sustained human attention can be expected to find it.

Once you frame it that way, the shape of the problem changes. Subrogation identification is not a matter of adjusters exercising better judgment about the files they see. It is a matter of every file being read carefully, in full, against a consistent set of recovery indicators, including the files that were closed last quarter and the ones nobody had a reason to look at twice. That is a reading workload measured in tens of millions of pages across a book of business, refreshed constantly as new documents arrive, and no staffing model in the industry has ever been able to sustain it. Carriers have known this for decades. The response has largely been to sample — audit a slice of closed files, run periodic recovery sweeps, apply rules-based triggers to structured fields — which is a reasonable adaptation to a workload you cannot fully absorb, and which by construction leaves the majority of files unexamined.

The discipline is systematic re-examination of closed files

The instinct is to fix this at the front of the process: better training, better prompts in the claims system, a checkbox that asks the adjuster whether subrogation potential exists. These help at the margin and they inherit the same flaw, which is that they depend on a person under time pressure recognizing a pattern at the exact moment the file passes through their hands. Recovery indicators frequently do not exist yet at that moment. The repair invoice that names the failed component arrives after the estimate is approved. The police report supplement arrives weeks later. A pattern that is only visible across several claims — the same product, the same contractor, the same premises — is by definition invisible from inside any one of them. A process that only looks at a file once, while it is open, will structurally miss everything that becomes knowable afterward.

The more useful discipline runs the other direction. Treat the closed-file population as the primary corpus, not the exception queue. Re-read files after closure, re-read them again when new documents land, and read them not against an adjuster's recollection but against an explicit, versioned definition of what a recovery indicator looks like in each line of business. This is unglamorous work and it is precisely the kind of work that rewards machine-scale reading: the same careful pass over every document in every file, applied with the same standard on the ten-thousandth file as on the first, with no fatigue, no caseload, and no incentive to close.

This is what an AI Mission is actually for, and it is worth separating from the marketing that surrounds the category. Gartner has predicted that more than forty percent of agentic AI projects will be canceled by the end of 2027, citing unclear business value, escalating costs, and what the firm calls "agent washing" — existing tools relabeled as autonomous without any change in what they can do unattended. A subrogation review mission is the opposite kind of proposition, because its value is not speculative and its scope is narrow enough to be honest about. It reads claim files. It surfaces the ones that carry recovery indicators, with the specific passage that triggered the flag quoted back to the reviewer. It maintains a queue that a human recovery specialist works through. That is the whole of it, and it is enough, because identification was always the binding constraint.

The line the machine must not cross

Everything downstream of identification belongs to accountable people, and this boundary should be drawn hard rather than left to emerge. A system of this kind does not determine liability, does not decide whether a claim is worth pursuing, does not contact a third party or their carrier, and does not initiate anything resembling legal action. Those are determinations with legal consequence, professional responsibility attached, and real people on the other end of them, and they belong to the adjusters, recovery specialists, and counsel whose job it is to make them and answer for them. The machine's contribution is narrower and, precisely because it is narrower, defensible: it read the file, it found the sentence, and it put the sentence in front of someone qualified to decide what it means.

Building it this way is not only an ethical posture but a practical one, because it is what makes the output auditable. When a mission surfaces a file, the reviewer should see the passage, the document it came from, and the indicator definition it matched, so that a wrong flag is visibly wrong and the definition can be corrected rather than mistrusted wholesale. In platform terms this is the ordinary human-in-the-loop pattern — specialist agents reading against enterprise knowledge, a reasoning core assembling the case for review, and a human gate on every action that leaves the building. In StudioX's vocabulary that is a mission with observations attached; in any vendor's vocabulary it should mean the same thing, and a carrier evaluating one of these systems should insist on seeing exactly where the gate sits before anything else. This posture is central to what the autonomous enterprise literature describes as the difference between autonomy and abdication: the work of reading is delegated, the authority to act is not.

The mental model worth carrying out of this is that subrogation leakage is not a performance problem in the recovery department. It is an attention budget problem in the claims organization, and attention budgets do not respond to exhortation. Every carrier already owns the evidence it needs; it is sitting in the notes, in the invoices, in the reports, in files that were closed correctly by people doing their jobs well. The question is not whether the organization is good at pursuing recovery. It is how many files it can afford to read, and how many times, and for the first time that number does not have to be a function of how many people it employs.