The claim that was scored before anyone read it

Take a bodily-injury file and follow it as a sequence of events. The models act at particular moments, and the moments are not the ones most discussions of AI in claims talk about.

Notice. The claim arrives with an injury reported. Within minutes, before any human has read anything, two predictions are attached: how likely this file is to end in litigation, and what it is likely to cost. The inputs are what intake collected, what the policy record holds, and whatever the insurer has learned from files that resembled this one — the injury described, the mechanism of the collision, whether an ambulance attended, the delay in reporting, the venue the loss occurred in, the claimant’s claim history, sometimes the vehicle, sometimes the text of the notification itself.

Routing. The severity prediction decides who gets the file and how much authority they carry. This is the least controversial use and probably the most valuable: a file predicted to be expensive going to an experienced handler on day one, rather than being discovered to be expensive months later by a handler who cannot settle it, is a straightforward gain for everyone, claimant included.

The reserve. The insurer must estimate what this claim will ultimately cost and hold that amount. Historically the initial reserve was a handler’s judgement made on almost no information, and it was reliably wrong in both directions. A model fitted across thousands of closed files does this better. That is not a small improvement: under-reserving produces low offers and reserve strengthening later, over-reserving locks up capital and distorts how the account is rated, and both have real consequences for real files.

The first contact. Somewhere in the first days the handler speaks to the claimant, and by now the file already carries a propensity score. If it says this claimant is likely to instruct a lawyer, the handling changes. Contact is faster and more attentive, an offer may be made earlier, and the file is worked to reach a resolution before representation arrives.

Here is where the first genuinely difficult question sits, and it is not answered by whether the prediction is accurate. Treating a claimant more attentively because a model predicts they will lawyer up is either good service or a manoeuvre, and which one it is depends entirely on whether the early offer is adequate. An early, fair offer to a claimant the insurer expects to litigate is a better outcome than a slow, contested one. An early, low offer made specifically to close a file before the claimant learns what it is worth is a different act, and both are produced by the same score and look the same in the workflow.

Representation. A lawyer comes on the file. From that moment the insurer’s own data says the claim will cost more, which is true as a prediction and circular as a justification: claims with lawyers cost more partly because lawyers obtain more, and partly because claimants who instruct lawyers had worse injuries in the first place. The model does not separate those and does not need to for reserving purposes. It matters greatly if the same number is used to argue about what the claim deserves.

Plaintiff counsel, meanwhile, is running an inference of their own. Counsel who handles volume against the same insurers learns their patterns: which files get quick generous offers, which get silence, where authority sits, how the posture changes at particular stages. Segmentation that is legible from the outside becomes a strategy from the outside. The early-offer policy for high-propensity files is, from a claimant firm’s perspective, a reason to make representation visible early.

Reserve movement. The prediction updates as the file develops, and the reserve moves with it. Those movements are recorded, dated, and later readable as a history of what the insurer thought this claim was worth at each point. Handlers have always known that reserve history is sensitive. What is new is that the movements are now partly produced by a system whose reasoning is not written down in the file, so the history shows a number changing without showing a mind changing.

The settlement window. The offer is made, and the authority behind it came from somewhere. If the severity prediction informed the authority, the insurer has a recorded estimate of the claim’s ultimate cost sitting alongside its offer, and the difference between them is a number that now exists in a system somebody may one day read. It may be entirely explicable — ultimate cost includes defence expense, interest and outcomes that may not occur, and is not the same quantity as today’s fair settlement. But the explanation has to be in the file, written at the time, because reconstructing it two years later in response to a bad-faith allegation is the weakest possible position.

Litigation. Suit is filed, and the internal documents become a subject rather than a background. What the insurer predicted, when it predicted it, what it reserved, what it offered, and what the handler recorded between those events are all potentially in play. What is actually obtainable, and what privilege or work-product protection covers it, is a question of procedural law that this piece answers for no jurisdiction; read it as a reason to build the file carefully rather than as a statement of any rule.

Two features of the file are worth attending to regardless of how that question resolves. The first is whether the claim-specific reasoning exists at all: liability assessed on the evidence, injury assessed on the records, quantum assessed against comparable outcomes, written by the handler. The second is whether the sequence of events reads as reasoning or as compliance with a number. Those are different files, and only one of them is defensible.

What the models actually learned

Everything above assumes the predictions are good. They are, in the narrow sense that matters to an actuary: they predict what closed files cost. That is also the limitation, and it is structural rather than a defect of implementation.

A severity model fitted on settled claims learns settlement outcomes. Where past outcomes were depressed — unrepresented claimants who accepted early offers, claim types the insurer historically undervalued, venues where claimants rarely pursued — the model learns those amounts as the value of such claims. Used for reserving, that is exactly right: the model is predicting what the insurer will pay, and it will pay that. Used as a view of what a claim is worth, it encodes the past into the present with the authority of a computation.

The propensity model has a sharper version of the same problem, because acting on the prediction changes the label. Predict litigation, settle early to prevent it, and the file closes without litigation, which the model records as a case where litigation did not occur. Over enough cycles the system is learning about its own interventions and reporting the result as a property of claimants.

The judgement

A severity model is a reserving tool that has been handed a second job nobody wrote down. As a reserving tool it is better than what it replaced and its adoption is unambiguously good. As an input to negotiation it carries the insurer’s own history into a conversation about what one injured person is owed, with none of the visible reasoning that a handler’s valuation used to leave behind.

The governance that follows is not complicated and mostly is not in place: the handler’s valuation recorded before the prediction is seen, the reason for each reserve movement written in words, the difference between predicted ultimate cost and offered settlement explained on the file rather than in a policy document, and retention of enough of the model’s state to reconstruct a score months later. All four are things an insurer can do without waiting to be told.

The prediction and the offer now live in the same system, a few fields apart. Whatever the gap between them means, it is written down, and somebody will eventually read it.