AI in Insurance: Claims and Underwriting

Insurance is the vertical where the explanation is part of the product. A claim decision or a premium usually has to be communicable to the person it affects, and that single requirement eliminates the architecture most teams reach for first. This page is not legal or regulatory advice; insurance is regulated in detail and differently almost everywhere.

A decision someone is owed a reason for

Declining a claim, applying a sublimit, rating a risk higher than expected, non-renewing a policy — these are decisions that affect a person’s money at a bad moment, and across most markets there is some expectation that the reason can be stated, questioned and appealed. The expectation may come from regulation, from the policy contract, from a complaints scheme, or simply from a customer with a lawyer. The engineering consequence is the same in every case.

It is not satisfied by asking a model to explain itself. A generated justification for a generated decision is a plausible narrative produced after the fact, with no reliable causal relationship to whatever produced the decision — the faithfulness problem again. If a complaint is upheld and someone asks what the system would have done had one fact been different, a narrative cannot answer, and a rules layer can.

What that rules out, and what it leaves

Ruled out: the file goes in, the decision comes out. What is left is a separation that is worth stating precisely, because the value is in which component owns which job.

claim file ─▶ MODEL: facts only
                 date and time of loss, peril class, location, parties,
                 damaged items with claimed values, injuries alleged,
                 documents present / missing
                 — each with a citation into the source document

           ─▶ RULES: the policy wording, encoded
                 coverage in force on date of loss?
                 peril within insuring clause? excluded?
                 sublimit for this item class?
                 deductible, per-claim and aggregate
                 conditions precedent satisfied (notice period, reporting)

           ─▶ OUTPUT: decision + reason codes + the facts each code rested on
                 → adjuster reviews facts, not conclusions
                 → letter to the insured is generated FROM the reason codes

The model never reads the policy and decides. The policy wording is the authority; it is versioned, it differs by form and endorsement, and it is the thing a complaints body will read. What the model contributes is turning a fax, a phone transcript and thirty photographs into the structured facts the rules need — which is the part that was expensive and which no rules engine could ever do.

One benefit is easy to miss: this architecture makes the failure mode legible. When the answer is wrong, it is wrong because a fact was extracted incorrectly, and you can point at the citation and see it. A single-model design fails opaquely and identically whether the cause was a misread date or a misunderstood exclusion.

What a claim file actually contains

• First notice of loss as a phone call, so a transcript with speech recognition errors on exactly the words that matter — street names, policy numbers, vehicle registrations. Confirm those against the policy record rather than trusting the transcript.
• Adjuster and handler notes, written in an internal shorthand, containing opinions and reserve figures that must not reach the insured or a generated letter.
• Photographs, taken in bad light by someone who is upset, of a damaged vehicle or a flooded room, occasionally of the wrong property.
• Repair estimates from estimating platforms, which are structured, line-coded and should be parsed rather than read by a model.
• Medical bills and reports on bodily injury claims, carrying diagnosis and procedure codes and free text, with all the handling constraints of health data.
• Third-party documents — police reports, weather data, contractor invoices — arriving as scans of printouts of scans.

Where the money concentrates

Damage assessment from photographs is the most-demonstrated capability in insurance and the metric usually shown for it — average agreement with an adjuster on damage severity — hides where the risk is. Claims cost is not smoothly distributed across severity. It concentrates at two boundaries.

The first is the total loss threshold: the point at which repair cost plus salvage considerations tips a vehicle or property from repair to write-off. Being one band out in the middle of the range costs a little; being one band out at that boundary changes the settlement by the difference between a repair and a market value. Report agreement at the boundary separately, and set the review policy from it.

The second is sublimits and exclusions. A domestic contents claim may be fully covered in aggregate and severely limited for jewellery, cash or bicycles; a commercial property policy limits by peril. A system that reads a schedule of damaged items and applies the headline limit produces a number that is right for most claims and badly wrong for the ones with an expensive item in a limited category. Item classification therefore needs the same per-class recall treatment as the taxonomy work in manufacturing — rare, expensive classes reported on their own.

Fraud detection is a routing problem

Fraud is the application everyone asks about and the one that most needs a boundary drawn around it. A model flagging a claim as fraudulent, with that flag having any effect on the outcome, creates an adverse decision about a person on the basis of a pattern, plus a feedback loop: the model is trained on claims that were investigated, which were selected by prior suspicion, so it learns what previously attracted suspicion.

The defensible form is narrow. The model produces a triage score that routes a file to a human investigator and does nothing else — it does not decline, does not delay beyond normal handling, and is not visible to the customer. The investigation reaches its own conclusion on its own evidence. Measured that way, the metric is investigator yield at a fixed capacity: given that you can investigate a fixed number of claims a week, does this ordering find more than the previous rules did. That is a comparison you can actually run, and it does not require anyone to defend a score attached to a claimant.