AI-Assisted Insurance Claims Operations Platform
Property & Casualty Insurance, Applied Machine Learning, Document Intelligence, Model Risk & Market Conduct Governance
The Problem
Claims leakage is not a fraud problem. It’s a data problem that everyone has agreed to call a fraud problem.
The industry number for leakage — the gap between what was paid and what should have been paid under the policy — sits somewhere between 2% and 5% of indemnity spend depending on who’s counting and how honestly. Fraud is a slice of that. The larger slice is duplicate payments, missed subrogation, rental and storage running unchallenged, invoices settled without line-item review, reserves set on day one by an adjuster with eleven minutes and no comparable claims to hand, and third-party costs that were negotiable right up until the point nobody negotiated them.
A typical operation makes this hard in specific ways. Auto sits in one claims system, property in another, general liability in a third that was migrated off a mainframe in 2016 and still carries the mainframe’s field lengths. Program and delegated authority business arrives as monthly bordereaux from MGAs and TPAs — spreadsheets, one schema per partner, twelve partners, none of them stable across renewal. FNOL arrives by phone, portal, agent email, and increasingly by represented-party demand package with a pre-built medical narrative attached. Supply chain data lives with the vendors.
The operational ask is easy to state: settle the simple claims faster, put adjuster time where indemnity is actually at risk, and stop paying money the policy never owed.
The constraints are what make it an engineering problem rather than a procurement one:
- Entity resolution with no shared key and active adversarial pressure. The same claimant appears across policies, as a third party on someone else’s claim, and under a variant spelling with a different date of birth. The same body shop, treating clinic, and attorney recur across unrelated claims — that recurrence is the single most useful signal you have, and it only exists if resolution works. Deterministic matching on policy number, VIN, and standardized address gets you most of the book; the residual is where organized activity lives.
- Label delay measured in years. Ultimate claim cost is only known at settlement. Auto physical damage closes in weeks. Represented bodily injury runs two to five years, longer where a policy-limits demand, a structured settlement, or litigation is in play. Any model trained on closed claims is trained on a book that has been filtered for speed of settlement, which is precisely the variable you’re trying to predict.
- Targets that are decisions, not facts. Case reserves are adjuster judgment. SIU referrals are investigator judgment. Train on them and you have built a model that predicts what an adjuster would have done in 2022, complete with their unit’s escalation culture and their supervisor’s authority limits.
- Regulatory surface across two axes at once. Market conduct: state Unfair Claims Settlement Practices statutes, prompt-pay and acknowledgment deadlines that vary by state, first- and third-party bad faith exposure, and periodic market conduct exams that will ask how the automated paths made their decisions. Prudential: anything influencing case reserves feeds statutory unpaid claim estimates, Schedule P development, and the Appointed Actuary’s opinion. A model that quietly changes reserving behavior has changed a reported number.
The Architecture
Everything lands raw and immutable before anything touches it. Claims system extracts from ClaimCenter, Duck Creek, or the legacy platform; ACORD messages where the agent and carrier channels support them; bordereaux as-received with the original file preserved alongside the parsed output; FNOL call transcripts; correspondence; and the document estate. ISO ClaimSearch responses are ingested as first-class records, not as screenshots pasted into a claim note.
Identity resolution runs probabilistically — Fellegi–Sunter scoring over name, date of birth, address, phone, vehicle registration and VIN, with CASS/DPV address standardization first and blocking on ZIP to keep the comparison space tractable. Three outputs, not two: matched, unmatched, and a review band that goes to a human queue with the evidence shown. Auto-merging a marginal pair is how you end up with two unrelated claimants sharing a claims history, and that is a consumer harm and a privacy incident, not a model quality issue.
The entity graph is a first-class artifact, not a by-product. Claimants, policies, vehicles, addresses, body shops, rental and towing vendors, medical providers, independent adjusters, and attorneys as nodes; claims as the edges between them. Most of the durable value in the platform comes from asking questions of this graph that no single claims system can answer.
The data layer
Bitemporal throughout — valid time and transaction time on every claim state, reserve movement, and payment. Two reasons, both non-negotiable. First, model training requires point-in-time correctness: features must be reconstructed as they stood at FNOL, or at triage, not as they stand today. A reserve that was revised upward in month nine is not a feature available in week one, and a training set that includes it will produce a model that looks extraordinary in backtest and useless in production. Second, the Appointed Actuary and the market conduct examiner will both ask what the system knew on a specific date. “We can rebuild the state” is an acceptable answer. “We overwrote it” is not.
Medallion structure — raw, conformed, curated. The conformed layer carries a canonical claim model that survives source system replacement, because the source systems will be replaced and the models should not have to care.
Document intelligence
The document estate is where the recoverable money is: CCC, Mitchell, and Audatex estimates, Xactimate property scopes, independent adjuster and engineer reports, medical bills and CPT-coded treatment records, rental agreements, demand packages, and schedules of loss.
Layout-aware extraction rather than flat OCR, because the unit of interest is the line item, not the document. The question is never “what does this estimate total” — it’s whether these refinish hours are consistent with this panel schedule, whether the rental period is consistent with repair duration and parts delay, whether these treatment dates are consistent with the reported mechanism of injury, and whether this line has already been paid on this claim under a different reference.
Extraction confidence routes the work. High confidence flows to automated validation. Low confidence goes to an adjuster with the source region highlighted, and their correction is captured as training data. No extracted value drives a payment decision without either passing a confidence threshold or being confirmed by a human — the marginal saving from automating the uncertain tail is small, and the tail is where the expensive mistakes are.
The model layer
Complexity triage at FNOL: Predicts handling path — straight-through, fast-track, standard, complex or litigation-likely — from what is genuinely known in the first hour. Trained on claims that have run to settlement, with survivorship handled explicitly rather than by hoping. Attorney representation probability is the single highest-value output here, because it changes the handling economics of the file more than severity does.
Ultimate cost as a distribution, not a point: Quantile regression, not conditional mean. A reserving philosophy is a statement about which percentile you sit at; a point estimate cannot express one. It also makes the uncertainty visible to the adjuster, which is the difference between a decision support tool and a number that appears in a box.
Recovery and subrogation identification: Structurally the highest-return model in the stack and consistently the least deployed, because nobody’s bonus depends on it. Liability indicators from FNOL narrative, third-party carrier presence, product and premises exposure, PIP recovery rights in no-fault states, and arbitration-eligible files that currently close without a demand ever being filed.
Anomaly and network risk: Two components: claim-level features, and graph features from the entity network — shared providers across unrelated claimants, unusual density between a body shop and a specific clinic, staged-loss patterns in no-fault jurisdictions, velocity of first-party claims against newly written policies, and post-CAT vendor patterns in property. The graph component is what finds organized activity; the claim-level component only ever finds opportunists.
Four models doing four different jobs, deliberately not one model doing all of them:
Complexity triage at FNOL. Predicts handling path — straight-through, fast-track, standard, complex or litigation-likely — from what is genuinely known in the first hour. Trained on claims that have run to settlement, with survivorship handled explicitly rather than by hoping. Attorney representation probability is the single highest-value output here, because it changes the handling economics of the file more than severity does.
Ultimate cost as a distribution, not a point. Quantile regression, not conditional mean. A reserving philosophy is a statement about which percentile you sit at; a point estimate cannot express one. It also makes the uncertainty visible to the adjuster, which is the difference between a decision support tool and a number that appears in a box.
Recovery and subrogation identification. Structurally the highest-return model in the stack and consistently the least deployed, because nobody’s bonus depends on it. Liability indicators from FNOL narrative, third-party carrier presence, product and premises exposure, PIP recovery rights in no-fault states, and arbitration-eligible files that currently close without a demand ever being filed.
Anomaly and network risk. Two components: claim-level features, and graph features from the entity network — shared providers across unrelated claimants, unusual density between a body shop and a specific clinic, staged-loss patterns in no-fault jurisdictions, velocity of first-party claims against newly written policies, and post-CAT vendor patterns in property. The graph component is what finds organized activity; the claim-level component only ever finds opportunists.
The Calls We'd Defend
Never automate a denial. Automate approval, automate straight-through settlement within limits, automate routing and reserve recommendation. Denials, reservations of rights, and coverage positions stay human, with the model output visible as one input among several. The asymmetry is total: a false positive on approval costs the value of the claim; a false positive on denial costs a DOI complaint, a market conduct finding, and — in the wrong state, on the wrong file — a bad faith verdict that dwarfs the policy limit.
Referral thresholds are set by SIU capacity, not by the ROC curve. If the investigation unit can work forty referrals a week, the threshold is the one that produces forty referrals a week, and the model’s job is to make those forty the best available forty. A queue running at three times capacity gets ignored, and an ignored queue is worse than no model at all — because in states with mandatory suspected fraud reporting, it manufactures a documented record of files the carrier flagged and then did nothing about.
Publish nothing that reveals a threshold. The moment a straight-through settlement limit is inferable from outside, estimates and demands cluster immediately beneath it. This is not hypothetical and it is not slow. Thresholds get randomized within a band, a proportion of below-threshold claims are sampled into full review regardless of score, and the sampled population doubles as the unbiased evaluation set — the only claims where you observe the outcome the model didn’t influence.
Calibration over discrimination. A severity model with a strong Gini and poor calibration will set reserves that are wrong in a consistent direction, and consistent directional error in reserving is a Schedule P conversation with the Appointed Actuary. Isotonic calibration, monitored by segment and by state — not just in aggregate, where a well-calibrated overall model routinely hides two badly calibrated halves.
Reason codes on every output, at adjuster comprehension level. Not SHAP values in a tooltip. “Flagged: this claimant’s treating clinic has appeared on 14 claims with this attorney in 90 days.” An adjuster who cannot explain the flag in a claim note, to a supervisor, or in a deposition will stop using the system — and they will be right to.
The adjuster override is a monitored signal, not a failure. Override rates by unit, by segment, and by model are the earliest available warning of drift — they move before performance metrics do, because adjusters see the change in the book before the labels arrive to prove it.
Governance
Model risk management aligned to the existing three-lines structure rather than assembled beside it: first line owns the model in operation, second line validates and monitors, third line audits. Every version carries documented lineage of training data, feature definitions, evaluation, calibration by segment, and known limitations — mapped to ASOP 56 where actuarial work product depends on it.
Where model output influences case reserves, the Appointed Actuary reviews and signs off before deployment, and the reserving impact of a version change is quantified before it ships rather than discovered in the next Schedule P triangle.
On AI-specific regulation: the NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers has now been adopted in a large majority of states, and it expects a written AIS program covering governance, risk management, testing, documentation, and third-party model oversight — including for claims, not only underwriting and pricing. Colorado’s quantitative testing framework and New York DFS Circular Letter No. 7 currently reach underwriting and pricing rather than claims, but they establish the direction of travel and the evidentiary standard examiners will arrive with. Build the file now; producing it retroactively during an exam is how a good program becomes a finding.
Fairness testing runs on protected and proxy characteristics across triage path, referral rate, settlement ratio, and cycle time — with proxies treated seriously, because ZIP code, vehicle age, and language of the FNOL call are proxies whatever the model card says. Unfair discrimination is a state statutory standard, not a philosophical position, and disparities that survive controls stop the deployment.
Data handling under GLBA, the state insurance data security regime, and CCPA/CPRA where applicable — with retention and purpose limitation applied specifically to the entity graph. It is the most powerful asset in the platform and the one most likely to be found disproportionate if it is retained indefinitely without a defensible basis.
What Good Looks Like
Modeled targets for a mature deployment, in the order they usually arrive:
- Straight-through settlement rate on eligible low-complexity claims: 25–40%
- Cycle time on fast-track auto physical damage: reduced by 30–50%
- Adjuster time reallocated from routing and data entry to indemnity-critical work: 20–30%
- Subrogation and PIP recoveries identified and pursued: 10–20% uplift over baseline
- Indemnity leakage identified in file audit: reduced by 1–2 points of spend
- ALAE per closed claim: reduced on the modeled cohort without a corresponding rise in indemnity
- Reserve development on the modeled cohort: tighter and less directionally biased at 12 and 24 months
- DOI complaint rate and reopened-claim rate on automated paths: flat or lower than the manual baseline
The first four fund the program. The last four are what stop it being switched off after the first market conduct exam.
What This Unlocks
The resolved entity graph and the bitemporal claim history are the durable assets. The models are not — they will be retrained, replaced, or made obsolete within eighteen months, and that is fine.
Once identity, temporality, and document line items are solved, the next use cases stop being programs and become integration work: pricing feedback from realized claims experience rather than modeled expectation, vendor and DRP network performance measured on indemnity outcome instead of invoice volume, CAT exposure and surge staffing forecasting, and portfolio-level early warning on emerging litigation patterns and social inflation by venue.
The claims system will be replaced. The data foundation won’t be.