Claims Fraud Detection

Identifying claims that are exaggerated, fabricated, or organised, without slowing down the genuine claims that make up the overwhelming majority.

AI is suitable with conditions

The problem has the shape AI handles well, which is ranking many cases by risk using weak signals that no single rule captures. It is conditional rather than strong because the value depends on confirmed outcome labels and reliable entity data, and many insurers do not have either in usable form. Without those, a rules engine and better data plumbing deliver more than a model will.

Problem Statement

A small fraction of claims are exaggerated, fabricated, or organised.

Finding them is not the hard part.

The hard part is finding them without imposing delay and suspicion on the overwhelming majority of claimants who are telling the truth.

Business Context

Claims fraud detection is a ranking problem constrained by a fixed investigation budget.

An insurer cannot investigate every claim, so the practical question is never “is this claim fraudulent”.

It is “of the claims we could investigate today, which ones should we spend our limited investigator hours on”.

This framing matters because it changes which metric counts.

A model that finds more fraud in theory but fills the queue with cases nobody has time to work has not improved anything.

Why This Is Hard

Fraud is adversarial.

Detection patterns lose value once they become known, because fraudsters adapt to them.

Labels are also delayed and incomplete.

A claim confirmed as fraudulent today may have been paid eighteen months ago, and claims never investigated are unlabelled rather than legitimate.

Treating unlabelled claims as negative examples is a serious modelling error, because a claim nobody investigated has no verdict at all.

How The Solutions Fit Together

Rules encode what the business already knows and provide the explainable floor.

Supervised scoring ranks the remainder by learned risk, combining signals too weak to justify a rule of their own.

Anomaly detection supplies discovery, surfacing patterns nobody has described yet and generating candidates worth labelling.

Graph analysis changes the unit of investigation from the claim to the network, which is how organised rings become visible.

None of these replaces an investigator.

Each one changes what lands on the investigator’s desk first.

Implementation Considerations

Start by fixing outcome recording, because every later approach depends on it.

Then tune and retire the existing rules, which is unglamorous and, in this playbook’s judgement, worth doing before any model.

Only then introduce scoring, and introduce it as a ranking over the existing queue rather than as a new queue.

Size the referral threshold to the capacity that actually exists.

Measure the model against the queue it replaced, not against a holdout set nobody will ever work.

ROI Considerations

The return comes from three places.

Leakage prevented on claims that would have been paid, investigator hours redirected from low-yield referrals to high-yield ones, and reduced friction for genuine claimants.

The third is the easiest to ignore, and in this playbook’s judgement it can be the largest, because delay drives churn among exactly the customers an insurer wants to keep.

Security, Privacy, and Regulation

Claims data is sensitive personal data, and in many lines it includes medical information.

Access should be limited to the investigation function, scoring should be logged, and any model influencing a customer outcome should be explainable in terms a regulator will accept.

Check for disparate impact across protected groups before deployment and on a regular cycle after it.

A detection system that is more suspicious of some customers than others is a legal problem before it is a technical one.

Symptoms

Loss ratio is worse than pricing assumptions and the gap is not explained by frequency or severity trends.
Investigators spend much of their time on referrals that turn out to be legitimate.
The same repair shops, medical providers, or addresses recur across apparently unrelated claims.
Fraud is often discovered after payment rather than before it.
Genuine customers complain about delays caused by checks that find nothing.

Root causes

Detection rules were written for patterns that are now widely known and avoided.
Claims data is fragmented across intake, policy, and payment systems, so connections are invisible.
Investigation capacity is fixed while claim volume is not, so referrals are triaged by queue order rather than by risk.
Confirmed fraud outcomes are recorded inconsistently, so nobody can measure what detection is actually catching.
Speed targets and accuracy targets are owned by different teams with different incentives.

Business impact

Direct leakage on every fraudulent claim that is paid.
Investigator time consumed by low-yield referrals.
Delay and friction imposed on genuine claimants, which drives churn.
Regulatory and reputational exposure when checks are applied inconsistently across customer groups.

Traditional approaches

Fixed rules and red-flag checklists applied at intake.
Manual referral by adjusters based on experience and instinct.
Random audit sampling of settled claims.
Industry-wide shared fraud registers.

Solution approaches

Each of these is written for this problem specifically, not borrowed from a generic catalogue.

Recommendation

Run rules and a supervised risk score together, feed both into one ranked investigator queue sized to actual capacity, and use graph analysis for organised activity. Keep the settlement decision with a human investigator. Automated refusal is the wrong place to spend the accuracy you have.

Data requirements

Claim record high sensitivity: First notification of loss, claim type, amounts claimed and paid, and the timeline of status changes.
Policy and customer history high sensitivity: Policy inception, prior claims, endorsements, and payment history, joined to the claim at scoring time.
Confirmed outcome labels high sensitivity: Investigated cases with a recorded verdict and the date the verdict was reached, which is what supervised learning trains on.
Entity identifiers high sensitivity: People, vehicles, addresses, phone numbers, bank accounts, repair shops, and medical providers, resolved consistently across claims.
Investigation effort data medium sensitivity: Time spent per referral and its outcome, which is how referral precision is measured against capacity.

Evaluation metrics

Precision at the referral volume investigators can genuinely work, not precision in the abstract.
Value of confirmed fraud prevented per investigator hour.
False positive rate measured against genuine claimants, segmented to check for disparate impact.
Time to settle for claims that are never referred.
Score drift and label delay, tracked continuously.

Human oversight

A trained investigator makes every accusation and every refusal. The model orders the queue, and it explains why a case was ranked highly, but it never declines a claim on its own. Any coordinated action against a suspected ring is confirmed by a specialist first.

Common pitfalls

Training on adjuster referral decisions rather than confirmed outcomes, which teaches the model to imitate existing bias instead of detecting fraud.
Using features that only exist after an investigation has started, which produces excellent offline metrics and a useless production model.
Optimising recall when investigation capacity is the binding constraint, which floods the queue and lowers yield.
Ignoring that fraud is adversarial, so a static model decays as fraudsters learn what it catches.
Treating a low-precision anomaly queue as a referral queue and burning investigator trust in the first month.

When not to use AI

Only a small number of confirmed fraud outcomes exist, as a rule of thumb fewer than a few hundred, which is not enough to train or to evaluate honestly.
Claims data cannot yet be joined across systems, in which case data integration is the real project.
The insurer operates where state regulators have adopted the NAIC model bulletin, which expects AI decisions to be accurate and not unfairly discriminatory, and the organisation cannot yet explain a model score.
Existing rules have not been tuned or retired in years, because that cheaper work is, in this playbook's judgement, worth doing before any model.

Sources

Agent context

A distilled brief for an AI agent or an engineer scoping this work. It carries the constraints and pitfalls that change an implementation decision, not the whole page.

Get the agent context as Markdown

Objective

Rank open claims by fraud risk so that a fixed number of investigator hours is spent on the highest-value cases, without automating any refusal decision.

Constraints

Never decline, delay, or reduce a claim automatically. The system ranks, a human decides.
Train only on confirmed investigation outcomes, never on adjuster referral decisions.
Use only features available at scoring time, and prove it by reconstructing the feature set as of the scoring timestamp.
Size the referral threshold to real investigator capacity rather than to a target recall.
Log every score with the feature values that produced it, for audit and appeal.
Test for disparate impact across protected groups before deployment and on a recurring schedule.

Inputs

Claim records with amounts, timeline, and status history.
Policy and prior claim history joined as of the scoring timestamp.
Resolved entity identifiers for people, vehicles, addresses, providers, and bank accounts.
Confirmed investigation outcomes with verdict and verdict date.
Investigator capacity expressed as referrals per day.

Outputs

A calibrated risk score per open claim.
The ranked referral queue truncated at investigator capacity.
The top contributing factors for each referred claim, in language an investigator can act on.
Cluster summaries for suspected organised activity, listing the shared entities.

Success criteria

Referral precision improves against the queue the system replaced, measured on the same capacity.
Confirmed fraud value prevented per investigator hour increases.
Time to settle for non-referred claims does not get worse.
No measurable disparate impact across protected groups.
Score drift and label delay are monitored with an alert threshold agreed before launch.

Last updated 2026-09-07