Introduction
Fraud analytics is the use of data, rules and statistical models to find the policies and claims that deserve a closer look. It does not find fraud. It finds risk, ranks it, and hands the result to a person.
That distinction shapes everything else: which techniques suit which problem, where they sit in the claim, and how a claims or counter-fraud leader should judge whether they are working. Some people call the same thing fraud analysis, and this article treats the two as one.
What fraud analytics means (plain English)
In insurance, fraud analytics combines configurable rules, advanced analytics and machine learning to surface referrals for fraud or claims teams.¹ The usual output is a risk score. A score indicates how likely fraud is, relative to other claims. It is a measure of risk, not a definitive judgement.²
A threshold then turns the score into work. A claim is referred automatically only if its score passes a defined level, or if a specific rule fires.² ⁸
The term is also used in banking, where it describes the prevention of online financial fraud such as account takeover.³
The main techniques, and what each is good and bad at
This article groups the techniques into five families.
Rules. A rule encodes what your own people already know. Rules give control and transparency, and they enforce policy and appetite.¹ Their weakness is rigidity. Overly rigid rules are a common cause of false positives, and rules work from patterns that are already known.⁴ ⁵
Predictive models. Predictive modelling uses statistical and machine learning techniques to find relationships between input data and known outcomes. It scales across large claim volumes, applies consistent logic, and can identify subtle patterns that rules miss. It also carries risks. Models can degrade as behaviour changes, which is known as model drift. They can inherit bias from historical data, and they become hard to explain if they are overly complex.⁶
Anomaly detection. Anomaly detection builds a baseline of normal behaviour from historical data and flags activity that falls outside it. Most approaches are unsupervised, meaning they do not rely on pre-labelled fraud examples, which makes them useful where a new type of fraud has not yet been formally identified. The cost is noise. A poorly calibrated model flags too much, and unusual is not the same as dishonest. Seasonal patterns, life events and changes in a customer’s circumstances all produce anomalies that are entirely genuine.⁵
Network and link analysis. Network analysis maps the relationships between people, addresses, devices, vehicles, repairers and claims. Its value is in revealing organised fraud that looks legitimate when viewed one claim at a time. Network signals can also enrich scoring, so a claim that is part of a wider ring is escalated even when it does not look high risk on its own. The technique depends on clean identifiers. Without entity resolution and sensible matching thresholds, a network view over-links and misleads.⁷
Text. Claims files are full of unstructured content, including statements, reports and correspondence, and traditional analytics struggle with it.⁹ kbs lists text mining among the techniques in its claims detection.⁸ More recently, large language models have been used to summarise claim circumstances, flag unusual narrative elements, and identify inconsistencies between a statement and the supporting evidence. They bring their own risks, including output that is plausible but wrong, and reasoning that is difficult to reconstruct.⁹
The published kbs position is that most insurers get the best results by combining AI with rules: AI to spot subtle patterns, and rules to enforce policy and appetite.¹
Where it sits in the life of a policy and a claim
Analytics is not a single checkpoint. It appears at four points.
At quote and inception. Policy fraud detection looks for suspicious behaviour at quote, application and inception, and across the life of the policy. Applications and policy events can be scored in near real time to support a referral or step-up verification. Catching a problem here protects pricing accuracy and reduces downstream claims losses.¹⁰
At the claim. In the kbs platform, claims are grouped into segments, for example by product, channel or claim type, and scored using a combination of rules and predictive models. Scoring is tailored by segment so that a single blanket model is not applied to everything. Handlers can still refer a claim manually when something looks unusual, and those referrals are tracked alongside the automated ones.⁸
In the investigation. A referral should arrive with its reasons. That means reason codes, the key features behind the score and the linked entities, so an analyst can validate it quickly.⁸ From there, link analysis lets an investigator move from a single claim to the wider network around it.⁷
After the outcome. Confirmed outcomes feed back into thresholds, rules and models. kbs describes this loop as the way to keep alert volumes manageable and to improve hit rate over time.¹
What it cannot do
It does not make the decision. Automated systems generate scores, alerts and recommendations, and people make or validate the final decision. That applies most of all to high-impact outcomes such as a fraud investigation or a claim denial, where accountability cannot be fully automated.¹¹
It will sometimes be wrong. False positives and false negatives are both unavoidable in fraud detection. Pushing one down tends to push the other up, so the real question is where to set the balance, given investigator capacity, claim value, customer experience and regulatory expectations. A false positive is legitimate behaviour flagged as suspicious, and in volume it frustrates genuine customers.⁴
It needs maintenance. Threshold tuning is not a one-time activity. Thresholds need review as volumes, behaviour and resources change.²
It depends on its data. Incomplete or inconsistent data distorts the patterns that detection relies on. The result is more false positives, more false negatives and a heavier workload for investigators.¹²
How to judge it
Alert volume is the wrong measure. Generating more alerts does not equate to better detection, and low-quality alerts overwhelm the people who have to clear them.¹³
kbs lists the common success measures as:
- Hit rate, meaning the share of referrals confirmed as fraud
- False-positive rate
- Time-to-triage and investigator throughput
- Leakage prevented and recoveries
- Customer impacts, such as complaint rates and cycle time¹
Two refinements are worth the effort. Track hit rate by source, so that rules, model scores, handler referrals and network signals are each measured on what they deliver. That shows which rules are over-firing and which thresholds need tuning. And segment the results by claim type, product or channel, to show where detection is making the biggest difference.¹³
The Knowledge Hub covers the buying side of this in Choosing fraud detection software.
Governance in Australia
Two regulators, the privacy commissioner and the industry code each have something to say.
In its letter to industry of 30 April 2026, APRA named claims triage and fraud and scam disruption among the areas where regulated entities are trialling or introducing AI. It found that governance has not matured at the same pace. Among its minimum expectations are an inventory of AI tooling and use cases, ownership and accountability across the AI lifecycle, and human involvement for high-risk decisions. APRA also observed that point-in-time and sample-based assurance is ill suited to probabilistic models that learn, adapt and degrade over time, and that few entities had continuous monitoring in place to detect model drift or bias.¹⁴
ASIC made a related point in October 2024, after reviewing AI use by licensees that included general and life insurers. It warned that governance could lag adoption. It also said licensees must consider their existing obligations and duties, and avoid simply waiting for AI laws to be introduced.¹⁵
The General Insurance Code of Practice reaches the detection rules themselves. Under paragraph 195, an insurer that has adopted the Code must have its quality assurance programme review its non-genuine claims indicators at least once a year, to make sure they remain relevant, appropriate and do not discriminate.¹⁶
Privacy law applies as well. The OAIC’s guidance on AI products says privacy obligations apply to personal information put into an AI system and to the output it generates. Where an AI system infers personal information, that is a collection under Australian Privacy Principle 3. The guidance also asks organisations to make sure their records show where information is the product of an AI output, and is therefore a probabilistic assessment and not a fact.¹⁷ That is a sound discipline for any risk score attached to a named person.
For a director, these four sources point to the same working list: know which models and rules are running, know who owns each one, keep a person in the decision, and be able to show how the indicators were last reviewed.
Related Topics
Risk scoring and prioritisation
False positives and false negatives
The 5 fraud ops metrics that matter most
Graph neural networks in insurance fraud
Sources & further reading
¹ kbs Intelligence, Detect page and FAQs, kbsintelligence.io/detect/
² kbs Intelligence, Jargon Buster: Risk Scoring and Prioritisation; Threshold Tuning and Operational Alignment
³ kbs Intelligence, Fraud Analytics, Knowledge Hub, 26 March 2025
⁴ kbs Intelligence, Jargon Buster: False Positives and False Negatives
⁵ kbs Intelligence, Jargon Buster: Anomaly Detection in Insurance Fraud
⁶ kbs Intelligence, Jargon Buster: Predictive Modelling in Insurance
⁷ kbs Intelligence, Link Analysis / Visualisation page and FAQs; Jargon Buster: Network Analysis in Insurance Fraud
⁸ kbs Intelligence, Claims Fraud Detection page and FAQs, kbsintelligence.io/detect/claims-fraud-detection/
⁹ kbs Intelligence, Large Language Models in Claims, Knowledge Hub
¹⁰ kbs Intelligence, Policy Fraud Detection page and FAQs
¹¹ kbs Intelligence, Jargon Buster: Human-in-the-Loop Decision Making
¹² kbs Intelligence, Jargon Buster: Data Quality in Insurance Fraud Detection
¹³ kbs Intelligence, The 5 Fraud Ops Metrics That Matter Most for Insurance Fraud Detection, Knowledge Hub, 25 March 2026; Jargon Buster: Quality of Alerts in Fraud Detection
¹⁴ APRA, Letter to Industry on Artificial Intelligence (AI), 30 April 2026
¹⁵ ASIC, media release 24-238MR, ASIC warns governance gap could emerge in first report on AI adoption by licensees, 29 October 2024, announcing Report 798
¹⁶ General Insurance Code of Practice, October 2023 update, paragraph 195, as set out in kbs Intelligence, General Insurance Code of Practice and claims fraud, Knowledge Hub
¹⁷ Office of the Australian Information Commissioner, Guidance on privacy and the use of commercially available AI products, 21 October 2024
