AI is already used across insurance claims management, from FNOL (first notice of loss) and document handling to fraud detection and settlement. The more difficult question, however, is where it delivers measurable value without introducing unnecessary regulatory, reputational or operational risk.
In claims, the strongest applications tend to involve repetitive, evidence-rich work. The weakest involve ambiguous liability, vulnerable customers, incomplete evidence or decisions that are difficult to reverse. In practice, the decisive factor is usually the type of decision being automated, rather than the sophistication of the technology.
The gap between AI adoption rates and artificial intelligence outcomes in insurance industry
Claims generate large volumes of data: policy records, photographs, repair estimates, medical documents, correspondence, call transcripts and historical outcomes. This makes insurance attractive for automation, but it does not mean every part of the claims management process is suitable for AI.
Technologies such as machine learning, natural language processing, optical character recognition and computer vision can process large amounts of information quickly. However, models that perform well on historical data can still struggle in production, where they encounter inconsistent documents, legacy systems, changing regulations and unusual or complex claims.
BCG’s research reflects this gap: the insurance industry has moved quickly on AI adoption, but only 7% of surveyed insurers had successfully scaled their systems.
This is why an AI insurance claims processing business case should start with the process, not the technology. The key questions are:
- What does the process cost today?
- How long does it take?
- How often is human judgement required?
- What is the cost of an incorrect decision?
- Can the output be checked independently?
- Can a human intervene before a consequential decision?
- Can the insurer reconstruct how the recommendation was produced?
The answers determine whether a promising pilot has a realistic path to production.
What claims management processing involves?
The entire claims lifecycle contains several different types of work, with very different automation potential.
1. First notice of loss
AI can extract information from calls, forms, photographs and documents, identify missing information and route claims to the appropriate workflow.
The customer interaction itself is more sensitive. Someone reporting a serious accident, fire or death may need judgement and reassurance rather than a fully automated conversation. AI-assisted FNOL is therefore more defensible than attempting to automate every interaction.
2. Triage
Triage is one of the stronger candidates for AI. Systems can assess complexity, estimate severity, identify missing information, detect potential fraud and prioritise workloads.
This is also where AI adjuster tools can add value without taking responsibility away from the adjuster. Relevant policy information, previous claims and supporting evidence can be brought together before a human makes the decision.
3. Investigation
AI can search documents, summarise evidence, identify inconsistencies and surface similar cases.
The risk increases when the system moves from finding information to deciding what actually happened. Conflicting evidence, disputed liability and litigation require considerably more human oversight.
4. Settlement and payment
For low-value claims with clear coverage, predictable outcomes and complete digital evidence, straight-through automated claims processing can work well.
For complex, high-value or sensitive claims, the same approach can create disproportionate risk. Payment may be easy to automate; establishing that the payment is correct is the difficult part.
Where AI in claims processing delivers measurable results?
The strongest use cases generally combine high volumes, predictable outcomes, good historical data and a manageable cost of error.
Straight-through processing for simple claims
A claim is a strong candidate for automation when:
- coverage is clear;
- the loss is easy to establish;
- the value is relatively low;
- evidence is available digitally; and
- there are no significant fraud or liability concerns.
Lemonade provides a useful public example. Its AI Jim system has handled FNOL, triage, fraud checks and payment, with human escalation where required. In its 2021 annual report, Lemonade reported that AI Jim handled FNOL 98% of the time and took around one-third of claims through the entire process.
These figures are not a universal benchmark. Lemonade operates on a highly digital model. The more transferable lesson is that straight-through processing works best when claim complexity is low.
Document classification and extraction
Document handling is a less visible but potentially more dependable source of claims AI ROI.
AI can classify documents, extract fields, identify missing evidence and summarise correspondence. Technologies such as optical character recognition, natural language processing and machine learning can help turn unstructured data into information that can be searched and used within the claims workflow.
The system does not need to decide whether a claim should be paid. It can reduce the administrative work surrounding that decision. That distinction also reduces risk.
Computer vision for damage assessment
Computer vision has a clear application in motor claims. Aviva announced a partnership with Tractable to use visual AI to assess vehicle damage from photographs and support repair-cost calculations.
This is a constrained problem: identifying visible damage and likely repairs. It is easier to test than a system attempting to determine liability from the same photograph.
Fraud pattern detection
AI fraud detection insurance has particularly strong potential because the system can identify unusual patterns without necessarily making the final adverse decision.
It can flag:
- repeated addresses or bank accounts;
- unusual claim timing;
- inconsistent descriptions;
- duplicate imagery;
- unusual repair patterns;
- connections between apparently unrelated claims; and
- unusual provider or supplier behaviour.
Aviva reported more than 18,400 suspect claims worth £233 million identified across its brands in 2025. It also reported increasing use of AI-generated images and documents by fraudsters.
The value is therefore not simply faster claims handling. It is helping investigators focus their attention where the evidence warrants it.
Not sure which claims use case is worth the pilot?
We help insurers separate the processes AI can safely take on from the ones that still need a person.
That starts with your actual baseline: cycle time, cost per claim, exception rates, and what a wrong decision costs you.
Fraud detection: the use case with the strongest evidence base
Fraud is different from many other claims decisions because the model can operate as an investigative filter.
AI can identify claims that contain signals warranting further investigation, helping investigators focus their attention where it is most needed. This creates a clear separation between anomaly detection and customer decisioning.
Lemonade, for example, has stated that potentially fraudulent claims can be flagged for human investigation rather than automatically rejected.
For insurers, the operating principle is straightforward:
Use AI to help investigators find the right claims, rather than use AI to remove investigators from the process.
This approach can improve fraud detection while keeping the final adverse decision with a qualified human.
Where (and why) AI consistently underperforms in automated claims processing?
AI tends to perform poorly when the problem is not prediction but interpretation.
A recurring failure pattern is that the last 5–10% of accuracy is often the most difficult and most consequential part.
A system may classify 95% of documents correctly. The remaining 5% may contain handwritten amendments, unusual policy endorsements, contradictory evidence, foreign-language documents, litigation correspondence or precisely the information that changes the outcome.
Complex liability disputes
Commercial and third-party claims can involve conflicting accounts, contractual obligations, exclusions, expert evidence and litigation.
AI can accelerate evidence review. Giving it responsibility for determining liability is much harder to justify because the cost of a wrong decision can be financial, legal and reputational.
The issue is therefore decision accountability, not just model accuracy.
Emotionally sensitive FNOL
A claimant dealing with a serious accident, house fire or death may need a person who can recognise distress and explain what happens next.
Automation can still collect information and check completeness in the background. But reducing contact cost should not be the only measure of success.
The long tail of document processing
A document model can perform very well on common formats while failing on exceptions: poor scans, handwritten amendments, unusual endorsements, contradictory evidence or unusual contractual language.
This makes headline automation rates potentially misleading.
If 98% of documents are processed automatically but the remaining 2% contain the highest-risk cases, the economics change. The relevant measure is therefore safe automation plus the cost of exceptions, not automation percentage alone.
Clinical and coverage adjudication
Health insurance presents a particularly difficult case because statistical prediction can conflict with individual clinical judgement.
The controversy around UnitedHealth’s nH Predict system illustrates the exposure. Plaintiffs alleged that the system had an error rate approaching 90%; those allegations are disputed and should not be treated as an established finding. The case nevertheless demonstrates the risks of using automated prediction in high-stakes claims decisions.
Cigna has faced similar litigation over allegations concerning its PxDx system and automated claim-denial processes; Cigna disputed the characterisation of its system.
The common lesson is simple: speed does not compensate for poor decision quality.
The regulatory and legal exposure that most AI deployments are not priced for
The cost of a failed AI deployment can extend far beyond the technology itself. It can include remediation, appeals, regulatory investigations, litigation, compensation, reputational damage and the cost of rebuilding human processes after automation.
The U.S. National Association of Insurance Commissioners (NAIC) adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers in December 2023. It states that consumer-impacting decisions made or supported by AI must comply with applicable insurance laws, including rules on unfair trade practices and discrimination. It also sets expectations around governance, risk management and documentation.
The practical consequence is that insurers need to be able to show:
- what data the system uses;
- how it was trained or configured;
- what role AI played in the decision;
- how accuracy and potential bias were tested;
- where human oversight applies;
- how model changes and drift are monitored;
- what happens when confidence is low;
- how customers can challenge an outcome; and
- how a decision can be reconstructed later.
The NAIC is also developing a regulatory framework around third-party data and models. Using an external vendor therefore does not remove the insurer’s accountability.
Auditability is a production requirement, not a reporting exercise.
How to evaluate an AI powered claims use case before you commit?
A practical evaluation can be reduced to four areas.
1. Establish the baseline
Measure:
- handling and cycle time;
- cost per claim;
- manual intervention;
- straight-through processing;
- leakage and rework;
- escalation and complaint rates;
- fraud detection;
- false positives and false negatives;
- settlement accuracy; and
- employee time spent on administration.
Then define exactly which metric the system is expected to improve. That is the basis for a credible claims AI ROI calculation.
2. Test data readiness
Before a pilot, establish whether:
- historical data is complete and consistently coded;
- outcomes are reliably labelled;
- production data is accessible;
- rare events are adequately represented;
- sensitive data can legally be used; and
- outputs can connect to existing claims systems.
Good historical data does not guarantee good production performance.
3. Risk-tier the use case
A simple three-level model works:
Tier 1 — Assist: summarisation, extraction, classification, search and knowledge retrieval.
Tier 2 — Recommend: fraud prioritisation, severity scoring, routing and other recommendations where a qualified employee remains accountable.
Tier 3 — Decide: automated denial, coverage adjudication or complex autonomous settlement.
Most insurers should prioritise Tier 1 and selected Tier 2 applications. Tier 3 requires substantially stronger evidence and controls.
4. Define the NO-GO criteria before the pilot
A NO-GO decision is a risk-control mechanism, not a failed innovation project.
Stop or redesign the use case when:
- the baseline cannot be measured;
- data quality is inadequate;
- performance deteriorates for important claim segments;
- error costs outweigh the expected benefit;
- humans cannot challenge the output;
- vendor documentation is insufficient;
- audit trails are inadequate; or
- the business case depends on eliminating human oversight that production governance will still require.
The business case should reflect the actual operating model, including integration, governance and human-review costs, not the assumptions used in a vendor demonstration.
The decision-maker checklist
Before approving an AI claims initiative, leadership should be able to answer yes to these questions:
- Is the problem defined in operational terms?
- Do we have a reliable baseline?
- Is the process repetitive enough to justify automation?
- Is the claims data sufficient, representative and legally usable?
- Is the system assisting, recommending or deciding?
- Is there clear human accountability for consequential decisions?
- Can outputs be tested across relevant claim segments?
- Have false positives and false negatives been measured?
- Can material decisions be reconstructed through an audit trail?
- Does the vendor provide adequate model, data and testing documentation?
- Can the system fail safely when confidence is low?
- Have claims, legal and compliance teams reviewed the use case?
- Does the ROI calculation include governance and human-review costs?
- Are NO-GO criteria defined before production?
Build the business case before you build the system
Most AI claims projects fail on data, integration and governance, not on the model.
We work with insurers to test data readiness, risk-tier the use case and cost the human oversight that production will still require.
The real opportunity is not autonomous claims. It is controlled claims automation.
The strongest evidence supports a selective approach to AI in claims processing.
AI is well suited to repetitive classification, document handling, triage, fraud detection, damage assessment and straightforward claims with clear evidence. It is less dependable where outcomes depend on disputed liability, vulnerable customers, unusual evidence, clinical judgement or complex coverage interpretation.
That does not make these difficult areas off-limits. It changes the role technology should play.
For many insurers, the strongest model is AI around the adjuster: finding information, detecting anomalies, preparing evidence and prioritising work while people retain responsibility for consequential decisions.
The value comes from removing avoidable work without removing accountability. In other words, the goal is not to automate insurance claims management at any cost, but to use AI where it can strengthen claims operations, improve consistency and support better risk assessment without compromising human judgement.
AI works best in claims when it reduces operational friction, not when it replaces accountability.
