WIXLIB

The FourthGeneration:PredictiveanalyticsFourth generation solutionsintroduce advanced analytics.This is a giant leap forward. All riskbased decisions hinge on makingconnections between various datapoints and identifying patterns withinthem; however, these patterns aren’talways immediately discernibleby humans. Advanced analyticscan reveal these previously hiddenconnections and relationships with anexceptional degree of accuracy.By using AI and machine learningtechniques, insurers can, for example,identify which data-points in theapplication process are irrelevantor simply redundant, and whichapplications are now suitablefor immediate straight-throughprocessing (STP). The system canassess risk profiles and, in certaincircumstances and with appropriaterisk management, the number ofquestions asked can be significantlyreduced with little to no impact on risktaken on by the life insurer. Certainparts of the application process maysafely be eliminated altogether.To illustrate the impact that analyticscan have, consider the case of apotential customer aged 25-35,with an early form of diabetes. It’san unusual condition, but commonenough that there is plenty of dataabout other individuals with the samecondition in the same age group.That data shows this group is still alow-risk category with an average lifeexpectancy.The data also shows that this categoryof customer is usually approved forstandard life cover once all personalevidence has been reviewed. In otherwords, insurers require intrusivetests, ask multiple questions, and addroadblocks to the process when thecustomer will be accepted anyway.Diabetes is an obvious flag for lifeinsurers, and the request for medicaltests and evidence feels correct.Human judgements have been builtinto the system. However, whatappears to be a common-sensedecision is not necessarily the correctone. With AI and machine learningtechniques, insurers gain a deeperunderstanding of these kind ofcounter-intuitive realities as well asthe effects – past and present – of theirchosen approach.“ Use predictive modelsto eliminate questions,reduce variables, changerules and secure anoptimal result.”7Augmented Automated Underwriting: The next generation of underwriting

The FifthGeneration:AAUbecomesbusinessas-usualMany insurers have embarked onadvanced analytics initiatives, but facechallenges in effectively deploying themodels developed. Fifth-generationsystems leverage advanced analyticalmodelling by fully integrating predictivemodels into existing underwriting andnew-business processes. This is whatwe mean by augmented automatedunderwriting (AAU).To enable underwriters and datascientists to quickly and effectivelydeploy, integrate, monitor and evolvemodels over time, such that modelmanagement becomes business asusual, is a major challenge.Technology that can predict anapplicant’s risk with an acceptablelevel of accuracy significantly improvesthe consumer experience and reducesonboarding barriers. AAU therefore hasthe potential to draw more people intothe life insurance application process,and then keep them in that process forlonger, to the benefit of customersand insurers alike.Importantly, the adoption of this newtechnology does not require a ‘rip andreplace’ approach. Instead, insurerscan choose their preferred models,use existing rules-developmentcapability to define interactionswith those models, and incorporatethose models by reconfiguring theirexisting underwriting processes. Thefinal underwriting decision for lifeinsurance applications may thereforeuse a model as an input to support andaugment the rules-engine decision.Models can be selected and deployedto achieve one of a number of goals,including but not limited to: Streamlining theunderwriting process Transforming the consumerexperience Reducing manual underwriting andmedical evidence costsThe degree to which the modelresponse is used to determine thefinal underwriting decision will, ofcourse, vary by insurer, use case, dataavailability and model confidence,as well as geographic locationand regulation.For example, one insurance firm inNorth America is in the process oftransitioning to predictive modellingrather than opting for an overnightswitch. Its use of predictive modelsallows individual underwriters tocompare their decision-makingwith the model, and highlightwhere human and model are insync to build confidence in themodel itself as it increasinglyinfluences decision-making.On the other hand, two life insurersin Southeast Asia transitioned fullyto predictive underwriting withmodels and technology developedby Munich Re. Improving turn-around times Increasing STP rates Decreasing the number ofquestions asked Identifying fraudulent behaviour Implementing up- and crossselling campaigns8Augmented Automated Underwriting: The next generation of underwriting

The FifthGeneration:AAU becomesbusinessas-usualThis generation is not just aboutmodel development, however.Creating a product from the modelhas knock-on implications, including:understanding and quantifying therisk it may imply; quantifying anypricing implications; consideringthe explainability of the model; andobtaining executive sign-off before itgoes into production.It also requires that underwriters,operations staff and technologistswork closely with actuaries, datascientists and data engineers in thedevelopment and sign-off of the modeland its intended deployment. In manyinstances, insurers will want to sharethe risk of the model as it is eventuallydeployed, because it is such adeparture from traditional approaches.By deploying AAU, insurance firmshave already achieved significantimprovements to their STP ratesand costs, but with only negligibleincreases to their own risk. Someexamples from Munich Re’sexperience: Company A reduced the amountof medical evidence required ofits potential customers, shortenedits questionnaire significantly,and decreased the amount ofmanual processing undertaken byits underwriters, which increasedits STP rate from 30% to 70% Company B pursues a stringentunderwriting philosophy tosafeguard profits but stillincreased its STP rate from 35%to 60%, decreasing its manualunderwriting requirements andthe associated costs, as wellas minimising the amount ofmedical evidence demanded ofapplicants Company C increased its STPrate from an already impressive60% to 80%, again decreasingits manual underwritingrequirements and the associatedcosts, as well as minimising theamount of medical evidencedemanded of applicants Despite eliminating 80% of itsprevious underwriting questions,Company D was able to offerstandard rates to 50% of itsapplicants with only a minimalprice increase“ Model selection andseamless integration forstreamlined applicationsand accurate offers.”9Augmented Automated Underwriting: The next generation of underwriting

Part II:Questionsand answers10Augmented Automated Underwriting: The next generation of underwriting

Regions andregulationsThe fifth generation can thereforebe seen as the next stage in theevolution of insurance technology,streamlining the customer experiencewhile maintaining or even enhancingthe life insurer’s confidence in its riskevaluation processes.While all paths may lead to AAU, thejourneys are not identical. So far, wehave discussed AAU as a generalconcept. However, there are regionaldifferences which will determine howunderwriters will effectively transitionto AAU.If we consider regulatory jurisdictionsas a spectrum, the European model isat one end, while the general US model– allowing for variations within the 50states – is at the other.In general, the US market is fact-basedand relies on extensive proof to validatethe information that is provided. Thiscreates a complex, parallel and multistage underwriting process, whichrequires detailed questionnaires andmultiple data sources and leads tocomplex outcomes: preferred plus;preferred; standard; rated; declined;or postponed.11Automation is therefore also complex,with plenty of manual processes stillrequired. For example, the demandfrom regulators in New York andCalifornia that underwriting rules beregistered, places serious constraintson an underwriter’s ability to updatethose rules in response to changingdata patterns. In addition, thecontestability period of 24 monthsmeans the burden of proof remains onthe underwriter when decisions arechallenged. The primary goal for USunderwriters, therefore, is to accelerateand simplify this process.In contrast, most countries in Europeapply the principle of utmost goodfaith. Customers provide informationto insurers, who take it on trust whenassessing risk. The process is typicallysimpler and more linear: interview,decision, additional evidence, andfinal decision. The complexity of theunderwriting questionnaire dependson a number of factors, includingproduct, distribution channel,engagement model, pricing, andmarketing practice, but commonlyaccepted outcomes are standard;rated; declined; or postponed.The level of automation alreadyin place also depends on a numberof factors, ranging from marketmaturity and availability of structureddata, to the nature of the underwritingprocess itself.Augmented Automated Underwriting: The next generation of underwriting

Regions andregulationsIn Europe, the burden of proof liesmore with the customer than theunderwriter; as a result, contestabilityis usually shorter and more swiftlyaddressed. The goal here is to usedata models in a way that improvesunderwriting accuracy, minimises falsepositives and reduces ‘spikes’ in risk.The APAC region is different again.Japan and other countries in SouthEast Asia have similarly high levels ofautomation as their peers in Europe,while the level of automation inAustralia is still comparatively low.The journeys to greater automationand augmented automation withinthe region will therefore take slightlydifferent routes.Meanwhile, there are plenty ofcountries that do not have readilyavailable, coherent records.However, fifth-generationtechnology is effectively agnostic.It can incorporate an underwriter’sfundamental philosophy, bring inmodels that support that, and pulldata in from other sources if necessary.As a neutral data hub, it allows fordisparate data sources and formatsto be incorporated, from which AAUcan be achieved.The nature of medical records in aparticular region also has an impact onthe adoption of AAU. As an example,UK health files are digitised throughoutthe National Health Service (NHS). Thechallenge is that family doctors havetraditionally acted as a gatekeeperof those records and were paid byreinsurers to provide them. As a result,there is some resistance to change,which can slow progress.12Augmented Automated Underwriting: The next generation of underwriting

ExplainabilityandtransparencyDespite regional differences, thereare two universally recognisedrequirements for AAU: explainabilityand transparency. These key factorsoften differentiate life insurance fromother financial products.AAU places underwriting in the realmof complex models that leveragemachine-learning techniques.Insurers will therefore be increasinglyrequired to use expert judgementwhen selecting models for use inAAU, since they are the means bywhich they achieve their fundamentalunderwriting philosophy. However,insurers will still need to prove toregulators and ombudsmen that thechosen approach is sound and thedecision-making is valid. This raisestwo key questions:1. Philosophically, how does humanintelligence understand anddescribe even the earliest formsof artificial intelligence?2. Practically, how does an underwriterexplain in a court of law thata machine algorithm made thecorrect decision?13Transparency and explainability haveto be built into both the model andthe model-selection process. GUIs,easily configurable rules, transparency,reporting, and now advanced analyticsare not unique to life insurance, andwe have seen similar moves towardsgreater automation and machinelearning in areas like fraud preventionand detection. Yet the life sectorhas rejected some of the automateddecision-making seen in otherareas precisely because of its‘explainability’ requirements.Explainability is therefore one of thekey barriers to adoption. However, itis worth noting that the demand forexplainability has not always been metby first-, second- or third-generationtechnologies. Explaining the neteffect of thousands of rules is noeasy feat, and regulators have oftenexpressed interest in avoiding unfairdiscrimination against customers by‘black-box’ models.Fortunately, many of the latestmachine-learning models havebecome more ‘interpretable’ thanearlier iterations. As a result, it ispossible to explain decision-makingat an aggregate level and to drilldown into what drives a singleprediction for a specific customer.The gap between an explainablemodel and explaining the impactof 1,000 variables may not be asinsurmountable as many believe.Augmented Automated Underwriting: The next generation of underwriting

Models andcomponentsAs indicated above, achievingAAU is not about implementing asingle, monolithic system. Multiplecomponents, or modules, areinvolved. Claims engines, machinelearning algorithms, monitoringdashboards, underwriting rulesets, data integration, and naturallanguage processing (NLP) all formpart of the picture, as does a scalablearchitecture enabled by cloudcapabilities and APIs.Individually, each of these componentparts offers very little value. It is onlywhen they are integrated and overlaidwith in-depth knowledge of the lifeinsurance sector – its history, itsregulations, its requirements and itsbroader societal role – that they canbe transformed into sources ofbusiness value such as point-of-salepredictive models.The process of selecting the modelsthemselves must also be considered.Companies have to take into accountthe trade-offs between modeleffectiveness and performance,complexity and explainability, and findthe right balance for their organisationin its market and regulatory context.All model developers – whetherinternal or external – must buildtransparency into their own processes,from data accessibility and dataanalysis, to model integration,deployment and monitoring.Developers should also strive to avoidthe syntax and jargon of data science,so that statistical methodology canbe adequately communicated tonon-experts.Currently, independent companiesand AI consultancies are developingmodels for insurance and otherpurposes. Some are based on publiclyavailable data, while others are basedon insurer-specific data. However,models that are built using only publicdata generally cover only one pieceof the puzzle that is a human life tobe insured.As a result, models from within theinsurance business offer an advantage.Many insurers therefore may chooseto work with models developed byproviders with insurance, underwriting,data and actuarial experts who havethe ability to explain the implicationsto risk, compliance and actuarialcolleagues and get the modelssigned off.This is a big leap to make for therisk-averse insurance sector andrepresents a profound change in theway insurance works. That change willbe required at almost every level.Capital, modelling andtechnology expertiseAs one of the world’s largestreinsurers, Munich Re takes a risksharing approach when developingmodels for clients with the capitalto support it. It has business unitsthat are dedicated to developingmodels, which can be integrated intoour underwriting engine alongsidemodels from third parties. However,when Munich Re models are used,tested and validated, and all partiesare convinced that the model willmake the right decisions, it isprepared to take a significant portionof the risk through a reinsurancearrangement. This ‘skin in thegame’ is a significant advantagereinsurers have in the development of14Augmented Automated Underwriting: The next generation of underwritingmodels, unlike, for example, pure AIconsultancies. It requires reinsurersto provide utmost diligence in themodel development as any modeldeployment has a significant financialimpact if its risks implications arenot properly assessed by actuarial,claims, underwriting and data scienceexperts.

Managingthe datadelugeAlthough there are challengesin moving to AAU, there are alsopowerful reasons for overcomingthose challenges. There is also asense of inevitability surroundingits adoption. The onward march oftechnological capability, coupled withthe rapid evolution of data science,has had a significant impact on theinsurance industry as a whole, andis likely to do the same for the lifeinsurance sector. The choice to adoptAAU may, in the end, not be much of achoice at all.For example, new sources ofpertinent health data are becomingreadily available. Electronic healthrecords (EHRs), for example, canprovide structured information thatcould be readily integrated intomodel development, all withoutinconveniencing customers orphysicians. In addition, today’swearable fitness trackers are compilingcomplementary data sets that mayprove valuable to underwriters. Theinternal data that insurance companiesalready hold can also be integratedinto models and mined for previouslyhidden insights.These additional sources ofinformation raise questions about whatdata could be used in the future. Whenconsidering potential data sources, wemust ask: Will the regulators allow it? Will it truly add value?Of course, the answers are notalways easy to come by. Considerthe following: The technology behind facialrecognition, although not sufficientlyaccurate at present, could eventuallybe used to identify a smoker who isreluctant to admit their habit. Is thisadmissible? Genetic data, increasingly availableas DIY DNA and ancestry kits growin popularity, could provide helpfulrisk information. No regulatorcurrently allows them to be usedextensively, and they are unlikely todo so in the near future. But couldregulators at some point overcomethis reluctance? Personal lifestyle informationabout hobbies, education, diet andother factors is readily availablethrough social media accounts. Thisinformation could be incorporatedinto models, but will it proveproblematic when it comes tojustifying decision-making? On arelated note, should models minesocial media as a whole for usefultrend analysis and broader societalcontext, or should this too beinadmissible?These examples provide insight intosome of the questions that will beasked in the near future. Similarly, theasymmetry of information betweenthe applicant and the insurer will haveto be addressed by regulators and theinsurance industry. Will it undermine principles of eitherconsent or explainability?15Augmented Automated Underwriting: The next generation of underwriting

Data scienceand itstechnologyIn addition to the ethical questionsaround data use, there is the morepractical issue of whether data isreadily usable. Structured formatsare needed for model analysis anddetails tend to be gathered in formalprocesses, which necessarily excludesa vast amount of information. Theprocess of gathering structureddata is also prone to distortion, ascustomers interpret and filter theirown personal information to suit theirself-perception.Data science is also in the process offinding better ways to convert moreinformal or analogue informationinto structured digital formats. NLPis one method that converts thecomplexity of normal conversationinto more standardised, readableand analysis-friendly data. Equally,the insight gained from data analysiscan itself help to structure questionsmore effectively and to pre-populateforms in a way that provides the mostuseful information. Analytics can alsoidentify where data could and shoul

augmented automated underwriting (AAU) may also be described as augmented underwriting with AI, accelerated underwriting, or predictive underwriting, among other terms. For the sake of simplicity, we use AAU in this paper, in which we examine: How AAU builds on and evolves previous technology investments and implementations