Machine Learning for

Insurance Business
Using Data to Foster Innovation
Table of contents Get in touch
Andriy Zakharchuk
Executive summary 03
CEO @ Intelliarts

Data as the untapped [email protected]

resource in insurance 05 andriyzakharchuk

Unlocking the power of

insurance data with ML 07
Volodymyr Mudryi
DS / ML Engineer @ Intelliarts
Use cases of machine
learning in insurance 10 [email protected]

volodymyr-mudryi

Path to ML implementation 19

Create the most eicient Yurii Laba

ML-powered solutions with DS / ML Engineer @ Intelliarts
Intelliarts 23
[email protected]

Intelliarts Case: ML-powered yura-laba

solution to improve lead

quality and agent eiciency 24
Alexander Barinov
Looking ahead: Managing Partner @ Intelliarts
Transforming your insurance
business with ML 26 [email protected]

alexander-barinov
About us 28

Oleksandr Stefanovskyi
Head of R&D @ Intelliarts

[email protected]

oleksandr-stefanovskyi

02
Executive summary
The most critical asset in insurance
Data has always played a central role in the insurance industry. Today insurers have access to
more data than ever before — in the past few years, we have produced twice as much data as
decades before.

Despite this, a recent study1 revealed that most insurers are struggling to maximize the benefits
of data. They process not over 10-15% of the data they have access to. Moreover, most of this
data is structured and housed in traditional databases. This means insurance companies fail to
benefit from all structured data available to them while also ignoring the value of semi-structured
and unstructured data.

Analysis of this unstructured data, as well as better use of structured one, requires the
introduction of advanced data science techniques. Innovative data analytics technologies
centered on artificial intelligence (AI) and machine learning (ML) can bring order and purpose to
insurance data and help an insurer drive better business decisions.

Machine learning to increase the value of insurance data

The 2021 report published by GlobalData2 informs that 80% of insurance executives expect
AI/ML to play a central role in the digital transformation of their company. Between 2019 and
2024, the AI/ML platform revenues within the insurance sector are expected to grow by 23%,
reaching $3.4 billion, according to the report.

AI/ML in Insurance

80% 47% 56%

of insurers plan to implement of insurers project the of insurers believe AI/ML

AI/ML in their digital growth in AI/ML investments would improve their
transformation in insurance over the next operational efficiency over
year the next three years

Source: 2021 Global Data Report

03
Executive summary

This doesn’t surprise, though. AI/ML technology is widely known for its potential to bring greater
efficiency and innovation across the entire insurance value chain. The potential applications of
ML are plentiful, from better risk assessment to faster claims processing and more accurate
fraud detection. ML is also useful for improving the insurer’s customer service with efficient lead
management, more personalized marketing offers, and customer churn prediction.

Together, these improvements in business operations allow insurers to enhance productivity and
generate better revenue streams. Check the survey results below that compare the revenue
distribution of traditional insurance organizations vs. so-called data master insurance
organizations with the technical capabilities, culture, and behaviors to support their data-driven
programs, which often rely on ML/AI technology.

Revenue distribution of traditional Revenue distribution of data master

insurance organizations insurance organizations

$20-$50 $50 billion $50 billion $1-$5

billion or more or more billion
14% 7% 27% 5%

$10-$20 $1-$5 $5-$10

billion billion billion
10% 57% 8%

$20-$50 $10-$20
$5-$10
billion billion
billion
51% 8%
12%

Source: Capgemini Research Institute, Data Mastery in Insurance survey, September-October 2021, N=204 insurance
organizations with N=37 data masters.

As we observe a global push toward digital insurance, this white paper will
help you understand the potential gains of introducing ML into your
business operations.

The role of data in the insurance industry

Specifically, Use cases of ML in insurance
we discuss: Steps of ML implementation
And how to create the most eicient ML solutions

Let’s dive deeper into how ML is contributing to the insurance sector!

04
Data as the untapped resource in
insurance
Insurance has always revolved around data. Its traditional data sources ranged from financial
data to actuarial data to historical claims and risk data. These used to affect every single
important decision that an insurer made.

Over the years, the critical role of data in the insurance sector hasn’t changed. With no physical
products to manufacture and sell, data remains one of the most valuable assets in the industry.
Yet, we can boldly say that the nature of insurance data has altered, with data growing in
volumes, variety, and velocity.

The 3 V’s of Big Data in Insurance

Today insurance companies measure the amounts of data they’re dealing

with in gigabytes or even petabytes. Data is being created at an exponential
rate, primarily because of the availability of alternative data sources, such as
Volume IoT devices that are capturing data to send it over the internet. Telematics in
car insurance is one of the greatest examples here, where a device is fitted
into the motor vehicle, and then the insurer can check the veracity of the
claim being made by using IoT data.

The times when data was stored in rows and columns are long gone. Now
data comes in structured, semi-structured, and unstructured formats. And
insurance companies could use any of these to put insight into perspective.
Variety Claims and policy data, as well as consumer information, are well-known
examples of structured data. Insurers also use semi- and unstructured data,
such as real-time weather feeds, public records, social media posts,
geospatial data, cold calling records, and emails.

They say that, estimatedly, the volume of data doubles every two years.
Velocity Insurance data is growing really fast, with businesses aiming to make data
flow as quickly as possible — real-time data is preferred.

Velocity can be more important than volume because it can give us a

bigger competitive advantage. Sometimes it’s beer to have limited data
in real-time than lots of data at a low speed.
MetLife Inc.

05
 Data as the untapped resource in insurance

Insurance data to contribute to If paired with machine learning (ML), data

generates insights that help insurers
insurer’s competitive edge transform business models and improve
relationships with customers. Data also has
At least 47% of insurance businesses consider
the potential to increase business efficiencies
data and analytics a very important factor
by streamlining insurers’ business processes.
when it comes to competitive advantage.3 And
this percentage is projected to grow up to
Unfortunately, not all insurers are able to
58% in the next three years. Today, insurance
leverage and optimize the value of big data.
data can be used for much more than
You might be sitting on volumes of data but
traditional historical analysis.
have no idea how to use them in a more
effective way. Or you might not know about
Here we can mention the development of new
the piles of data accessible to your company
products, marketing and sales, assessment
because your business hasn’t implemented a
and mitigation of risks, optimization of
data strategy and data pipeline yet. Either
premiums, and more. (See many other
way, don’t you think it’s finally the right time
examples of how data is reshaping insurance
to unleash the potential of your insurance
below.)
data?

Data is reshaping insurers’ approach to new markets and old assumptions

Percentage of executives who agree to the below statements or have implemented the below initiatives at partial or
full scale.

With advanced analytics, we are able to 44%

Data-driven product development

enter niche markets or segments deemed 49%

too risky in the past 38%

40%
Developing new features and coverages in 46%
traditional products 33%

Data and analytics are facilitating our 39%

company’s shift from protection to 34%
prevention 43%

Developing emerging insurance solutions 26%

using real-time data (micro-insurance, 30%
parametric insurance, etc.) 24%

We are moving towards deeper 44%

segmentation of consumers based on 47%
39%
data-driven risk pricing

We are able to create new risk-transfer 38%

solutions based on advanced analytics 40%
Risk insights

35%
We continually revisit actuarial
assumptions (such as loss ratios, 35%
39%
emerging-risk scenarios at regional level,
30%
etc.) through data-derived insight

We are capable of modelling new/ 33%

emerging risks based on data from diverse 43%
(internal and external) sources 29%

Source: Capgemini Research Institute, Data Mastery in Overall

Insurance survey, September-October 2021, N=300
Life and Health Insurance
business executives and 210 technology executives
across 204 insurance organizations. Property and Casualty (P&C) Insurance

06
Unlocking the power of insurance
data with ML
As consumer expectations are growing bigger in terms of efficiency and customer service, we
see the shift to data-driven insurance businesses. By unlocking the power of data, ML
technology has a big chance of becoming the frontrunner in the industry.

A brief overview of machine learning

When it comes to ML, we’re basically talking about extracting knowledge from data. According to
IBM,4 ML can be defined as:

A subfield of AI and computer science that focuses on the use of data and algorithms to
imitate how humans learn by gradually improving its accuracy.

Deriving from historical data, an ML algorithm can learn on its own and make predictions and
conclusions without being explicitly programmed. So, an ML model fuels data-driven decision-
making, where an insurer moves from data to insights to decisions. It generates the results as
accurately as possible, but the model is important to be fed with enough quality data.

Data-driven decisions thanks to ML

Data Sources Data Analytics Decision Making

Data Insight Decision

Figure: Moving from data to insights to decisions.

There is a substantial difference between AI and ML, yet the two concepts are frequently
confused in business. As implied by its definition, ML is a branch of artificial intelligence. While
both belong to computer science, AI is a broader concept than ML that is designed to simulate
human thinking capability and behavior.

Another critical concept to mention is deep learning. Alike, it’s a subset of ML that handles
algorithms inspired by the structure and function of the human brain. The major difference
between the two lies in how the data gets presented to the machine.

While traditional ML algorithms rely on structured data, deep learning can work with both
structured and unstructured information working based on the multiple layers of artificial neural
networks.

07
 Unlocking the power of insurance data with ML

Artificial intelligence can deliver on industry expectations through machine

learning and deep learning

Artificial intelligence Machine learning

The science and engineering of
A major approach to realizing Deep learning
making intelligent machines A branch of machine learning
artificial intelligence

1950s 1960s 1970s 1980s 1990s 2000s 2010s

Artificial intelligence (AI) Machine learning (ML) Deep learning (DL)

Intelligence exhibited by machines, Major approach to realizing AI by Branch of ML in which algorithms

whereby machines mimic cognitive learning from and making attempt to model high-level
functions associated with human data-driven predictions based on abstractions in data. DL connects
minds; cognitive functions include data and learned experiences. ML artificial, software-based
all aspects of learning, perceiving, compromises several categories, calculators that approximate the
problem solving, and reasoning. including reinforcement learning, functions of brain neurons. Neural
supervised learning, and networks, formed by these
unsupervised learning. calculators, receive, analyze, and
determine inputs and are informed if
determination is correct.

Source: The History of Artificial Intelligence by Rockwell Anyoha

The value of machine learning in insurance

Since ML has a great potential to unleash the power of insurance data and speed up the tempos
of digital transformation, insurers believe ML is important — and are investing in it. The survey of
the top 100 US insurance companies proves that more than half of insurers (66%) are investing in
this technology now, and nearly three-quarters (74%) are thinking about this in the next three
years.5

Statements of the top 100 US insurance carriers about AI/ML importance

63% 74% 66% 34%

Important Today Important in 3 Years Investing Today Not Investing Yet

Source: 2019 LexisNexis Survey

08
Unlocking the power of insurance data with ML

The key drivers of ML in insurance vary. We can distinguish between the main ones:

Today, connected consumer devices, such as smartphones, smart

Increase in data TVs, or fitness trackers, are becoming increasingly popular. This
explains the growing number of data in insurance. Insurers can use
volumes this data received from IoT devices and evaluate their customers’
profiles more accurately.

With tons of data accumulated in the industry, open-source

Open-source protocols are becoming mainstream to make sure data is shared
and used across. Also, private and public sectors join forces to
everywhere create reliable ecosystems where data is shared safely and
securely.

McKinsey predicts 25% of the insurance industry to be automated

Strong potential by 2025.6 The industry indeed has lots of areas to be automated,
for automation from claims management to policy cancellation. And ML is very
useful when it comes to automation.

The pandemic has taken a great toll on insurance businesses. Still,

those insurers that have incorporated intelligent technologies
Beer response
appeared better prepared for Covid-19. For example, they were able
to Covid-19 to process claims fast and accurately, although a large part of the
workforce worked from home.

The surge in ML has altered expectations around customer

experience and personalization of insurance services. Now
Shift to
customers demand more advisory, faster, and efficient services like
personalization instant quotes, dynamic pricing, and policy offers tailored to their
needs.

So, what are specific ways to unleash the value of insurance data?

09
Use cases Machine learning in insurance

ML is changing the face of competition and urging innovation in insurance.

Data-driven businesses are using data and ML to provide a fully digital,
personalized, customer-centric experience, with automated processes and
more eicient business operations.

Today, we can speak about at least eight use cases of machine learning in
insurance that can bring insurance services to an entirely new level:

01. Lead management

02. Personalized marketing and sales

03. Price optimization

04. Churn prediction

05. Data extraction

06. Intelligent underwriting

07. Fraud detection

08. Claims management

10
Use cases of machine learning in insurance

01. Lead management

Lead data is the greatest trove of knowledge about customer behavior and market opportunities.
Insurers can use this info to extract valuable insights and get a fuller picture of an ideal customer
and the customer journey. This information is also useful for choosing more quality leads and
spending less time on those with potentially low chances of conversion.

Use-case examples:

Lead analysis and classification for Calculating the sales threshold to

improving lead generation increase lead conversion

Improving cold calling effectiveness Lead scoring optimization

See it in action: Lead scoring optimization

Case Study
A midsize insurance company has reached out to us to optimize their resources with better
lead scoring. Since the company got their leads from multiple sources, they wanted to
prioritize them efficiently.

Throughout our cooperation, the Intelliarts team has built a predictive scoring model that could
forecast the probability of how likely the lead would buy a policy. With this system in place,
insurance agents could call leads with high scores first and leave those leads with low scores
for later or skip them entirely in case of a lack of resources.

NewLeads

Your
Data

Lead Scoring

Models Segmentation
Ordering
External Machine Prioritization
Data Learning
Action

The ML-based lead scoring model we’ve built has shown over 90% accuracy and good results
in production. The model divided leads into several groups based on their buying capacity. The
most promising leads had an 80%+ probability of purchasing a policy, and the expected
conversion rate was 3.5 times bigger than on average. The insurer became able to process
leads more efficiently and cut costs, saving time on less profitable prospects.

11
Use cases of machine learning in insurance

02. Personalized marketing and sales

Personalized marketing is another way to reap the full benefits of ML. As said, ML can get
insights from large amounts of data and see patterns in customers’ behaviors, attitudes, and
preferences. An insurer can then use these insights to provide individual offers,
recommendations, loyalty programs, and messages to the end-users.

Besides, insurers can use ML for market segmentation. By analyzing consumer data, such as
income, age, gender, and locations, ML algorithms can classify customers into different groups.
An insurer can then develop specific attitudes or adjust marketing tactics based on these groups
to target different customer segments.

Use-case examples:

Building innovative products and services that would target specific customer
segments

New cross- and up-selling Data analysis to choose the best

opportunities based on customer channels and touchpoints to contact
insights the leads

See it in action: Data-driven marketing

Case Study Example Of Underperformance Group Of Calls

Distribution of success rates of similar calls

The insurtech company has 0.40

contacted us to help them 0.35

research and solve the problem 0.30
Probability Density

of low cold calling performance 0.25

so the company could optimize
0.20
their sales approach based on Success rate of group of calls
0.15
the findings. A data science
approach was used to test the 0.10
0.13% 2.14% 13.59% 34.13% 34.13% 13.59% 2.14% 0.13%
hypothesis of underperforming 0.05

cold calls caused by the 0.00

μ - 4Q μ - 3Q μ - 2Q μ -Q μ μ +Q μ + 2Q μ + 3Q μ + 4Q
company’s phone numbers
being flagged as spam.

Our research proved a zero correlation between the company’s numbers being marked as spam
and the low success rate of the cold calls. Instead, Intelliarts has discovered multiple other
factors affecting the effectiveness of cold calls, such as the insurance agent who was calling
and the way how the calls were managed. The timing of the original and follow-up calls, as well
as the demographic state of the lead, also mattered.

Our data engineers offered the customer to build an ML system that could detect the
underperformance of phone numbers, track the factors causing the underperformance of cold
calls, and avoid such patterns in the future. We created the MVP for this solution that provided
valuable outcomes for the customer and created opportunities for the company to optimize
their cold calling approach. Now we’re implementing this full-scale solution.

12
Use cases of machine learning in insurance

03. Price optimization

ML-based pricing models have been gaining popularity recently in insurance as a great way to
optimize pricing policies. In this scenario, insurers include a variety of variables in the premium.
When defining prices, ML algorithms process more sources and channels (even alternative ones,
such as previous claims or social media). Finally, an ML-based pricing model allows insurers to
adjust premiums dynamically, for example, by reacting timely to market fluctuations.

Use-case examples:

More accurate pricing models, which discern patterns from data, integrate extra
sources and information, and notice trends and demands at the early stages

More dynamic prices that respond to monetary policies and market fluctuations

See it in action: Better pricing

AXA, global insurance company

According to the company's statistics, most of the car insurance policyholders that get into
road traffic accidents cause minor accidents, while only around 1% of accidents are classified
as large-loss cases with huge payouts.7 AXA’s R&D team decided to apply deep learning to
predict if a driver may cause a large-loss case during the insurance period.

The company identified 70 parameters and risk factors, including age range of the driver,
annual insurance premium range, previous accidents, etc. They entered these features into a
neural network and were able to predict large-loss traffic accidents with 78% accuracy.

This approach helped the company to build a strategy to optimize prices for car insurance
policies by offering the possibility of real-time analytics and dynamic pricing at a point of sale.

Features 3 hidden layers Output class

neurons neurons neurons

Driver
Large-loss
accident
PDL Limit

PPA Limit

Insured Age range

Small
… accident
Car Age range

Region

Previous accidents

Source: Schematic visualization of AXA's deep learning model for insurance price optimization

13
Use cases of machine learning in insurance

04. Churn prediction

Detecting early churn signals and recognizing at-risk customers provide insurers with an
exclusive opportunity to prevent customers from leaving the company. With an ML model, a
company can track at-risk customers based on their behavioral patterns and predict the churn
rate. As a result, the insurer interferes early and prevents reductions in revenue.

Related reading: Using ML for Customer Churn Prediction in Insurance

Use-case examples:

Efficient churn prediction based on Discount offers to at-risk policy

customer data keepers to gain back their trust

See it in action: Efficient prediction of churn rate

Swiss Life

In their research8, Swiss Life insurance company investigated opportunities to predict churn in
terms of a customer buying back their life insurance.

In order to accomplish this task, the research team used an anonymized excerpt of the
historical data about more than 150 000 customers and their past or ongoing contracts,
including life insurance, pension insurance, health insurance, etc.

Initial modeling has shown that just using the customer's description data does not entail the
relevant information for predicting early contract termination. So the research team has
incorporated additional time-series data into the training dataset, including data about
changes in the policy or its components related to buying a house, marriage, birth of a child,
etc. They also used the TF/IDF representation for compiling time-related features of the data
set, as it allowed them to gain significant improvement in terms of ML model's prediction
accuracy.

As a result, ML solution built on top of the SVM algorithm has shown striking prediction
accuracy of 99.7%. This allows the company to discover customers with a high risk of churn
and take marketing actions to improve customer relationships.

Looking for a technology consulting partner to

implement a machine learning project and
become closer to digital insurance?
Contact us
Share your desired business outcomes, and
together we will produce the best solution for
your insurance company.

14
Use cases of machine learning in insurance

05. Data extraction

Another interesting use case of ML is document-based digitalization with optical character
recognition (OCR). This technology allows recognizing handwritten and printed texts, which
helps insurers process documents faster and improve operational efficiency.

ML-powered data extraction allows insurers to forget about manual re-typing. When paired with
computer vision, it can also accurately render every single pixel and translate the info.

Physical Documents Data extraction Structured or semi-structured digital data

algorithms

Related reading: ML-Based Data Extraction in Insurance

Use-case examples:

Simple physical document processing Reduced manual data extraction, which

by extracting information seamlessly leads to data accuracy and integrity

Faster customer onboarding and claims settlement since the information about the
customer is extracted automatically

See it in action: Automatic data extraction

The Hanover Insurance Group

A data extraction tool has become a real finding for the Hanover Insurance Group to
accelerate their daily legal and insurance document intake and processing. The solution that
the company’s using combines computer vision, natural language processing, and ML models
to handle complex unstructured reports widely used in the insurance sector.

Tokio Marine
Tokio Marine uses an ML-based OCR service to handle claims. The tool is cloud-based and
helps the company to process handwritten claims by “reading” the documents and submitting
them to the existing system. Its recognition rate is over 90%, which allowed the company to
reduce human error by 80% and processing time by 50%.1

15
Use cases of machine learning in insurance

06. Intelligent underwriting

In insurance, traditional manual underwriting is associated with piles of physical and digital
paperwork, back-and-forth communication between clients and agents, and entering and
rechecking lots of info submitted into the system. And we don’t even mention the cumbersome
and time-consuming process of assessing clients’ risk to decide whether to issue the contract
and under which conditions.

ML-based intelligent underwriting can help insurers better analyze risks, provide more accurate
pricing, and process applications faster. With ML and deep learning, the underwriting process
time is reduced to a few minutes (instead of hours or days) because most activities are
automated and data-driven.

Related reading: Using Machine Learning to Increase Underwriting Efficiency

Use-case examples:

A more accurate risk assessment by Access to extra data sources (social

checking an applicant’s profile against media, bank information, or
loads of data third-party) to improve risk selection

Prediction of potential failure rate and other operational risks

Optimized premium rates based on geospatial information, projected real-estate market

fluctuations, geopolitical risks, weather and climate forecasting, and IoT

Improved coverage recommendations that align with customer needs and risk profiles

See it in action: Faster and more intelligent underwriting

Progressive Insurance
Progressive Insurance is using telematics to collect data from drivers and improve the accuracy
of risk assessment. For example, the car insurer buys the driving data from the Snapshot
mobile app, which provides the insurer with billions of data monthly. Based on this, the insurer
can assess risks better and offer discounts to less risky drivers.

Midwest Life Insurance

Midwest Life Insurance has introduced intelligent underwriting to identify which applications
were worth quoting faster. The company has analyzed attending physician statements and
internal and customer information. In a few months, they noticed the first results: fewer errors,
better underwriters’ productivity, and reduced turnaround times.

16
Use cases of machine learning in insurance

07. Fraud detection

The Insurance Information Institute informs about $38 to $83 billion in annual losses related to
insurance fraud.9 And this is excluding health insurance fraud, which costs the US nation an
extra $68 billion and is the most costly type of fraud.

Since ML algorithms work great for anomaly detection and classification of large datasets,
machine learning is a good fit for fraud detection and prevention. An ML system seeks patterns
and analyzes consumers’ behaviors. As soon as it detects any abnormal activity, it warns the
insurer immediately.

How an ML fraud detection system works

Block

Review

Сustomer files an ML-based system generates Model predicts Allow

insurance claim customer features fraud risk score

Use-case examples:

Easy detection of inconsistencies in the submitted claims-related files to find fake and
duplicate claims

Use of alternative data sources, such as public data or IoT, to improve fraud detection
results

Image recognition as an additional filter for fraud detection at the submission stage

See it in action: More accurate fraud detection

Anadolu Sigorta
The Turkish insurance company10 tells about their ROI improvement by 210% in one year after
they switched to an ML-based predictive fraud detection system. The ML model was
developed based on the 3-year historical claims data provided by the insurer plus enhanced
with 260 business rules for fraud scoring. ML introduction has saved the company $5.7
million.

ZhongAn Technology
ZhongAn Technology, a Chinese insurer, has employed image recognition to mitigate claims
fraud. The company asks users to upload a photo of their cracked phone screen. Then ML
algorithms help the insurer understand whether the actual damage took place.

17
Use cases of machine learning in insurance

08. Claims management

By implementing ML, insurers can get a unique advantage of how they handle claims — in a
faster, more efficient, and more accurate way.

The case study by Insurance Thought Leadership11 reveals that claims settlement usually takes
around 11.5 minutes, from submitting the receipt to updating or closing the record. With an ML
model in place, this time shortens to three minutes. This means a 74% reduction in time needed
for insurance claims management and an 80% increase in productivity. The same case study
confirms that ML can manage claims with 80% of accuracy. The only thing that affects accuracy
refers to database limitations and spelling errors. However, since an ML solution tends to “learn”
over time, insurers should expect fewer errors in the long run.

ML-powered claims management process

1 2 3 4

Digital claim Auto evaluation and Automated document workflow Insurance

submission cost estimation guided by machine learning system payment

Related reading: Automated Claims Processing with Machine Learning

Use-case examples:

Smart claims triage to identify more

Claims volume forecast to reduce the
complex claims that need extra
overall claims settlement time
review

Easy detection of inconsistencies in Use of alternative data sources, such

the submitted claims-related files to as public data or IoT, to improve fraud
find fake and duplicate claims detection results

See it in action: Automatic, error-free claims processing

Grinnell Mutual
Grinnell Mutual uses computer vision to assess damage in property claims. The company
offers aerial intelligence for roof inspection instead of involving many parties (insurance
agents and roofing contractors). An insurance agent then uploads the pictures from drones to
an ML-based system and gets the claim processed in a few days rather than spending weeks
or months for damage estimation.

Fukoku Mutual Life

Fukoku Mutual Life handles claims data with the help of deep learning. Technology helps the
insurer automatically find and access medical documents related to the case, as well as
calculate the pay-offs. As a result, the Japanese insurer can boast of a 30% increase in
productivity and cost savings of around $1 million a year.12
18
Path to ML implementation
Obviously, ML and predictive models are worth integrating into your insurance business.
Regardless of the ML solution you plan to build (e.g. it could be a fraud detection model, churn
prediction, a claims management system, or a pricing solution), the key steps for implementing
an ML-based insurance project are generally the same:

1. Goal definition
Before an insurer moves forward with building an ML solution, they should understand the
existing problem and define the value they’re expecting to get from the solution, ideally including
the measurement criteria. Data scientists that the company will work with will need this
information to understand what type of ML problem they are going to solve: classification,
regression, clustering, or the combination of any:

Defines to which class or category a data point belongs to. For

example, if we’re building a churn prediction model, an ML algorithm
Classification will be trained to divide customers into churners and non-churners
and answer questions like “Will the customer leave the company?”
“Renew the insurance policy?” “Downgrade it?”

Aims to evaluate the relationship between the variable and the data
values that influence it. In simple words, the outcome of regression
Regression always includes a number. Taking the same example with a churn
model, it could be the period when the customer is projected to leave
the company.

Groups the unlabelled dataset. Let’s say the insurer has no idea about
the characteristics that define customers that are about to leave the
Clustering company. The algorithm can divide the customers into clusters
(churners and non-churners) on its own.

Classification & Regression & Clustering

Classification Regression Clustering

19
Path to ML implementation

2. Data collection
The next step will be to decide what sources to use for data collection and actually gather this
data. In ML, data matters a lot, and the quality and relevance of the insurer’s data will directly
affect the results that the ML models produce.

Most insurers would use historical datasets of their insurance claims as a backbone of data. But
you can also think about extra data sources, such as public records, demographic and geospatial
data, customer reviews, and so on. Although the general rule states “the more data, the better,” it
is also important to make sure the quality of data is good — the data science team will gladly
help you with this.

Even if the company doesn’t have enough historical data, they shouldn’t worry. Experienced data
scientists can help insurers with data collection, for example, by producing synthetic data or
advising on data collection, data storage, and labeling.

3. Data preparation
Data scientists will also need to prepare the raw data and convert it into a suitable format for an
ML system to digest. They should structure data points according to the same logic and get rid
of any inconsistencies. Pre-processing could include:

Drops out the most discriminative and irrelevant information

Feature extraction and, thus, shortens the number of variables, i.e. attributes that
impact the results.

Determines a set of features that describe raw data and that

Feature engineering ML algorithms will base on in their prediction results.

Feature selection Revises the extracted features and chooses

Feature selection only those that correlate with the variables that the company
wants to predict.

All Features

Feature Selection

Final Features

20
Path to ML implementation

4. Modeling and testing

At this stage, the team moves to the actual development of a predictive model. Different ML
algorithms are used to build models in insurance. Think about ML algorithms as the engines of
ML, helping to turn a dataset into a model. For example, its idea is to use the insurer’s data
described by labels and features and learn to make conclusions, e.g. fraud vs. not a fraud.

Most popular ML algorithms in insurance

ed
t aliz & t
en n ng n nt ing en
so eti tio n n e it
em r a rn tio tio lli
g r d ion s m
im e
ad ag Pe ark s e z
ir c imi h u i c a
at c te erw au ect la nag
Le an m le P pt C red D xtra n
I nd r
F et C a
m sa o p e u d m

Gradient
Boosting

Random
Forest

Decision
Trees

K-means

Neural
Networks

Reinforcement
Learning

Computer
Vision

Clustering
Analysis

Natural
Language
Processing
(NLP)

Logistic
Regression

Also, one should know that data scientists usually test several models before choosing ones that
are the best candidates to solve the business problem. Building an ML model engages multiple
iterations: specialists have to train the model, tune parameters, evaluate, and monitor
performance. They stop only when they notice that the model makes the most accurate
prediction based on the training data.

21
Path to ML implementation

After the model is trained, the team needs to evaluate its results. The primary goal of testing is to
check that the logic learned by the machine remains consistent. Insurers widely apply A/B
testing as an efficient approach to make sure the new solution performs well and brings the
desired outcomes.

ML models training process

Dataset

Train set Model

Data Cleaning Feature Pre-processed

& Filtering Selection Data

Pre-processing Data
Test set Trained Model

Output

5. Deployment and monitoring

The final step is to put the prediction model into production, which also means testing its
performance and integrating it into the current system. Data scientists should also regularly track
the model performance and retrain it if needed, for example, in case of significant data changes.

6.Personnel education
Personnel education is also a critical part of a full-scale ML solution. At this stage, the data
science team, together with the insurer, teaches the personnel how to use the newly built
solution in the most efficient way. The HRM department should also make sure that employees
understand that the ML-powered solution is there to help them solve their day-to-day tasks
more productively and reduce the manual work, rather than replace the workers.

72% of business leaders said AI can enable workers to focus on meaningful

work, while 34% of business leaders said AI can help free up time for more
meaningful, valuable work.
PwC study

22
Create the most eicient ML-powered
solutions with Intelliarts
As evident from the six steps we’ve covered, the success of an ML-powered project largely
depends on the proficiency of your data science team. Intelliarts is a trusted partner for many
businesses across domains. We build end-to-end data-driven and ML-powered solutions and
help our partners with digital transformation, innovation, and expansion of engineering
capabilities.

The company has vast domain expertise in the insurance sector. This allows our engineers to
become involved in the project faster and naturally contribute to its outcomes. Having solved
different business challenges in insurance, we know the nuances of the industry, can provide
value from the business perspective, and help to evaluate risks better.

Our expertise includes but is not limited to:

Lead scoring and analysis Forecasting customer acquisition and
Sales process optimization retention

Demand prediction and price optimization Insurance settlement process optimization

Fraud detection and prevention Smart recommendation and assistant

systems
Churn prediction

Building or scaling your team with top engineering talent

Intelliarts is one vendor to cover your every

The Intelliarts team is well-experienced in
need. We build a full-scale ML-based software
insurance industry business challenges and
solution that flawlessly integrates into the
how to solve them. Our business acumen
existing ecosystem of software products that
helps us build efficient ML solutions across
the company uses right now. Similarly, we
domains.
develop a full-scale solution from scratch.

Strong
Focus on
AI/ML Domain Full-cycle CRISP-DM
business
engineering knowledge development methodology
outcomes
team

Our certified data science

and ML engineers leverage We use an industry-
We never do ML modeling
solid AI expertise and standard iterative approach
for the modeling itself, but
powerful research in data science that drives
focus on business results
capabilities to uncover the us through all the questions
and do our best to improve
full picture and offer the and obstacles to the
the outcomes.
most innovative AI/ML workable ML solution.
solutions.

23
 ML-powered solution to improve
Case Study
lead quality and agent
eiciency

Business challenge:
Our customer (under NDA) is one of the fastest-growing health tech insurers. The company
specializes in selling Medicare-eligible insurance coverage by working as an insurance broker.
They’re a big supporter of disruptive innovation and always seek ways to improve business
operations.

The insurer has contacted us to improve sales effectiveness, and after a series of the first
meetings with the customer, we came to an agreement that the best way for them to convert
leads into sales will be through:

Improving lead quality

Increasing insurance agent efficiency

Solution:
To meet these demands, the Intelliarts team has built the ML-powered solution composed of two
parts. The solution aimed at helping the customer detect regions with the most promising leads
(1) while also assigning them to high-performing insurance agents (2) to raise the chance of a
successful sale.

Here is how the solution was built step by step:

First, we conducted sales data analysis at a zip code level. The Intelliarts team tried to find a
pattern between the demographic and financial data and sales data provided by the customer.
The goal was to predict the average profit in the regions to choose those where sales would be
the most profitable.

However, the preliminary research proved that the team didn’t have enough sales data for
quality analysis at the zip code level. So, we moved to the data analysis based on a destination
sectional center facility (SCF). As a result, we’ve gotten two key insights into the lead quality:
the conversion rate was a bit bigger in sub-urban areas (1); there was no linear dependency
between AGI (Adjusted Gross Income) and conversion rate (2).

0,16 0,25
Conversion rate

Conversion rate

0,14
0,20
0,12
0,10 0,15
0,08
0,06 0,10

0,04
0,05
0,02
0,00 0,00
Rural Suburbs City Very low Low High Very high Medium

Density_rank Agi_rank

24
 ML-powered solution to improve lead quality and agent eiciency

To detect SCF codes with high conversion, our data engineers labeled SCF as high- and
low-performing based on conversion and built a PCA visualization. We then built the ML
classification model using decision tree and clustering analysis techniques, which should help
the customer choose the most profitable SCF regions for lead selection.

Total Explained Variance 71.02% 2 component PCA with 61.44%

Principal Component 2
6

4
8
6 2

4 0
-4
2 -2 -2
0
PC 3

PC 2
0
2 -4
-2 4
-4
-6
15 0 -5 0 5 10 15

Principal Component 1

Color: 1 0 Color: 1 0

To increase the efficiency of insurance agents, we built an algorithm that could detect agent
efficiency and connect the top-performing agents with the most promising leads.

Based on the collected data about agent calls, we created a statistical model that predicted an
agent conversion rate per week and contrasted it with the actual conversion rate. According to
the results, the agent could be promoted to a higher tier and, thus, receive more quality leads
and earn more. This motivated agents to work better and increase the chance of a successful
sale.

Business value
The ML-powered solution we’ve built helped the customer establish a better lead selection
strategy. Now the insurtech company spends much less time on SCF with possibly low
conversion and focuses most sales efforts on those high-performing ones. Respectively, the lead
quality has already increased by 5%.

Also, agent efficiency has grown by 3% because agents are more motivated to work to be able to
earn more. Altogether, these changes have improved overall conversion rates and maximized
company profitability.

Read more: Case Study on Improving Insurance Lead Quality and Agent Efficiency

25
Looking ahead: Transforming your
insurance business with ML
No doubt data is the single most critical differentiator for insurance organizations. Today,
information is being used for much more than traditional historical analysis. Insurers are
leveraging data to attract and retain customers with more customized offers and personalized
messages; assess and mitigate risks; optimize premiums; process claims; and improve fraud
detection.

With competition heating up for insurance companies, becoming a data-driven business seems
the only reasonable solution. And one of the most reliable ways to do it is to transform your
insurance business with machine learning. Here are also a couple of best practices you can
consider when deciding to add value to your insurance company with the help of ML:

Start with specific business goals for an ML project and a clear

Build a coherent
implementation plan. A long-term strategic plan will help you track
strategic plan progress easily and analyze the outcomes.

A good idea is to choose one or two use cases you’d like to

introduce to your business, such as an ML-based fraud detection
Start small model. Only after this project is implemented, your company can
scale up and move to building other ML solutions.

Integrate skills, technology, and insights important for building

Enhance talent efficient ML-powered solutions. Build your in-house ML expertise
and by hiring professionals directly or by reaching out to a trusted
external ML vendor.
infrastructure

Never make change a one-time initiative. As you see, emerging

technologies are vastly embraced in the insurance sector. New use
Scale up cases appear regularly, which can add more value to your business
and place your company ahead of the competition. Develop an agile
mindset and keep up with these trends!

The volumes of data used for ML projects usually create an

Think about data additional security risk for the insurer. Always keep in mind data
security safety and add an extra layer of security measures to avoid data
breaches and leaks.

Intelliarts offers insurers a strategic partnership — with our vast expertise and domain knowledge,
we can help you build full-cycle ML-powered solutions and take your business to the next level.

Regardless of whether you’re just getting started with ML or want to innovate your business
operations, email us and let’s partner.

26
Endnotes
Ravi Malhotra and Swati Sharma, “Machine Learning in Insurance”, Accenture,
1 https://www.accenture.com/_acnmedia/pdf-84/accenture-machine-leaning-insurance.pdf

2 “Artificial Intelligence (AI) in Insurance – Thematic Research”, Global Data, 2021.

3 Forrester, “Unlock the Power of Data to Transform Your Business,” IBM, 2019.

“What is Machine Learning”, IBM Cloud Learn Hub,

4 https://www.ibm.com/cloud/learn/machine-learning

“Hype or Reality: The State of Artificial Intelligence and Machine Learning in the
5 Insurance industry”, LexisNexis Risk Solutions, 2019.

Sylvain Johansson and Ulrike Vogelgesang, “Automating the Insurance Industry”,

McKinsey,
6 https://www.mckinsey.com/industries/financial-services/our-insights/automating-the
-insurance-industry

Henry Burton, “AXA Predicts Large Loss Claims with 78% Accuracy”,
7 https://www.linkedin.com/pulse/momentum-gathers-axa-turns-machine-learning-optimis
e-henry-burton/

Katharina Morik and Hanna Köpcke, “Analysing Customer Churn in Insurance Data – A
8 Case Study”,
https://sfb876.tu-dortmund.de/PublicPublicationFiles/morik_koepcke_2004a.pdf

“Background on: Insurance fraud”, Insurance Information Institute,

9 https://www.iii.org/article/background-on-insurance-fraud

“Anadolu Sigorta Success Story”, Friss.com,

10 https://knowledge.friss.com/download-success-story-anadolu-sigorta

Silke Genuit, “How to use AI in claims management”, Insurance Thought Leadership,

11 https://www.insurancethoughtleadership.com/claims/how-use-ai-claims-management

Tony Harrington, “Disrupting the insurance landscape through artificial intelligence and
deep learning”, ITIJ,
12 https://www.itij.com/latest/long-read/disrupting-insurance-landscape-through-artificial-in
telligence-and-deep-learning

27
About us

Intelliarts uses data science and machine learning to help

companies optimize, scale, and innovate their businesses to
unlock organizational value and gain a competitive advantage.
Seek assistance from the Intelliarts team to reveal the ways to
transform your insurance processes, build, and implement
eicient ML-powered solutions.

ai.intelliarts.com
If you’re ready to turn your
insurance company into a
data-driven business, let’s have
[email protected]
a chat to discuss

Book a call