HOW AI IS CHANGING THE
WORLD OF INSURANCE
Dr. Nilesh N. Karnik
Chief Data Scientist, Aureus Analytics
Seminar Name: 15th Seminar on Current Issues
in Life Assurance (CILA)
Venue: Hotel Sea Princess, Mumbai
Date: 20-12-2019
www.actuariesindia.org
WHY AI?
2
CHALLENGES FACED BY
INSURERS
www.actuariesindia.org
Tapping into potential
customers at the right time
Reduce fraud
Providing the right set of products/services
that meet customer requirements
Giving customers a hassle-free claim
experience
AI helps re-define Customer Experience
3
www.actuariesindia.org
WHAT IS AI?
4
www.actuariesindia.org
Natural language processing
Deep learning
Neural networks
Machine Learning
Self driving vehicles
Computer vision
Fuzzy logic
Expert systems
Robotics
5
WHAT IS ARTIFICIAL
INTELLIGENCE?
www.actuariesindia.org
THE THEORY AND DEVELOPMENT OF COMPUTER SYSTEMS
ABLE TO PERFORM TASKS NORMALLY REQUIRING HUMAN
INTELLIGENCE
6
WHAT REQUIRES HUMAN
INTELLIGENCE
78829
173467 + 2663747
283
÷
1
3
Finding the fastest check-
out line at the super market
Identifying your spouse in
their school photograph
www.actuariesindia.org
7
WHAT IS ARTIFICIAL
INTELLIGENCE?
www.actuariesindia.org
THE THEORY AND DEVELOPMENT OF COMPUTER SYSTEMS
ABLE TO PERFORM TASKS NORMALLY REQUIRING HUMAN
INTELLIGENCE, SUCH AS
VISUAL PERCEPTION, LEARNING
FROM EXPERIENCE, DECISION-MAKING, AND UNDERSTANDING
HUMAN LANGUAGES.
8
VISUAL PERCEPTION
www.actuariesindia.org
ABILITY TO COMPREHEND IMAGES AND VIDEOS
IDENTIFYING OBJECTS
DETECTING MOVEMENT
GETTING A 3-D UNDERSTANDING OF THE ENVIRONMENT
9
LEARNING FROM
EXPERIENCE
www.actuariesindia.org
LEARNING FROM HISTORICAL INFORMATION (DATA)
ABILITY TO ADAPT TO CHANGES IN ENVIRONEMNT
ABILITY TO GENERALIZE
10
DECISION MAKING
www.actuariesindia.org
ABILITY TO USE EXISTING EXPERT KNOWLEDGE
COMBINE WITH KNOWLEDGE FROM EXPERIENCE
RESOLVE CONFLICTING RULES
11
UNDERSTANDING HUMAN
LANGUAGE
www.actuariesindia.org
UNDERSTANDING SPEECH
RESPONDING IN HUMAN LIKE LANGUAGE
RESPONDING IN HUMAN LIKE SPEECH
UNDERSTANDING WRITTEN LANGUAGE
12
www.actuariesindia.org
AI IN INSURANCE
13
www.actuariesindia.org
New facial analysis technology helps find indication of:
• BMI
• Age
• Gender
• Smoking
Useful for better underwriting of life insurance policies.
WHAT’S IN A SELFIE?
14
Complex features
HOW DOES IT WORK?
www.actuariesindia.org
Feature
Extraction
Risk
estimates
Selfie
Models trained
on past trends
LEGAL & GENERAL AMERICA
15
Other applications of image
and video analysis
www.actuariesindia.org
Automatic analysis of accident pictures for faster claim processing
Analyzing Geo-Spatial imagery for better estimates of property and
home insurance premiums
Real time analysis of driver behavior for road safety
ALLSTATE
LIBERTY MUTUAL
AGRICULTURAL INSURANCE
COMPANY OF INDIA
16
REAL-TIME CAR DAMAGE
ASSESSMENT
www.actuariesindia.org
Tractable technology uses
image recognition
technology for automated
damage analysis.
The technology is
expected to shorten the
process for assessor to
visit, inspect and evaluate
the expenses for the
damaged car -
significantly from weeks
to one day.
17
SATELLITE IMAGES FOR
AGRICULTURAL INSURANCE PRICING
www.actuariesindia.org
The use of satellite images
helps to survey and monitor
a large agricultural area day
and night.
The satellite images allow
insurers to receive real-time
updates of potential perils in
the fields. The data from the
images, with the boundary
of the insured, will help
insurance to price risks
more accurately, increase
efficiencies and lower
operating costs
18
USE DRONES TO TAKE
PHOTOS OF HOUSE ROOFS
www.actuariesindia.org
The use of drones in the
Property & Casualty
insurance will soon
become the standard
procedure for quoting,
inspection and damage
assessment.
A drone can take hundreds
of images in 10 to 20
minutes for quoting
purpose. The use of
drones provides speed and
service.
19
RISK MODELING WITH IMAGE
DATA
www.actuariesindia.org
Facebook can identify 98% of its images
to the right person.
Facebook uses its imaging technology to
identify and remove fake accounts. Such
image-based fake-identification has
immense potential in banking and
insurance. There is numerous potential
in using the image data for fraud
identification.
A fraud model can be enhanced by the image
score to identify a false account and transaction.
20
LOOK WHO’S TALKING?
www.actuariesindia.org
Chatbots have been used
successfully to achieve
• Improved customer
response times
• Cost savings
21
HOW DOES IT WORK?
www.actuariesindia.org
Natural language understanding
Natural language generation
ALLSTATE / ABIE LEMONADE
22
CLAIMS PROCESS
AUTOMATION
www.actuariesindia.org
Allstate Business
Insurance has also
recently developed ABIe in
partnership with EIS. ABIe
(spoken as Abbie) is an
AI-based virtual
assistant application
designed to cater to
Allstate insurance agents
looking for information on
ABI’s commercial
insurance products.
23
www.actuariesindia.org
RECOMMENDING THE CORRECT
PRODUCT
• Product recommendation models are
getting more and more popular.
• They improve lead conversion.
• The customer benefits from an
unbiased recommendation and is likely
to be more persistent.
24
www.actuariesindia.org
HOW DOES IT WORK?
Matching customer profile with available choices
Predicting purchase propensity
Customized coverage as per customer needs
Right time to offer
External data can be very useful.
INSURIFY CLEARCOVERINSHUR
25
www.actuariesindia.org
PREDICTIVE MODELS
Prediction is difficult,
Especially so when it is about the future !
26
www.actuariesindia.org
HOW DOES IT WORK?
Machine learning algorithms
Learning repeating patterns from historical data
27
www.actuariesindia.org
CASE STUDY:
PREDICTING THE RISK OF AN EARLY CLAIM
28
www.actuariesindia.org
OVERALL PICTURE
Submission of
insurance
application
Policy
issuance
Model data
Real time
response :
Prediction of early
claim risk
Real time request
: sent for every
submitted
proposal
Insurer systems
Cloud
Predictive model for
identifying the risk of
early claims
End of day data feed to
update metrics used by
model
Request for scoring a
proposal includes
information about that
proposal, such as
premium, sum assured,
etc
Scored response by
the predictive model
generally includes a
numerical score, a
category label (such as
RAG) and a list of
influencers (detail
about how variables
affect the prediction)
Claims
29
www.actuariesindia.org
PREDICTIVE PROBLEM
DEFINITION
Predict the risk of a early claim – claim within 3 years of issuance.
Universe for prediction : All submitted proposals
Prediction at proposal submission
Data available
at the time of
policy
issuance
Predictive
model
Green – Low risk
policies
Amber – Medium risk
policies
Red – High risk policies
30
RESULTS
4 in 1,000
2.5%
6 in 10,000
5%
14%
100%
~ 6x of the
average
probability.
Captures nearly
half of risk in a
small set less
than 5% of the
portfolio size
~ 1/8 of the
average
probability
18 in 10,000
39%
85 in 10,000
14%
38 in 10,000
28%
www.actuariesindia.org
31
HOW WAS THE MODEL
CREATED ?
Prediction
indicating
early claim
risk bucket
Proposal
record to
be scored
Model A
Model B
Model C
Highest risk
from all 3
algorithms
A composite model created by combining 3 different models:
Model 2: Uses Gradient Boosting algorithm
Model 1: Uses Random Forest algorithm
Model 3: Uses a neural network
www.actuariesindia.org
32
www.actuariesindia.org
SIGNIFICANT PREDICTORS
1. Age of Customer as on Submission Date
2. Product Category
3. Ratio of Premium Paying Term to Benefit
Term
4. Agent’s Claims to policies Issued Ratio
5. Marital Status of Customer
….
33
www.actuariesindia.org
CASE STUDY:
PREDICTING RENEWAL PROPENSITY
34
www.actuariesindia.org
OVERALL PICTURE
Policy
admin
Model data
Incremental
data shared
at the start of
every moth
Insurer systems
Model
infrastructure on
premise
Predictive model for renewal
propensity
Incremental monthly data
includes new policies issued
since last month, new payment
transaction, status changes,
incremental CRM records and
agent details since last run.
Scored response by the
predictive model generally
includes a numerical score,
a category label (such as
RAG) and a list of
influencers (detail about
which particular variables
affect the prediction for a
particular policy)
Renewal
propensity for
policies following
due in the next 3
months
Payment
Management
CRM
Agent
admin
35
www.actuariesindia.org
PREDICTIVE PROBLEM
DEFINITION
Predict whether a given policy (which is nearing its due date) will
pay the premium before the end of its grace period.
Universe for prediction : All non-monthly policies*
Prediction is done periodically – at the start of every month
*Note: A separate model was created for policies with monthly payment frequency.
Data available
at the time of
prediction
Predictive
model
Green – Low risk
policies
Amber – Medium risk
policies
Red – High risk policies
36
www.actuariesindia.org
RESULTS
66%
30%
91%
30%
20%
100%
~ 1.4x of the
average probability.
Captures ~41% of
renewals
Less than
1/2 of the
average
probability
66%
50%
Note: This is a supervised ML model created using Gradient Boosting Machine algorithm.
37
www.actuariesindia.org
SIGNIFICANT PREDICTORS
1. Time of last payment
2. Did the policy ever miss it’s payment date
in the past?
3. Historical in-force ratio for the product
4. State
5. Vintage
…
38
www.actuariesindia.org
INDUSTRY TRENDS
39
www.actuariesindia.org
SIGNIFICANT PREDICTORS
WITH ONLY 1.3% OF INSURANCE COMPANIES INVESTING IN AI
COMPARED TO 32% IN SOFTWARE AND INTERNET TECHNOLOGIES,
THE INSURANCE INDUSTRY IS STILL LAGGING BEHIND IN THE AI
MOVEMENT.
THE VALUE OF GLOBAL INSURANCE PREMIUMS UNDERWRITTEN BY
ARTIFICIAL INTELLIGENCE WILL EXCEED $20 BILLION BY 2024, UP
FROM AN ESTIMATED $1.3 BILLION IN 2019.
INSURANCE INDUSTRY COST SAVINGS FROM AI WILL GROW FROM
$340 MILLION IN 2019 TO $2.3 BILLION BY 2024.
Source: Juniper research
40
www.actuariesindia.org
40
ANY
Q UE STIONS?
