© 2021 IBM Corporation

Processing Insurance Claims with Automated, Scalable and Fair AI

Maxime AllardData Scientist, IBM Data Science & AI Elitemaxime.allard3@ibm.com

Kristen McGarryData Scientist & Industry CTO, Financial Serviceskristen.mcgarry@ibm.com

Kenneth LimSenior Data Scientist, IBM Data Science & AI Elitekenneth.lim@ibm.com

July 22, 2021

Agenda

• The new horizon of claims management

• Infusing AI across the process

• Deep-dive into our latest accelerators

• Successful case studies

• Q&A

2

Acquisition

• Digital channels & online marketplaces

• Retail Agents, Brokers, advisors

Capture, classify, and extract data from

applications

Automate u/w decisions with business rules

Design and manage end-to-end workflow

Securely manage content

Automate repetitive tasks with RPA

Improve operational Intelligence

Automate with low code / no code tools

Policy Issuance• Digitized policies, riders,

terms and conditions

• Predictive underwriting models against claims data

IT Management

• Proactive incident management resolution

• Enforcing policies & cost controls across cloud

Invoicing & Billing

• Modern digital payments

• Optimized transactions, across channel ecosystem

1

6

2 3

© 2020 IBM Corporation 3

Provide an enhanced customer experience throughout customer journey

Improve and optimize with Process Mining

Manage API lifecycle

Integrate with rating and underwriting

systems

Claims / Service• Contactless mobile / virtual

claims processing

• Detect & avoid fraud, while supporting legitimate claims

4

Compliance• Automated regulatory

assessments,

• Improved agility and risk management

5

Observe and optimize

environment

Improve resilience and avoid downtime

Inefficient claims processing is costly and a major priority for insurers this year.

Processing claims make up 30% of operating costs on average.1

Inefficient processing leads to claims leakage, costing the insurance industry more than $30B/year.

Insurers experience leakage up to 25% vs. 3% industry benchmark.2

This is becoming the #1 priority for insurers in 2021.3

4

1: https://www.mckinsey.com/industries/financial-services/our-insights/from-transparency-to-insights-mckinseys-insurance-cost-benchmarking-20162: https://www.pwc.com.au/industry/insurance/assets/stopping-the-leaks-jan15.pdf3: https://go.forrester.com/blogs/europe-predictions-2021-financial-services/

AI can help process claims more efficiently and save insurers substantial amounts of money.

Reducing claims leakage can save insurers 5%--10% of their costs, translating to millions of dollars.4

AI can minimize potential claims leakage prospectively vs. diagnosing retrospectively (status quo).

Up to 70% of claims can be handled automatically.5

54: https://www.pwc.com.au/industry/insurance/assets/stopping-the-leaks-jan15.pdf5: https://www.digitalistmag.com/customer-experience/2019/04/10/digitally-fully-automated-claims-is-not-2030-dream-its-possible-today-06197774

6

Claims TriageFirst Notice of Loss/Benefit

Assess complexity of incoming claims instantaneously for effective routing to surveillance teams, claims handlers, or automated resolution.

Claims LeakageAssignment

Automatically distribute claims to minimize claims leakage and efficiently optimize workloads across handlers.

Claims ProcessingManagement

Equip handlers with transparent and debiased AI predictions on the most likely outcomes and predicted settlement amounts.

Our Approach: Enhance claims management with speed, cost reduction and fairness

Speed Cost Reduction Fairness

©2021 IBM Corporation 24 July 20216

Input: Claim Records; Claims History; Complexity Assessment

Output: Complexity Rating

Inputs: Claim Records; Claim Adjuster Profiles

Output: Claim Adjuster Assignment

Inputs: Claim Record; Analytical Models

Output: Unbiased Claim Settlement

User Interface Business Use Case Technical Assets

• Repeatable, customizable,

modular

• Notebooks, AI Models,

business terms and more

• Demonstrable on Cloud

Pak for Data

• Address specific real-world

problems

• Show tangible business ROI

from applied data science

projects

• Interactive visualization of AI

output in business context

• Example customizable

business application

A development-ready AI project starter

customizable to your use case

8

8

A packaged set of assets that address common business issues to kickstart your own implementation. End-to-end implementation, from data prep to deployment.

Data sources

Input your data using our sample catalog

Data schema

Define and map your data to the machine learning model

Watson Knowledge Catalog

Business terms and categories with predefined mappings

Machine Learning Developers

• Watson Studio• AutoAI• SPSS Modeler• Notebooks• CPLEX• Scripts

Machine LearningModels

Watson Studio – Deploy and manage the Machine Learning lifecycle with online or batch deployment

Model Monitoring• Explainability• Bias Mitigation• Model Drift

API Endpoints

Sample UI

Collect AnalyzeOrganize Infuse

Analytics Project

RETRAINING MODELS

✓ Increase speed to market

✓ Provide proven use cases for successful AI adoption

✓ Eliminate deliverable creation - tailor vs. create

✓ Deploy ready to scale assets

✓ Promote adherence to industry best practices

IBM Data Science and AI EliteDemo

9

Insurance Claims Leakage Accelerator

Industry Accelerator on IBM Cloud Pak for Data

VIDEO: INSERT cwtl_leakage.mov

Data Science and AI/ March, 2019 / © 2019 IBM Corporation 10

IBM Data Science and AI EliteDemo

11

Insurance Claims Processing Accelerator

Industry Accelerator on IBM Cloud Pak for Data

VIDEO: INSERT cwtl_processing.mov

Data Science and AI/ March, 2019 / © 2019 IBM Corporation 12

Data Science and AI/ March, 2019 / © 2019 IBM Corporation 13

1. Automating Claims Classifications and Regression

Tabular Claims data, containing both categorical and numerical features (e.g. age, dates, covered amounts)

2 Predictive Steps in the solution:

– Predicting likely outcome of claim (i.e. Settled or Rejected)

– Predicting possible claim leakage for assigned claim handlers

We let Watson Studio handle imputations and model choice as well as Feature engineering through Cognito1 and ADMM2

Cognito finds relevant features (e.g. through PCA or other simple transformations)

1: https://ieeexplore.ieee.org/abstract/document/7836821

2: https://arxiv.org/abs/1905.00424

13

(c) IBM Corporation 2021

IBM Data Science and AI EliteDemo

14

Watson Studio - AutoAI

Industry Accelerator on IBM Cloud Pak for Data

VIDEO: INSERT Watson_AutoAI.mov

Data Science and AI/ March, 2019 / © 2019 IBM Corporation 15

Data Science and AI/ March, 2019 / © 2019 IBM Corporation 16

2. Fairness in Claims Predictions

Certain features induce bias against a certain group of people (group fairness) or individuals (individual fairness)

Both human specialist and Watson Studio monitor this bias

For our example, bias against younger and older age-groups has been detected

Bias mitigation can be performed to assist the Claims Handler (by using for example Reject option classification1)

In addition to LIME2, we use Contrastive Explanations3 to explain the model’s decisions to help Claim Handler understand the decisions

1: https://ieeexplore.ieee.org/document/6413831

2: https://arxiv.org/abs/1602.04938

3: https://www.ibm.com/downloads/cas/0ZRZNR8E

16

(c) IBM Corporation 2021

IBM Data Science and AI EliteDemo

17

Monitoring with Watson Studio

Industry Accelerator on IBM Cloud Pak for Data

VIDEO: INSERT Watson_Monitoring.mov

Data Science and AI/ March, 2019 / © 2019 IBM Corporation 18

Data Science and AI/ March, 2019 / © 2019 IBM Corporation

3. Optimization of scheduling

Based on predictions for the claims leakage, we solve a problem to find the optimal scheduling among claim handlers

Claim handler features and their workload are used as inputs to the claims leakage classifier

Optimal distribution between experience of claim handler, their case load and difficulty of claim case

CPLEX Optimizer allows for large-scale optimization across thousands of claims and handlers

4. Transparency is the Key

• It is not enough to model claims automatically but the whole data and modelling process needs to be documented

• AI Fact Sheets are important for any business using ML models

• Communicating what steps are used makes it easier for the consumer to understand how this could impact any decision

20(c) IBM Corporation 2021

Expected Outcome

If the claim agents could see the predicted

outcomes at point of claim notification, it would

lead to:

— Improved Customer Satisfaction and

Experience

— Greater consistency in claims routing.

— Reduced Time and Cost

Solution Components⎻ IBM Cloud Pak for Data⎻ IBM Watson Studio

“The Data Science Elite team really understood the persona of the claims handling agent, the outcomes from a customer perspective and the challenges we face from a data perspective and worked with the internal teams to design and build a proof of concept model that could produce tangible outcomes – great work”. Barry Hostead

Head of Data Management, MI and Analytics Insurance and Wealth

Lloyds Banking Group

IBM’s Data Science Elite team strives to improve customer experience by helping insurance claims handling agents make appropriate claim routing decisions by reducing time and cost

Industry: Banking and Financial MarketsGeography: Europe

Unique Challenge

— Siloed data across organization.

— First data-driven approach to predicting claims to assist agents.

— Claims-specific data is in both structured and unstructured formats.

Use Case

To embark on its digital transformation journey, Lloyds Banking Group wants to improve customer claims experience using a data-driven approach to support claim handling agents to make more informed routing decisions by leveraging their existing data

Conclusion

Inefficient claims processing is costly.

AI can reduce inefficiencies and costs but face ethical and scalability concerns.

We developed a solution that automates the creation of predictive and prescriptive models with explainable and fair results.

View our accelerators: https://community.ibm.com/accelerators

22