Moderator: Ann Weber, JD. Presenters: Dan Adamson Angela Nelson

Transcription

1 Session 103 PD - Use of Artificial Intelligence in GI: Industry Opportunities and Regulatory Challenges Moderator: Ann Weber, JD Presenters: Dan Adamson Angela Nelson SOA Antitrust Compliance Guidelines SOA Presentation Disclaimer

2 AI in General Insurance SOA 2017 Dan Adamson

3 Themes 1. State of the Union 2. What is AI and all that jazz? 3. Lots of data but is it usable? 4. Where is it usable? Examples. 5. How to get started 6. Challenges & recommendations 2

4 The industry is changing All these new technologies have emerged creating a real tipping point potential. 3

5 Themes in FinTech Competitors Cooperators Technologies powering FinTech Blockchain AI / Machine Learning / Cognitive Computing Analytics 4

6 Themes in InsurTech Competitors Cooperators Blockchain AI / Machine Learning / Cognitive Computing Analytics Confidential. All rights All rights reserved. reserved. Copyright Copyright 2017 Exiger (c) Inc Not for Not redistribution for redistribution. 5 5

7 Tipping Point Example: e-discovery Similar to the revolution of e-discovery Zubulake v. UBS Warburg (2004) was the tipping point. imagine a world today without e-discovery How fast does this happen? In 2000, discovery was mostly done by print out, with teams of lawyers scouring over file folders In 2010, lawyers proposing that approach would be fired for insanity by their clients who want auditable, cost-effective discovery Will AI will do the same to insurance? 6

8 AI and Associated Buzzwords 7

9 AI in Insurance: Examples Move to (more) automated underwriting in new lines of business Claims monitoring (fraud detection) Investigations Customer service Marketing Adjusting 8

10 AI in Insurance: Approaches Areas AI Applications Strengths Issues Predictive model replacement Neural nets, ML, Deep learning Fraud detection, pricing, underwriting Less data required, pushing better detection, classification, scoring, messy data handling Suitability without explanations, regulatory controls Automation RPA Claims handling, policy handling, Rapid deployment ROI compared to traditional automation? Complex scenarios Client Support Chatbots (NLP, feedbackbased learning) Web response times Response times, controlled messaging Effort to develop in sophisticated scenerios Breakthroughs LIME, frameworks Speed of implementation Price points, major players, specialists 9

11 Use Case Example Automated Underwriting DDIQ employs the same cognitive processes to evaluate content as a researcher without the constraints of human-based research. The platform s iterative, learning approach allows it to question allows it to question and lock onto the subject and signifcantly reduces false-positives. Subject Information Data Input Enriched Seed Data Structured & Unstructured Data Acquisition W W W Investigation of Open Web and Premium Sources DDIQ Cognitive Engine Comprehensive & Accurate Research 10

12 Use Case Example Automated Underwriting with Context Interpreting unstructure d data is a hard problem for people... and computers! Too Many Results Wrong Subject Wrong Subject again Right subject But not a risk Wrong subject again Sample InsuranceIQ- Generated Report 11

13 Example Problem Should we underwrite this rare book collections? Top ranked (?) results are not relevant in this context More sources = the longer the list to wade through Not actually the Toronto Reference Library Highly ineffective and time consuming. And the problem is getting worse. THINKING OUTSIDE TOGETHER 12 12

14 AI for Commercial Underwriting Should we underwrite this rare book collections? Uncovering contextually relevant risks 13 13

15 How do you get started with Underwriting? Mode Phase 1 Phase 2 Phase 3 Underwriters use as assisted underwriting Exception-based underwriting Incorporated into pricing Starting point Feedback collected Classifiers trained initially on standard events Input from Phase 1 into exception-based model Input from Phase 2 results Underwriter feedback Underwriter feedback Predictive modeling w/backtesting 14

16 How do you get started with Customer Service? Example case study: RBS Users Input Phase 1 (June 2016) Phase 2 (Jan 2017) Phase 3 Provide tool for client Rollout to early Broader rollout support agents to use adopter segment of clients Trained initially on client support materials Input from Phase 1 Input from Phase 2 Feedback Client support agents offer alternatives, suggestions Client feedback, Trained on escalations to human 15

17 Regulatory Concerns and Recommendations Regulators of have concerns rolling out new technologies: An old measure of success was headcount New technologies can be unproven and some are black box FI s can implement incorrectly Not always clear what variables are being used Regulators need to be won over with: Consistency/reproducibility of results Same underlying effectiveness tests, e.g. sampling of traditional approaches Greater explainability and auditable records Increased performance and timeliness in some systems that are automated, for example realtime monitoring for adverse events 16

18 Take home messages The same evaluation approaches apply (a classifier is a classifier ROC curves, confusion matrices, etc) Regulators need to be won over design from the beginning Pick tracks where you can get started on the journey Concrete business case Support from multiple layers of the organization A project that isn t Boolean journey and not a destination 17