Artificial Intelligence in Insurance

Transcription

1 Artificial Intelligence in Insurance Serhii Kondraiuk & Jake Diner November 2017 V. 4.6

2 AI in Insurance Nov Jake Diner Elafris, CEO Elafris, Assurant, Driveway, Belkin, Dell, Computer Science from Michigan State University Serhii Kondratiuk Elafris, Lead Scientist Elafris, Azzurro, Samsung R&D Kiev, Computer Science from Taras Shevchenko National University

3 Insurance is old and very, very big Over 1,500 Over 1 million Over $600B 10% First Insurance policy written (fire protection) Oldest operating insurer formed, Philadelphia Contributionship First regulations adopted to maintain adequate reserves Insurance companies Insurance agents, broker and service employees In annual revenue Agent s average commission

4 Artificial Intelligence Applications Segmentation Claim Adjudication Claims Automation Virtual Agents Repair Estimation & Model Detection Utilizing new unstructured data for risk analytics Understanding exclusions and fighting fraud OCR for Robotic Form Automation Using NLP for customer support and sales Image classification

5 Elafris Focus on Virtual Agents How did we get started Building a business case before anything else Positioning is the key Communication Channel Customer Support Insurance Agent Select technology to deliver on the business needs

6 The Solution Our AI-powered virtual insurance agent (chatbot) provides quotes, collects premium payments, sells policy add-ons, and helps with claims using

7 Product in the Market For Agents and Brokers Quotes Payments Customer Support Cross-sell & Up-sell For Insurers Policy Claims Roadside

8 Vertical Workflows Insurer-defined process Dialog design Dataset creation NLP for Quotes

9 Challenges Encountered From a technical standpoint, a typical chatbot is a set of rule-based actions with the user confined to the dialogue flow dictated by the bot creator. This solution works for a limited, simple cases. We elected to design our AI module to genuinely recognize the user s intent based on his input. But this approach gives rise to multiple challenging obstacles: Users phrase choices are extremely diverse, so sticking to a dictionary is not adequate or efficient. Users word choices and inputs differ widely depending on the specific context. Unrelated user input is common and must be managed correctly and elegantly. Users must have instant access to change submitted info almost impossible with rulebased flows.

10 Challenges Encountered Our fuzzy model is able to work with non-linear dialog flow but requires more data to train compared to classic approaches. 1,000+ phrases per intent are usually required for the AI module to recognize it. ЫЩООО! One challenge we conquered was the building a sustainable process for collecting verifying extending the data. We ve assembled a linguist team for this task, working closely with our development and data science teams. We created and implemented specialized tools allowing fast and convenient intent creation and data processing.

11 Considerations & Results Our approach grants users a comfortable, intuitive experience. Non-linear dialog flows Instead of the interrogating the user, our chatbot offers a dynamic conversation guided by the user s mindset, not the chatbot s. Multiple intents from a single input Adaptable to an infinite user communication styles. Wordy or short, formal or casual, our chatbot recognizes intents as they are communicated, prompting for the missing info when it naturally fits the conversation. With our AI approach we we re able to reach intent recognition of 98% accuracy.

12 Customers

13 Elafris Growth 50, % 4x Unique user conversations Messages read rate Insurance Quotes conversion rate compared to web forms 300% 2 Visit YoY company growth rate Open positions Sr. Python and Sr. NodeJS Elafris.com/careers

14