.png){kind=small}
As an insurance analytics executive, you know that you need to lean into GenAI, but where and how should you begin this journey?
Demystification
Productionizing GenAI involves more than typing prompts into ChatGPT or using a code co-pilot. There’s application programming, data set creation, often tight coupling with math-oriented machine learning models, and other facets to build non-trivial applications.
Moreover, it’s not uncommon to overestimate what today’s generally available GenAI tools can do by themselves. For example, in our first foray, we wanted AWS’s Claude to tell us, based on ~1 million claim notes, which property features were most predictive of large losses. To do that, we’d need tools from AWS Sagemaker, Claude, Titan, and other programming tools to create the application pipeline.
So, based on what’s available now for general use, what is GenAI good at regarding typical P&C workflows?
- Summarizing policies, documents, and other unstructured content
- Synthesizing and summarizing summarizations
- Answering questions based on what it has learned from summarizing
- Translating between languages such as English and SQL
- Creating and repairing computer code
P&C Use Cases
The day will come in the not-too-distant future when GenAI is indispensable to customer experience, underwriting, claim processing, rate determination, etc. But we’ll have to crawl and then walk before we run. So, what are some “starter” use cases today? Here are some excellent examples:
Underwriting:
- Generate a risk profile based on underwriter documents
- Generate assessment of risk alignment to exposure appetite
- Flagging and identifying missing material data
- Validating submission data against additional sources (e.g., inspecting tree overhanging)
Claims:
- Damage Assessment
o Scan policy terms to assess claim coverage quickly
o Generate more timely alerts
o Make recommendations for adjusters
o Couple with computer vision technologies to improve damage assessment
o Summarize court rulings
o Gather evidence for negotiations
- Fraud Detection
o Flag suspicious claims
o Analyze data from multiple sources – customer profiles, historical claims, external databases
- Preparation of mailings to claimants
Actuarial:
- Data manipulation
- Creating test data
- Generate summaries: recent financial experience, product descriptions, and market evaluations.
- Documentation of model steps
Customer Service:
- Sentiment analysis
- Domain-relevant Q&A
- Personal recommendations
- Contextual assisted chatbot
IT:
- Code generation based on text prompts
- Code review, bug fixing, and suggested improvements
- Generating code comments to explain and document code
- Generating unit and regression tests that can be integrated into the build process
- Amazon Code Whisperer; Github Copilot; Microsoft Copilot
How To Mobilize
Any department that doesn’t get on board will be left behind. Our recommendation is to empower a cross-functional steering committee to prioritize use cases for each department and, behind that, an operating committee that includes business domain experts, IT specialists (many software tools will be required), and data scientists to guide the journey. Prioritization should be based on cost, benefit, and readiness for adoption. A two-month project to inventory and prioritize potential use cases, may be a great way to start.
Start by building easy use cases. Most importantly, develop your operating model, i.e., roles, responsibilities, gates, and hand-offs across the stages of ideation, design, engineering, testing, deployment, and monitoring. You'll move from walking to running as you burn in your operating models and get use case experience.
By Mike Lamble
Partner and Founder, PremiumIQ
Comments