GenAI and the Pursuit of ROI in Insurance
Introduction
Extracting business value from GenAI is the new gold rush. In 2023, the expectations from this gold rush peaked (ref: Gartner hype cycle). This meant tremendous pressure from shareholders and stakeholders to demonstrate that you weren't falling behind and missing out. Large investments were made in people, technology, hardware and time to identify and build use cases that utilized emerging GenAI models and techniques.
Now the bill has come due, and most companies are stuck in a POC quagmire. They have ideated on and built several successful POCs utilizing GenAI but are realizing that deploying these at scale is much more complex and expensive than they had initially estimated. A recent Lucidworks report said that despite initial hype, slow deployment and low success rates are commonplace, with only 25 percent of planned projects becoming fully implemented. This lag is stalling anticipated ROI, with 42 percent of companies yet to see a significant benefit from generative AI initiatives. And now the very same stakeholders are asking – "where is the business value that was promised." This problem is even more acute in insurance because of a few factors:
- With combined ratios hovering near 100 percent, there isn't much wiggle room to play around with; i.e. margins are tight. Any investments made need to show a quick path to ROI.
- Regulatory bodies are aware of the opportunity and the risks posed by GenAI and are moving quickly trying to craft new, effective guidelines in anticipation of the industry's embrace of even more disruptive GenAI tools. For example, in December 2023, the National Association of Insurance Commissioners (NAIC) released its model bulletin that provides guidance allowing for responsible use of AI by insurers with an emphasis on transparency, accuracy and lack of bias in deployed AI models along with robust governance and risk management controls. Regulation in insurance also happens at both the federal and state levels, making this even more complex. This means additional pressure on the models to be accurate and without bias.
- Generative AI systems that bring in proprietary data through RAG or fine-tuning rely on large, high-quality datasets. Insurance firms are often faced with the challenge of inconsistent, incomplete, and potentially inherently biased data, which can lead to biases in decision-making, perpetuating unequal treatment and unfair outcomes. Another GenAI-specific risk driven by incomplete or bad data is hallucinations, i.e., where the GenAI model fabricates authentic-looking data. All of these outcomes could lead to liability risks. For example, if a decision is made on denying coverage or denying a claim based on a hallucinated or biased output, it opens organizations to regulatory scrutiny and lawsuits. Data siloes and large swathes of non-digitized data (especially in commercial insurance) further exacerbate this problem and hence data management (cleansing, aggregation, normalization, governance) becomes crucial.
- GenAI use cases can represent a significant change in existing workflows, which at insurance firms can be quite complex. Not everyone in the organization will be comfortable with this. Hence, successful implementations with widespread adoption often require significant cultural and process shifts.
Realizing value
So how can insurance organizations navigate this maze of challenges to realize the promise of business value from GenAI? How can you take successful POCs and deploy them to production at scale?
Use case discovery & prioritization
Firstly, to even get to this stage, you need to ensure you are working on the right ideas, i.e., use cases that have a direct impact on the financial metrics, either in terms of driving revenue growth or reducing expenses or losses. You need a well-defined process to identify and prioritize the right use cases. You also need to validate those use cases not only in terms of business value but also from an adoption, deployment complexity, risk and operational disruption perspective.
For example, models that help you with new product development or improved underwriting may have the largest impact but also carry with them increased implementation complexity and higher risks from a regulatory and liability perspective, whereas a claims processing or policy audit automation use case may have lower value, but reduced complexity could mean ROI impact at a higher velocity. Once you select the right ideas, you should be able to rapidly spin up, iterate, evaluate and abandon POCs as needed.
Deployment & scaling
Now that you have a library of POCs, you need a scalable, repeatable, continuous LLMOps process to deploy, measure and refine models and data pipelines. Given the unique regulatory and data challenges in insurance, there needs to be an enhanced focus on data engineering, data governance and model evaluation.
Adoption
A deployed scaled-up solution will only deliver value if it is adopted by users. There are three things to focus on to ensure that:
- Incentivize: Demonstrate tangible impact for users in day-to-day activities, display executive level buy-in and adoption and connect to measurable impact on key financial metrics like the expense ratio, loss ratio, total premium, etc.
- Remove adoption barriers: Insurance workflows are inherently complex, with processes like underwriting and claims assessment often involving multiple internal and external systems, back and forth with external partners like brokers and claims assessors and parsing of a lot of data. Hence, any new functionality driven by GenAI should, as much as possible, be integrated into these existing workflows, rather than creating one more system that users need to access. It is also essential to allay fears around the accuracy of model output by providing transparency into the provenance of model output and providing formal feedback processes and mechanisms.
- Educate: Provide detailed documentation on use case features and functionalities and conduct organizational change management training and education to improve AI literacy, highlight how to avoid the common pitfalls of GenAI and promote AI adoption across the organization.
Industrialization
The next question to tackle is, once you successfully deploy this process, how do you industrialize this across your organization such that you have a conveyor belt of use cases churning out business value for you? The key here is in developing accelerators, either through investments or partnerships. Some of these accelerators can be:
- Build out a GenAI COE, A Centre of Excellence responsible for overseeing the organizational GenAI strategy and developing frameworks for security, governance, infrastructure strategy, RAG implementation, and model evaluation.
- Work with partners to gain access to a lab where you can experiment, test and innovate, with hands-on access to the latest AI hardware, software, and reference architectures.
- Invest in building out an LLMOps capability: A scalable and flexible framework for managing LLMs with a defined set of tools and techniques to enable model data preparation, model training, model deployment and monitoring.
- Develop a measurement and "bank-back" framework: Build the ability to have a clear line of sight to measure and quantify business value, not just improvement in operational metrics. For example, a Big Bang use-case is one that dramatically saves people time for a large population. But that time can just get converted to coffee breaks, so you need a measurement method to track how the time gets used and might have to do organizational re-engineering to "bank" the savings you get, e.g., set a specialized R&D team if R&D is what has been automated via GenAI.
- Build out a GenAI Experience Studio: A dedicated team of designers, engineers and product managers focused on integrating GenAI into day-to-day workflows.
- Establish frameworks focused on data management, specifically for Gen AI use cases, with an emphasis on data quality and provenance, data infrastructure and data governance.
In conclusion, while organizations may face a difficult uphill challenge in terms of extracting meaningful ROI out of their GenAI efforts, some key pillars, if invested in and developed, can help not only unlock this value for use cases currently under consideration but also establish a foundation to increase the velocity to ROI for future use cases.