Why underwriters must set guardrails for AI: Scale exec

• Distribution, underwriting and claims are ripe for improvement
• Be selective about what to feed AI for underwriting
• Regulatory sandboxes could help AI development
  

AI shouldn't just replace existing systems in the insurance industry, it should orchestrate end-to-end processes, says Nirmitee Shah of Scale AI.

AI can be developed in sandbox environments to better serve insurers, she adds. Shah joined Scale AI as its general manager of financial and professional services in April, bringing experience from several major financial services firms to the role. 

Scale began as a data modeling company about 10 years ago. Two years ago, it remade itself as an enterprise-level AI provider for the insurance and healthcare industries, serving clients including Allianz. Its change mirrors the impact of AI on insurance companies. Shah, general manager of financial and professional services at the company, spoke with Digital Insurance about how Scale AI sees the insurance industry and where AI fits for insurers' operations.

This article is excerpted from a longer interview and edited for clarity. 

How is the insurance industry positioned going into this AI boom?

Insurance is a few years behind compared to hedge funds or even the top global banks in AI adoption. It has old processes and old ways of collecting data. There is an opportunity for insurance to leapfrog where it is today into the AI world, without the painful transformation other financial institutions have gone through. We are working with insurers and have a number of ways we help them get the outcomes they want.

What are some areas where the insurance industry can derive value using AI?

Insurance is a risk business. First is distribution, how insurance is sold to people in the marketplace. Second is underwriting. Third is claims. Claims hits the bottom line, and is probably the place where the insurer has the most risk. In claims, we understand there's a lot of regulation. In claims, we can start with subrogation or the claims process. At the end of the day, the idea is to improve the ratio and feed claims information in ways that are appropriate and regulated to other parts of the organization to make underwriting better, to make the connected enterprise work.

Where does the insurance industry stand in accepting AI?

As a whole, the industry is where it has always been with new technology, which is waiting to watch and see what happens and then jump in. There's been some real AI solutions built in some of the largest insurers. We are working with some very large insurers to change the way they work and the way their entire business operates. We're seeing some early conversations and some midway through implementation.

How do you see AI getting applied in underwriting?

AI is like a very smart 10-year-old. If you let the AI run amok, it's going to make inferences that lead to things that we don't want, that can cause real risk to the operations of an insurance company. There could be weird things that the AI does because it enforces things based on pattern recognition. The critical thing about AI in the underwriting space is to actually tell it what not to do. Everyone can tell AI what to do, but to tell AI not to infer based on X, Y, Z criteria is critical to successful AI implementations in insurance.

Would a regulatory sandbox environment help insurers implement AI successfully?

If regulators create a sandbox environment for insurers to work and understand what AI can and can't do in sensitive functions like underwriting, that might go a long way in accelerating adoption of AI in insurance. A sandbox environment to figure out if things work or don't work would be a real accelerator to the industry. If the NAIC [National Association of Insurance Commissioners] could do that, that would accelerate some adoption for sure.

How do you view AI in an insurance operations context? 

I don't think of AI as something we want to rip and replace existing solutions on. I think of AI as a way to create an orchestrated end-to-end process. There are aspects of claims AI does really well. Where I see AI coming in is to create an end-to-end process with an agent in the workflow, and using available data to pass context from one process to another to create better outcomes. 

The second aspect is being able to mine information — taking the data created through that end-to-end process to start pattern mapping. If I can replicate that behavior for any single claim I have, I'll have better outcomes. The idea is to not replace what exists, but to create value by creating a workflow around what exists.

What leaps or assumptions are made with AI and how should those be addressed?

Get the right evidence, stop wrong data privacy layers and create gates for regulatory purposes. The level of expertise and skill people bring to the table is vastly different. Data privacy, especially in insurance, especially in health insurance, can be a very big gate.

As organizations become more and more AI savvy, it is going to be critical for them to understand whether their AI models are performing or not.