Newsletters
Risk Central
Digital EditionDeciphering AI’s Promises and Perils to Ensure Value in Workers’ Compensation Analytics
Frank Huang, FCAS, MAAA is Managing Director and P&C Actuarial Practice Leader at Davies, formerly Merlinos & Associates. Frank has over 20 years of actuarial experience supporting insurers, self-insureds, private equity, agencies, and governmental entities in various risk and insurance issues. Frank resides in Atlanta, GA, and can be reached at [email protected].
Workers’ compensation has long been a prime candidate for predictive modeling and analytics, which can help to provide more clarity and confidence in the face of complex, long-tailed claims experience and shock losses.
As the Insurtech movement continues and greater startup capital enters the industry, it is no surprise that there are a bounty of SaaS and analytics firms offering a bevy of AI options.
While there are new types of WC-focused analytics firms popping up all the time (such as those addressing more niche issues or using specialized approaches to leverage video imaging and social media data), we will primarily look at the segment of the market that focuses on underwriting, pricing and claims-level solutions to answer two essential questions:
- How much of what’s being marketed as revolutionary AI is fluff?
- How can I tell which firm will be able to actually help me?
But First, a Message From Our Dictionary…
Before we delve into various market offerings, it’s important to understand what could be meant by a company when the term “AI” is used.
- Artificial Intelligence (AI) is usually defined as a computer’s ability to perform human-like tasks e.g. perception, reasoning, learning, etc.
- Machine Learning (ML) is a subset of AI and represents the way computers can learn how to do something without explicit instructions. Within ML, there is both supervised and unsupervised learning:
-
- Supervised learning covers the building of models based on a dataset that has both inputs and a desired output. The modeler will give some direction, but the computer is doing most of the computational labor. Classification and regression models fit into this category.
-
- Unsupervised learning covers the building of models based on a dataset that has only inputs and no identified output. The computer investigates and discovers patterns and groupings on its own. Examples of unsupervised learning include clustering, anomaly detection, dimension reduction techniques, neural networks and more.
In short, AI covers a lot of ground and a lot of different techniques. And because AI and ML are used interchangeably, it’s easy for people to get confused about what constitutes “true” AI.
For the purposes of this article, I define AI as models driven mostly by unsupervised learning.1
So, Where Is the WC Industry in AI Evolution? Is It Style or Substance?
While there’s no way to address every company and every solution offered, I’ve often found that when a SaaS or analytics firm says they have AI, the firm may more likely be offering a model that is developed using supervised machine learning, with a small portion that is more truly AI.
Often, the model is trained on a large proprietary data set that incorporates none or very little of your own data.
Such a model — which is developed independently from your own risk exposure, with no additional adjustments made — can be referred to as “off-the-shelf,” while models that are trained more on your own data can be referred to as “bespoke.”
Whether off-the-shelf or bespoke, WC models needn’t be fully unsupervised to make a significant improvement in your WC program.
Fair Enough. How Do I Find the Right “AI” Firm for Me?
One of the first things you can do is to make sure you have the right support to walk you through the procurement, evaluation, negotiation and operationalization process.
An internal stakeholder with previous experience is the most ideal, but existing relationships like brokers or actuarial consulting firms may also suffice. Less ideal and more costly would be a new external consulting relationship.
The next thing you can do with your analytics advocate is to understand the model being offered. Sign an NDA and insist that the vendor trying to win your business explain the inner workings of the model until you are truly comfortable that you are getting what you pay for.
If the vendor insists that the model design is proprietary and can’t be shared, you can surely still continue discussions, but I’ve found many firms are willing to share at least basic design concepts to help you distinguish between buying a Porsche 911 flat-six or a Toyota Corolla 4-cylinder CVT.2
Macho car references notwithstanding, the differences between models could include not just performance and handling but also model assumptions that may or may not be appropriate to your situation.
In a similar vein, an additional ask from the interested seller is for references and to show an understanding of your particular situation and industry.
WC claims, even in the same state and class code and the same demographic claimant, can vary considerably depending on the claims TPA and other parties involved.
If you don’t feel comfortable that the vendor understands you and your business pre-sale, one can only imagine the type of support one would get post-sale.
Discernment & Due Diligence
While there is marketing bravado throughout any industry, there are ways to discern whether an analytics firm’s offering is more style or more substance.
Having the right help is an important piece, but so is having honest discussions about the model with those who understand and can explain it.
As with everything that could be a 5-, 6- or even 7-digit spend, do your due diligence to make sure you’re getting the right return for your investment. &
[1] It is common for a single model to contain both supervised and unsupervised techniques.
[2] To each their own, but I’ve driven both and would definitely recommend the flat six.
