Risk and Recommended Controls for Models and Artificial Intelligence Systems

Eric Bonnell, SVP, Manager - Technology and Asset Risk, Atlantic Union Bank

As insurance companies look to incorporate modern technology into their business plans, it is important to understand the risks and expected controls to protect maintaining models and Artificial Intelligence (AI) systems. There are potential consequences of not having sound models on which to base decisions. To address this, you must have a good understanding of the assumptions, rules, and data used within the models that make up artificial intelligence systems.

Virtual simulation of real-life events is a tricky concept. In theory, the number of variables for which you would account for within a perfect system would be infinite. As this is impractical, AI systems strive to reasonably operate within a degree of certainty by working with enough variables to cover the most impactful considerations. As this is an imperfect process, it is important to note assumptions to point out any significant variation due to the amount of historical data, the scope of data to be used, the level of certainty of the data, and how the data will be interpreted when the simulation is executed.

For example, how can a morbidity rate be simulated within a degree of certainty? It is important to have enough historical data to cover the potential values (e.g., age at death), a reasonably accurate distribution of likelihood (e.g., how many died at each age and cause to predict likelihood and timing), as well as the health conditions within the populations, the areas of concentration of each condition (i.e., location or other specific conditions under which a particular condition is likely), and many other variables as are understood to provide meaning and guidance within the model.

Know Your Assumptions

As actuaries can tell you, multivariable model analysis can be very complex. Understanding assumptions about the data and rules will assist in the prudent usage of models, including:
• The definitions of each data type and how well it represents the assigned variable or condition
• what key variables will be considered fixed or absent in the calculations
• the assumed level of quality of data and measurements
• the limits and potential outliers within the measured data and resultant model calculations

Establish Model Rules

Consistency and transparency are paramount to build a reliable model. Understanding the variables used, the conditions affecting each variable, and the calculations that will operate on each variable are all important to correctly operating the model and upholding the expected level or range of certainty of the results.

Those stepping into the world of model development and implementing AI technology should be aware of the risks around the quality and integrity of these systems

AI is in effect driven by code that establishes models. Whether your AI is driving customer policy underwriting or asking questions and providing guidance through a customer service portal, documenting, and testing the output and user experience of the program, given a representative sample of input data representing extensive conditions is prudent. This will not only help you to understand the program behavior and expected quality of results, but will also help you to identify and plan for upcoming improvement opportunities.

Procure Reliable Data

The term “GIGO”(i.e., garbage in, garbage out) is a term used to describe the outcome of poor data and code quality. In short, you get what you give. Errors or misunderstandings of data used by your calculations may result in driving poor decisions.If you do not scrub your data and test your code thoroughly with an abundance of test cases, you may generate questionable results which can be easily misinterpreted.

Also, the quantity of historical data will affect the level of certainty that you may expect. It may be okay if your model is based upon 10 years of financial data, given the understanding that there was a two-year economic correction within the data; only you can accept that level of model reliability or strive to procure a longer history of data to add more stability in calculations.

Putting it All Together -Assumptions, Rules, and Data

Artificial Intelligence is driven from this trinity: assumptions, rules, and data. These are the requirements that go into building and operating the model. Having clear and meaningful documentation to support the model is paramount to its sound operation within the expected level of certainty. This documentation also allows you to trace your steps when troubleshooting the model’s code and criteria and to facilitate further “what if” analysis and expansion of the model complexity in the future. Comprehensive and controlled testing of data and AI code results are very important to understand the model behavior and its expected level of reliability.

Conclusion

Those stepping into the world of model development and implementing AI technology should be aware of the risks around the quality and integrity of these systems. A comprehensive understanding of the assumptions, rules, and data within these systems will assist in gauging the reliability of these tools when making business decisions. Applying a high level of due diligence can protect the company from poor use of this technology and even provide regulators with a higher level of comfort, especially when this technology is used for making decisions that affect consumers and customers. Strong documentation will also support the company’s due care if challenged by any unforeseen litigation.