Git Out ‘Da Noise, Bring Home ‘Da Bacon!

Underwriting is one of the most critical functions of any insurance enterprise. At its core, it involves complex decision-making, often challenged by incomplete information, time constraints, and competitive pressures. The risk of error is always present. 

This article examines the inherent limitations of human judgment in underwriting. It explores how AI can effectively enhance, rather than replace, human risk assessment to improve consistency, fairness, and ultimately, your return on investment.

Errors in Judgment and Their Financial Impact 

Underwriters’ decisions must be accurate to ensure their company is profitable in the short term and solvent over the long term. The challenge, however, is that this decision-making relies on informed human judgment using data that can be incomplete, incongruous, or simply incorrect. This environment can lead to errors, which are often amplified by external pressures like deadlines and new business demands.

An error in judgment can result in a risk assessment that is either too liberal or too conservative relative to the “ground truth” of a case. Management often assumes these errors are offsetting—that they cancel each other out. Unfortunately, they are not; they are additive and cumulative. 

Your company will likely win underpriced business and lose out on overpriced business. Financially, you will probably end up with less revenue and lower profits than you otherwise would have. Even worse, you could suffer significant losses that permanently destroy value. AI-augmented underwriting may well be the key to mitigating these errors and their costly impacts. 

The Two Types of Error: Bias and Noise 

As detailed in the book Noise: A Flaw in Human Judgment by Daniel Kahneman, Olivier Sibony, and Cass R. Sunstein, our judgment is subject to at least two forms of error: bias and noise. 

• Bias is a systematic error—a consistent deviation from the true value. It can result from recency bias (favoring recent cases), anchoring (being disproportionately influenced by initial information), and confirmation bias (favoring evidence that supports your initial impressions). We all know of underwriting shops that have been characterized as overly aggressive or conservative; that's bias in action. 

  
  

• Noise describes random variability and is more nuanced. Different underwriters can arrive at different conclusions on identical cases using the same data. This can arise from factors like excessive coherence, workload fatigue, and even the order in which information is presented. 

  
  

These two types of error are independent, but they are equally weighted in determining the overall error in a decision. 

The Underestimated Impact of Noise 

Conceptually, this could be as “simple” as having 100 underwriters assess 100 cases using the exact same data. Their assessments would be recorded and analyzed for deviation from the mean, providing a clear indicator of the magnitude of noise in the system. While simple in theory, such an audit is impractical as it is costly in time, effort, and distraction. Besides, who has 100 underwriters to take offline for such an exercise?

Nonetheless, one unnamed insurance company reportedly did just that and was alarmed by the results. Their executives had assumed there would be some natural variation between underwriting decisions, expecting it to be around 10%. They envisioned a tight, controlled distribution of outcomes. 

Much to their chagrin, they discovered a variation that was five times greater than anticipated! One executive concluded that their noisy system had cost them hundreds of millions of dollars over time. As the authors of Noise point out, systems don't make multiple judgments on the same case in the real world; instead, they make single, noisy judgments on many different cases, which only adds complexity to assessing the degree of error. 

So, the question becomes: how do we reduce and manage our system noise? Can augmented underwriting be part of the answer? 

How AI Can Reduce Error and Empower Underwriters 

AI tools can dramatically improve underwriting outcomes by minimizing both bias and noise, empowering underwriters in several key ways: 

Harnessing Data Processing Power

Utilize AI tools to extract and surface structured insights from unstructured data embedded in Attending Physician Statements (APS), Electronic Health Records (EHRs), lab results, and other documents. This saves hours of manual review, which is itself subject to "scrolling" error. 

Establishing Consistency and Standardization

Present comparable cases in a comparable format. Once information is standardized, powerful studies can be performed that reveal hidden patterns and insights. 

Enabling Decision Support

Highlight missing or incongruous data, surface anomalies, and generate risk indicators that prompt an underwriter to dig deeper, thereby reducing the chance that a decision is made using faulty information. 

I believe this augmented approach will significantly reduce the range of outcomes and, in turn, reduce noise error. Crucially, the underwriter remains the final arbiter, with AI strengthening the very foundation of their judgments. 

Rethinking the ROI of AI 

When companies invest in innovative technology, they often project their J-curve and expected ROI based on time and expenditure savings. This is a limited view. 

I argue that if a company uses AI to augment its underwriting process, it can project a much higher ROI by factoring in the impact of error reduction. Your projections should reflect the immense positive impact of not winning “bad business” and winning more “good business.” This use of AI yields far more than just operational efficiency; it drives profitability. 

Conclusion 

Effective underwriting requires informed human judgment. When that judgment is paired with better data and decision support from AI tools, both risk assessments and financial results will improve. The winning organizations will be those that treat AI as a partner—not a replacement—to enhance their underwriting capabilities and build a more profitable, sustainable business. 

What are your thoughts on the role of AI in reducing noise and bias in your field? Share your experiences in the comments below. 

Author's Note

Please don’t think I am “picking on” underwriters. Noise in human judgment is not limited to underwriting decisions. It is endemic in many areas. As pointed out in the book Noise, it exists in medical assessments, child custody cases, personnel decisions, and, I might add, even weather forecasts. Please keep this in mind as you use your judgment in making your next big decision. 

Notes

¹ Kahneman, D., Sibony, O., & Sunstein, C. R. (2021). Noise: A Flaw in Human Judgment. Little, Brown Spark. The book reports on several noise audits, including the insurance company example mentioned in this article.