Use Data Responsibly in Insurance Underwriting and Claims Without Hurting Results

In order to achieve the right balance when it comes to regulating insurance AI with predictive capabilities, it is necessary to make transparency in feature engineering a priority, and keep the complexity of the model in check. Too often, this creates a black box that somehow ends up using wrongful bias proxies for the protected classes. The most reliable systems are the ones that ensure mapping every data input to a specific and useful risk factor which can be substantiated by a human underwriter. Relying only on correlation without deep understanding of cause and effect creates plenty of consistency problems with regulators, calling for their attention.
In our digital transformation projects with enterprise clients, we have chosen Explainable Output instead of Predictive Margin as a principle. Though complex and opaque models may give a small positive increase in predicting power, this is usually outweighed by regulations and associated risks if the models prove to be unable to explain the outcome of a certain decision. Now we are focused on developing systems which explain every automated adjustment or flag in simple words. This way, we will be able to provide an answer to the regulator when they want to know why a decision was made and point out specific features that played a role in a decision instead of algorithmic weights.
The main thing in the process is to create human in the loop strategy for decisions with a strong impact. When a person is required to approve or decline the claim denied by AI or the increase of premiums with the help of AI technology, we make sure the technology remains in the role of a decision-making assistant rather than the judge. This helps us not only to perform consistently but also earn trust of the customers who need to be sure that their assets are well protected with the help of both technology and a rational human mind.

When deciding which external data sources or AI outputs to use for underwriting or claims, I focus on whether the signal ties to stable, explainable drivers in our claims data over a multi-year window. Practically, we require two to three years of claims history and checks that any new input relates to observable cost drivers such as pharmacy trends or concentrated large claims. The principle I adopted is simple: data should drive decisions, not fear. We formalized that by modeling outcomes before making changes and committing to quarterly claims reviews so every decision can be traced back to the data that motivated it.

The principle we adopted is that no field gets used in a decision unless we can show where it came from and how much to trust it. I come at this from the provider side of claims: QuickIntell builds AI agents that handle eligibility, prior authorization, claim follow-up and denials for medical practices, so we sit across the table from insurers every day and face the same explainability problem in reverse.

In practice, every value our agents act on carries its source (the payer document, portal screen, call transcript or system record), a timestamp and a confidence score. The rules are then simple: act above a threshold, route to a person below it, and always show the reviewer the source. That one design choice does more for fairness and privacy than any model tweak, because it forces you to notice when a prediction rests on a stale, second-hand or inappropriate input. If you can't name the source of a field, it shouldn't be influencing a claim or a rate.

It also changed how we talk to customers and regulators. When a hospital's compliance team asks why an agent flagged a claim, we don't explain a model; we show the record, the evidence it cited and who approved the action. For insurers, I'd apply the same test: strong predictions are the ones you can defend line by line, not the ones with the best aggregate score.

As insurers increasingly adopt and scale AI usage, it's essential they ensure they're adopting a provably compliant, effective system of guardrails that mean they do actually move faster through implementing AI tools, but crucially that they don't lose anything in terms of governance or auditability.

Two traps I'll mention:

1. It seems most have tried to incorporate AI models into their workflows and then thought about governance second, or tried to consider both deploying AI and governance at the same time. Many companies in the financial services space have been in this situation recently. It's important that insurers avoid this trap, because it's a lot harder to put the governance layer in place after the fact. Starting with what you'll permit models to do makes this much easier.

2. Like many companies in the enterprise context, insurers often have had to get by relying on probabilistic enforcement of good policy, meaning they have to live with not knowing for sure if every action is fully and provably compliant. Insurers should demand their software vendors give them consistency, such that every action taken by an AI model is either acceptable or not allowed at all.