How AI Is Changing the Way Reserves Get Set

What you will find below:

• How Reserves Get Set Today
• What Traditional Methods Tend to Miss
• What Changes When AI Enters the Process
• Where Human Judgment Still Comes In
• Why This Matters for Claims Handling
• What to Ask Before Adopting a Reserving Model

Reserve accuracy shapes almost everything downstream on a claim, from authority levels to litigation strategy. This piece breaks down what changes when AI enters the reserving process, and where that differs from the methods most claims organizations still rely on today.

How Reserves Get Set Today

On most files, a reserve gets set early and revisited only when something on the claim changes. That initial figure is typically built from averages tied to claim type, injury severity or venue. Those averages come from large volumes of historical claims, grouped together on the assumption that a new claim in the same category will develop the way similar claims have developed in the past.

This method has been the industry standard for a long time, and for good reason. It is easy to apply and easy to explain to stakeholders who need to understand how a number was reached.

What Traditional Methods Tend to Miss

Because this approach relies on grouping claims together, it can be slow to register when an individual file starts behaving differently than the group it was placed in. A soft tissue claim that begins showing signs of surgical treatment is one example. A provider billing pattern that does not line up with the reported course of treatment is another. Average-based reserving is built to reflect these shifts once enough similar claims have shown the same pattern, but it typically can’t catch a single claim doing something unusual in real time.

Research published in 2023 compared traditional aggregate-level reserving against a machine-learning model built to evaluate claims individually. When claim-handling conditions shifted, the aggregate approach produced substantially larger errors, while the individual-claim model held up with greater accuracy.

What Changes When AI Enters the Process

Traditional reserving starts by placing a claim into a category and applying that category’s historical average. A machine learning model, on the other hand, evaluates the specific claim on its own terms, weighing a much larger set of variables at once: the injury and treatment pattern, the provider and billing history, the jurisdiction, prior investigation or utilization review findings, and dozens of other data points that a category-based average was simply never built to hold.

The model also does not wait for a scheduled review to update its estimate. As new information lands on the file — a new medical record, a bill review result, an updated IME finding — the estimate can shift immediately rather than sitting static until the next formal reserve review.

One 2025 study tested this against real claim data, comparing machine learning models directly to actuaries’ own reserve predictions across several major insurance lines. The models produced lower error rates on standard accuracy measures, and when actuaries’ own predictions were fed into the model as additional input, accuracy improved even further. This suggests the strongest results come from combining the two rather than treating either as sufficient on its own.

Where Human Judgment Still Comes In

A bulletin published by the Society of Actuaries (SOA) Research Institute in 2026 points to human judgment as the piece that still provides context, accountability and an understanding of claims-specific circumstances a model cannot fully replicate on its own.

Treating AI output as a second opinion gives claims organizations a practical way to use it. A model can flag a reserve estimate that has drifted from what the underlying claim data supports, giving the adjuster or examiner a reason to take a closer look before that gap becomes a larger problem later. The final call still rests with the person handling the file.

Why This Matters for Claims Handling

For adjusters, claims managers, and SIU professionals, this shows up in tangible ways. A reserve built from claim-specific detail holds up better under internal audit and reinsurance review. Authority requests are easier to support when there is a clear basis behind the number. And when a claim starts to deviate from its expected pattern, that signal surfaces while there is still time to act on it.

The quality of an AI-driven reserve estimate depends heavily on the quality of the claim-specific data feeding it. Verified information from investigations, medical record reviews, and independent medical exams gives both the model and the person reviewing its output something concrete to work from as opposed to relying on category averages alone.

What To Ask Before Adopting a Reserving Model

For a claims organization evaluating a reserving tool or vendor, a few questions tend to separate a genuinely useful model from one that just adds another layer of complexity:

• How often does the model update its estimate, and what triggers that update?
• What specific data feeds the model, and how much of it comes from verified sources versus self-reported claim notes?
• Where does human review sit in the process, and does the adjuster/examiner see the reasoning behind a flagged reserve or only the number itself?

The answers to these questions usually reveal whether a model is built to support claim handling or simply automate a piece of it.

Curious how better reserve accuracy fits into a broader claims strategy? Connect with our team to talk through how investigations and medical management support smarter decisions earlier in a claim’s life.

Sources:

• Actuarial Intelligence Bulletin. (2026, May). Society of Actuaries Research Institute.
• Okine, A. N.-A. (2023). Individual-Level Loss Reserving and Environmental Changes. Variance, 16(1).
• Zheng, Y., Li, X., & Zhong, W. (2025). An empirical study on machine learning-based future loss prediction for insurance firms. Applied Economics, 1–14.