The Last Mile of Claims: Why Correspondence is Structurally Harder Than Summarization
The dominant narrative of AI in insurance for 2025 focuses on compression. We are seeing a proliferation of tools designed to summarize 500-page medical records, classify FNOL photos, and flag potential fraud indicators. These are essential “read-only” or “read-then-suggest” workflows where the AI serves as a filter. If a model hedges or misses a minor detail in a file summary, the downstream human workflow often absorbs the noise. The stakes are internal, and the error is recoverable.
Claims correspondence inverts this pattern. When an adjuster clicks send on a Reservation of Rights (ROR) letter or a partial denial, they are performing an active legal act on behalf of the carrier. That document locks the carrier’s position. It is preserved in the claim file, discoverable in litigation, and subject to the scrutiny of state regulators. In the world of claims, correspondence is the “last mile” where automation moves from being a helpful librarian to a high-stakes drafting partner. This shift from reduction to synthesis is why correspondence is fundamentally more difficult to solve than summarization.
The Synthesis Challenge vs. The Summarization Trap
Summarization is a lossy process of reduction. It takes a large volume of data and distills it into a digestible format. Correspondence, however, is a process of synthesis. It requires an adjuster to pull from three distinct and often conflicting data silos: the specific facts of the loss, the precise language of the policy and its endorsements, and the statutory requirements of the jurisdiction.
This is why the “severity paradox” currently facing the industry is so taxing. While claim frequency has stabilized in some lines, severity is increasing, driven by social inflation and rising litigation costs. As we noted in our analysis of the Verisk Claims Report 2025, the legal stakes of every outbound letter have grown. A generic summary of a file does not protect a carrier from bad faith; a precise, timely, and policy-compliant letter does.
The Jurisdictional Maze
The complexity of correspondence is rooted in a patchwork of state-specific mandates that can turn a routine administrative task into a liability. In California, for example, the Fair Claims Settlement Practices Regulations do not just dictate what you say, but when you say it. Carriers are required to provide specific disclosures and brochures within strict timeframes from the initial notice of loss.
Across the country, these rules vary wildly. In Bismarck, North Dakota, an insurer’s failure to provide a proof-of-loss form within 20 days can result in the requirement being automatically waived, shifting the burden of proof entirely. In Maryland, if a sworn proof-of-loss is not collected within 15 days, certain recovery rights may be barred. In Rhode Island, there are strict prohibitions on using “final” or “release” language unless policy limits are paid or a specific compromise is reached.
These are not just formatting preferences. They are legal requirements where a single wrong word or a missed deadline changes the carrier’s legal position. This is why we created the Claims Correspondence Compendium, a public resource that maps these jurisdictional nuances. When you look at the sheer density of these rules, it becomes clear why generic Large Language Models (LLMs) struggle. They lack the “ground truth” of specific state statutes and policy-specific endorsements required for defensible drafting.
Drafting Speed as the Primary Lever
There is a common misconception that AI should solve the regulatory question by acting as an autonomous rule engine. In reality, the regulatory landscape is too fluid, and the human accountability required by the NAIC Model Bulletin on the use of AI is too high for full autonomy.
The real lever for productivity is not replacing the adjuster’s judgment, but collapsing the time it takes to draft the document. If an adjuster spends 45 minutes toggling between a policy PDF, a state statute guide, and a Word template, they have no bandwidth left for the critical thinking required for complex claims.
At Voltaire, we see this as the core of the problem. By automating the mechanical aspects of drafting, such as pulling policy exclusions or inserting state-mandated disclosures, we allow the adjuster to move from “doing the work” to “reviewing the work.” This approach recognizes that the adjuster is the licensee and the ultimate arbiter of fair settlement practices.
What Changes When Drafting Time Collapses
When the drafting burden is removed, two shifts occur within a claims organization. First, quality migrates upward on the most difficult files. Adjusters can give genuine attention to complex ROR letters, awkward partial denials, or catastrophic (CAT) follow-ups, rather than relying on standard templates that may not fit the facts of the case.
Second, carriers can absorb complexity growth without proportional increases in staffing. The regulatory environment is not getting simpler, and litigation is not decreasing. By focusing on the drafting layer, carriers buy back the capacity their teams need to navigate this complexity.
In the future of insurance, AI that actually works in claims will not be a black box that makes decisions in the background. It will be a tool that sits next to the adjuster at the moment of adjudication. It will handle the tedium of policy lookups and disclosure checks, leaving the judgment to the human. Correspondence is the hardest part of the workflow to automate because it is the most human part of the process. The goal is not to remove the person from the letter, but to give them the time to write it correctly.
See Voltaire in action — schedule a demo.
