AI 5 min read
The next wave of AI adoption in insurance claims starts with decisions, not workflows
Claims teams are under pressure to settle faster, control leakage, reduce handling costs, and improve customer experience. At the same time, experienced talent across the insurance sector is becoming harder to replace, increasing the value of specialist capacity.
AI is already helping automate parts of the claims process, but turning that progress into value at scale remains difficult. Capgemini’s 2026 P&C research found that 60% of insurers remain in the exploration or proof-of-concept stage, while only 10% are successfully scaling AI. BCG similarly found that just 38% of P&C insurers are generating value at scale from AI across core workflows.
For simpler, lower-variance claims, straight-through processing is achievable. Where coverage, evidence, liability, and settlement paths are clear, much of the claim can move without specialist intervention.
The harder challenge is the claims that fall outside that path. Ambiguity, conflicting evidence, missing context, or specialist judgment make full automation less reliable. In those cases, the more useful question is not how much of the workflow can be automated with AI. It's which decisions within it matter most. By decisions, we mean the points in the claims process that determine the outcome.
Those decisions are the most useful starting point for AI optimization.
Where the value sits: the decisions that shape the claim
Decisions in the context of insurance claims can be thought of as the concrete calls that determine what happens next. Is the loss covered? Where should the claim be routed? Does it warrant a fraud referral? Is the available evidence sufficient? How should the claim ultimately be resolved?
Naturally, some of these decisions carry much more weight than others. Coverage is an obvious example because it sits so early in the claim, but the same is true of decisions around routing, investigation, evidence, settlement, and recovery.
What makes these decisions so consequential is the effect they can have on everything that follows:
- Better routing can reduce unnecessary handoffs.
- More precise fraud referral can protect investigation capacity.
- Earlier clarity on evidence can avoid repeat requests and delay.
- Decisions around repair, settlement, and recovery can materially affect claim severity and cost.
Looking only at the workflow can obscure this difference in leverage. A claims workflow may contain dozens of steps, but improving each of them does not create equal value. Starting with the decision helps isolate the points where better information, earlier intervention, or more consistent execution can materially change what happens downstream.
That is what makes the decision a useful unit of optimization. Automating a task removes effort from that task; improving a consequential decision can influence what happens across the rest of the claim.
Agentic AI can change the work around the decisions that matter most
Agentic AI is particularly useful around consequential decisions because the work required to reach them is often not fixed. What needs to be investigated next depends on the evidence available, what remains unclear, and whether additional information is required. More predictable parts of the process can remain rules-based or conventionally automated.
The challenge is often getting the file into a state where a decision can be made reliably. Relevant context may be spread across the claims management system, policy administration system, document repositories, correspondence, third-party data sources, and previous actions on the file. Agents can help bring that context together, compare evidence, investigate discrepancies, and complete approved follow-up actions as the picture develops.
Crucially, that does not mean handing the consequential decision itself to AI. For higher-stakes decisions, the agent can produce structured outputs that capture the relevant evidence, findings, and unresolved questions. Deterministic rules can then consume those outputs to apply thresholds, render a decision, or route the claim for human review.
One German-based insurer uses agents to classify claims documents within the context of its own policies and guardrails. The agents can extract relevant data, identify any discrepancies in customer submitted documents, categorize the claim, and flag the cases that require further human review.
The opportunity is therefore less about making an entire workflow agentic and more about using agents to prepare the context around consequential decisions. Deterministic logic and people still govern how decisions are ultimately made.
Start with decisions where the conditions are right
A useful starting point with applying agentic AI to the claims process is to assess important decisions against four distinct questions:
- How much does the decision impact the outcome? Decisions that affect subsequent handling, loss, customer outcome, or specialist involvement offer more potential value than those where the impact ends with the immediate task.
- How variable is the path to the decision? Decisions that require multiple sources, unstructured evidence, investigation, or different follow-up actions depending on what is found are stronger candidates for agentic support. Where the sequence is already predictable, conventional automation may be sufficient.
- How much specialist effort is spent preparing for the decision? If experienced claims professionals repeatedly spend time reconstructing history, collecting information, or resolving routine uncertainty before they can apply their expertise, there may be an opportunity to move more of that work earlier.
- What should the agent be allowed to do independently? Some decisions may be suitable for an agent to prepare or recommend, while execution remains governed by deterministic rules, thresholds, approvals, or human ownership. Coverage denials, higher-value payments, reserve changes, and other consequential actions may require different controls depending on the insurer and jurisdiction.
Together, these questions help distinguish between work that should remain straight-through or rules-based and decisions where agentic AI can add more value. They also create a clearer basis for prioritization than simply starting with the most manual or complex workflow.
The result is a more deliberate AI roadmap: start with the decisions that matter, understand what makes them difficult today, then determine which parts of the surrounding work are best handled through rules, automation, AI, or an agentic approach.
For example, one insurer uses agents to standardize evidence and apply consistent checks. Straightforward cases progress while nuanced issues are routed for specialist review. Each step remains traceable to the original evidence, and specialist feedback can refine prompts, rules, and decision logic over time within defined guardrails.
In a regulated industry, every step has to be reviewable
For regulated insurers, preparing a decision is only part of the requirement. It must also be possible to understand what informed it, which actions were taken, which rules or thresholds were applied, and where human oversight entered the process.
Those expectations exist on both sides of the Atlantic. For example, in Europe, the FCA’s Consumer Duty requires firms to act to deliver good outcomes for retail customers. In the US, insurance remains primarily state-regulated. The NAIC’s Model Bulletin on AI sets expectations around governance, risk management, testing, documentation, and regulatory oversight for AI-supported insurance decisions.
For agentic claims processes, this makes the architecture around the model especially important.
An agent may be able to investigate a claim, assemble evidence, or recommend a next step. But it shouldn't necessarily be authorized to independently deny coverage, release a high-value payment, or make another consequential determination. Those boundaries should sit outside the model and remain explicit.
Better claims decisions create value beyond the individual file
Claims is one of the clearest places an insurer sees how risk behaves in practice. Coverage questions, fraud patterns, liability outcomes, settlement behaviour, and recovery results all reveal where assumptions are holding and where they are beginning to shift.
When the context and outcomes around those decisions are captured more consistently, that learning becomes useful beyond the claims function. Underwriting can see where risk may be changing, product teams can identify recurring friction in policy design, fraud teams can refine the signals that matter, and portfolio or reinsurance teams can respond to emerging loss patterns earlier.
That gives decision-level optimization a broader strategic value. The immediate benefit is better handling of complex claims and more effective use of specialist expertise. Over time, it also creates a stronger feedback loop between claims and the rest of the insurer.
Workflows will continue to matter. They are how claims operations run. But they are not always the most useful place to start when deciding where AI can create the greatest value. Starting with the decision makes it easier to identify where better context, adaptive investigation, deterministic controls, and human expertise need to come together.
For claims leaders, that is the shift: optimize the workflow around the decisions that matter, rather than treating the workflow itself as the unit of AI transformation.
