The AI Crossroads in Claims: A Practical Playbook from CLM

Claims leaders are being asked to do two things at once that can feel fundamentally at odds: move faster with AI and automation, while preserving the human judgment required for fair, compliant, and defensible outcomes.

That tension is now at the center of claims transformation. Carriers are operating in an environment shaped by rising complexity, growing documentation burdens, litigation risk, regulatory scrutiny, and constant pressure to improve efficiency without degrading the policyholder experience. AI has obvious appeal in that context. It can summarize records, surface patterns, draft communications, organize information, and reduce manual work across the claim lifecycle.

But claims investigations are not just information-processing exercises. They require context, skepticism, fairness, accountability, and reasoning that can stand up later in a file review, market conduct exam, deposition, or courtroom. That is why the most important question in claims is no longer whether AI can help. It is how to use AI in ways that strengthen investigations without weakening the judgment behind them.

That was the shared throughline of a recent CLM panel focused on AI, claims operations, and the future of investigation work.

The panel featured David Fernandez, Chief Claims Officer at Kingstone Insurance, where he oversees claims strategy, customer service, and operational efficiency; Eric Brandt, President of Second Horizon Solutions and a longtime claims executive who has led claims organizations at major carriers and now advises boards and leadership teams on technology and strategy; Matthew Scarfone, shareholder at Colodny Fass, whose practice spans insurance litigation, regulatory, and transactional matters; Steve Ellis, Vice President of Liability Practices at Sedgwick, bringing the perspective of claims operations at scale; and Arvind Sontha, Founder and CEO of Kyber, who works with P&C carriers on compliant, consistent, and personalized claims communications.

It was a cross-functional panel spanning carrier operations, legal defensibility, fraud and investigative strategy, claims administration at scale, and workflow automation. While each speaker emphasized different risks and opportunities, the panel’s message was remarkably consistent: AI can make claims organizations faster, more consistent, and more scalable, but it should be deployed as decision support, not decision replacement.

What follows is the panel’s playbook for using AI to modernize claims investigations while preserving defensibility.

The Panel’s View: Use AI to Strengthen the Workflow Around Judgment

The panel’s perspective was not anti-AI. If anything, it was notably pragmatic about how valuable AI already is becoming inside claims. The group discussed document summarization, recorded statement support, fraud detection, file review, data extraction, legal research, communication drafting, and workflow acceleration as real use cases with real operational value.

But the panel drew a sharp line between automation that supports professionals and automation that tries to replace them.

As Arvind Sontha put it, claims leaders should think about automation across two dimensions: task frequency and task risk. The best starting points are usually the high-frequency, lower-risk tasks around the core decision, not the judgment-heavy decision itself. That framing helps explain why the panel kept returning to the same principle. AI should reduce friction around the work of investigation so claims professionals can focus more of their energy on causation, fairness, coverage, severity, negotiation, and defensibility.

That shared view matters because it reframes the role of AI in claims. The goal is not end-to-end autonomous adjudication. The goal is to build a stronger operating system around human expertise.

Where Claims Leaders Get AI Adoption Wrong

Again and again, the speakers surfaced the same mistakes: beginning with the technology instead of the problem, trusting outputs too quickly, chasing speed before process quality is ready, underestimating legal exposure, and treating governance as something to solve later.

David Fernandez captured the first issue most directly when he said teams should “fall in love with the problem, not the solution.” That line could serve as a filter for nearly every AI initiative now entering claims organizations. When leaders start with the tool, they tend to optimize around novelty. When they start with the operational bottleneck, they build for value.

The panel also made clear that speed is not inherently progress. Faster workflows can still produce worse outcomes if they increase error rates, weaken review discipline, or create records that will be difficult to defend later. In that sense, AI does not remove the need for operating judgment. It raises the stakes for having it.

This is what makes the current moment feel like a crossroads. Claims teams can use AI to improve investigation workflows in meaningful ways. But they can also create new forms of risk if they automate too aggressively, document too loosely, or over-trust systems they do not yet fully govern.

Start With the Claims Problem, Not the Technology

The panel’s first practical lesson was simple: do not start with “how should we use AI?” Start with “what friction in our claims operation is worth solving?”

That distinction sounds small, but it changes the entire shape of adoption. A technology-first posture tends to produce pilots in search of a business case. A problem-first posture produces use cases tied to actual outcomes like faster triage, lower documentation burden, better file quality, earlier fraud escalation, improved consistency, or greater adjuster capacity.

Fernandez’s point about loving the problem instead of the solution was grounded in experience. Too many modernization efforts begin with excitement about the tool and only later confront the realities of timing, workflow design, adoption, and measurable value. By contrast, starting from the claims problem forces leaders to identify where the pain actually sits: repetitive administrative work, documentation overload, fragmented information, delayed handoffs, inconsistent communications, or missing data buried in the file.

The panel’s shared recommendation was to begin where the pain is concrete and recurring. High-volume work around intake, documentation, file review, communications, and investigation support is often a stronger first target than judgment-heavy decisions. These workflows tend to be easier to measure, easier to validate, and easier to improve without compromising accountability.

That is also where leaders can learn the most. A contained use case reveals how the organization responds to change, what level of human review is actually needed, how much trust the workflow earns, and where the next layer of operational improvement should sit.

Use AI Where It Improves Investigation Throughput and Clarity

The panel was especially useful in naming the kinds of investigation work AI can strengthen today.

Across the discussion, the strongest use cases were not abstract. They were practical and adjacent to the investigation itself: summarizing long files, organizing claim materials, extracting key information from unstructured records, accelerating reassigned-file review, transcribing and summarizing recorded statements, surfacing potential fraud indicators, supporting communication drafts, and improving how information gets routed or escalated.

Matthew Scarfone described one of the clearest examples when he explained the value of using AI on “voluminous records, narrowing it down so that you’re not relying solely on the summary, but it’s giving you a starting point.” That idea captures the panel’s broader logic well. AI is most helpful when it reduces the time spent finding the signal, without pretending the signal no longer needs expert review.

Other panelists reinforced the same dynamic from different parts of the ecosystem. Steve Ellis highlighted the operational value of AI-generated file summaries for reassigned claims, where time-to-context can shrink dramatically. Eric Brandt pointed to fraud-oriented use cases like image forensics and altered-document detection, where AI can help insurers respond to a threat landscape that is itself being reshaped by AI. And Arvind framed workflow automation as a way to remove non-core burdens from adjusters so they can spend more time on the analytical and empathetic work that actually defines strong claims handling.

Taken together, the panel’s message was not just that AI can make work faster. It was that AI can improve the structure of investigation work by helping teams find, organize, and act on the right information sooner.

Keep Human Judgment at the Center of the Investigation

If there was a non-negotiable principle in the discussion, this was it.

Claims investigations still require humans to assess credibility, apply policy language, understand causation, evaluate fairness, interpret nuance, and make decisions they can stand behind. AI can support that work, but it cannot absorb the accountability that comes with it.

The panel approached this from multiple directions. Scarfone focused on the legal risk of allowing AI outputs to become de facto decisions. Ellis focused on the behavioral risk of using AI “like Google,” where teams accept outputs too casually instead of treating them as tools that require thoughtful prompting, review, and verification. Audience participants raised the development risk that younger professionals may lose decision-making muscle if organizations let the machine do too much of the reasoning too early.

Those concerns all point to the same operating truth: investigations are not just about reaching an answer. They are about reaching an answer in a way that is fair, accountable, and explainable.

That is why the panel kept returning to the boundary between assistance and adjudication. AI can classify, summarize, recommend, and flag. But when a claims decision has legal, regulatory, financial, or consumer consequences, a qualified human still has to evaluate the facts, apply the policy, and own the outcome.

For claims leaders, this is not a conservative constraint that limits innovation. It is the discipline that makes innovation usable in a regulated environment.

Make Defensibility Part of the Workflow Design

The panel’s strongest throughline was not really about automation at all. It was about defensibility.

As AI becomes more embedded in claims operations, carriers will face increasing pressure to explain what the system did, what information it surfaced, how it influenced the workflow, what human review occurred, and why the final outcome was justified. That pressure can come from litigation, regulatory review, internal audit, market conduct activity, or broader scrutiny around fairness and bias.

Scarfone was especially clear here: defensibility cannot be bolted on after the fact. It has to be designed into the workflow. If AI helps summarize a file, recommend a next step, or generate a communication, the carrier needs to know how that output was used and how the human decision-maker engaged with it.

Eric Brandt underscored the same need in simpler terms: “traceability, transparency, verification.” Those three words are useful because they cut through the hype. Claims leaders do not need systems that merely feel advanced. They need workflows that can be explained.

That has practical implications for product design, file documentation, and vendor evaluation. The panel repeatedly implied that high-stakes claims workflows need more than strong output quality. They need traceable source material, preserved rationale, review checkpoints, and an operating record that can survive scrutiny later.

A useful internal test emerged from the conversation: if the organization would be uncomfortable explaining the workflow to a regulator, plaintiff’s counsel, or a judge, the workflow is not mature enough yet.

Govern Early So You Can Scale Safely

The panel’s final lesson was about sequence.

Many carriers understandably want to know where AI is going next. But the panel’s advice was not to chase the most ambitious use case first. It was to start in a way that teaches the organization how to scale responsibly.

Arvind recommended beginning with low-hanging fruit not because the use cases are trivial, but because they let the organization learn. A focused deployment reveals how people respond, where validation is needed, what oversight is required, and whether the workflow actually creates value. Those early lessons then become the foundation for broader adoption.

That sequencing is what turns experimentation into governance. Instead of treating governance as a drag on innovation, the panel treated it as the mechanism that makes sustained adoption possible. Human-in-the-loop review, clear thresholds for escalation, auditability, vendor accountability, change management, and cross-functional input from claims, legal, compliance, SIU, IT, and operations are not signs that an organization is moving slowly. They are signs that it intends to move without creating unnecessary exposure.

This was also where the panel’s multi-speaker format became especially useful. Different participants emphasized different risks, but they all pointed toward the same conclusion: the organizations that benefit most from AI in claims will not be the ones that automate the most. They will be the ones that create the clearest operating rules for where AI helps, where humans decide, and how the workflow stays explainable.

A Practical Playbook for Claims Leaders

The panel’s discussion ultimately points toward a practical operating model for claims organizations navigating AI adoption.

1. Start with a claims problem that is concrete, recurring, and operationally expensive.
2. Use AI first where it improves speed, clarity, and consistency around the investigation rather than replacing the judgment inside it.
3. Preserve human accountability for decisions involving coverage, fairness, severity, settlement, or denial.
4. Build traceability, transparency, and documentation into the workflow from the start.
5. Treat early AI deployments as governance exercises, not just productivity experiments.
6. Scale only after the organization can explain how the workflow works and why it can be trusted.

That is the real opportunity in front of claims leaders now. AI does not need to replace the human core of claims to create major value. It can reduce friction, improve consistency, increase investigative leverage, and help teams move faster through the parts of the workflow that do not deserve to remain manual.

But the carriers that benefit most will be the ones that modernize with discipline. They will automate around judgment, not away from it. And in a regulated industry, that distinction will matter more and more over time.

‍