A recurring theme from the Digital Insurance Summit, and why execution, not strategy, is what closes it. By Justin Goff, Director of Technical Delivery, Hylaine.
Picture the moment a claim begins. A policyholder has just been in a car accident. No one is seriously hurt, but the car is not drivable, and the stress is immediate. They file a first notice of loss, and then they wait. A carrier usually makes first contact within a day or two. The wait that matters is for the conversation after that, where an adjuster confirms the coverage that applies and walks the customer through an estimate. Reaching that point means assessing the damage, checking the policy, and preparing the estimate, mostly manual work that can stretch across a week or more. Every one of those days carries a cost: operational expense, customer anxiety, and the quiet erosion of trust that makes a policyholder shop around at renewal.
That lead-up is exactly the kind of work agentic AI is built to compress, which is why claims repeatedly emerged at the Digital Insurance Summit as one of the most promising places to begin. The work is well understood, the risk is manageable, and the return is easy to measure. Working behind the scene immediately after a first notice of loss, AI agents can assess the damage, identify applicable policy coverage, and prepare a draft estimate for a human adjuster to review and refine before speaking with the customer. The result is a dramatically shorter wait time, reducing what is often a week-long delay to just a day or two.
So why are so few carriers actually running this at scale?
The answer was the most honest theme of the entire event, and it remains largely unresolved across the industry. Insurers cannot simply point their claims data to a foundation model and let them talk. Compliance will not allow it. Sending regulated policyholder data to an external AI provider, or standing up ungoverned models alongside sensitive systems, is a non-starter in an industry accountable to state regulators, HIPAA, and an evolving patchwork of federal mandates. Most organizations pursuing AI at scale find themselves caught between the pressure to move quickly and the obligation to maintain strict governance. In practice, choosing speed at the expense of compliance is not a viable option.
That reality is forcing the conversation beyond AI strategy and into execution. “Adopt AI responsibly” may be a useful principle, but it offers little guidance when compliance teams begin evaluating real-world implementations. What organizations need is a practical framework that makes innovation and governance compatible. Increasingly, that framework is taking the form of a federated, walled-off environment for AI compute—an architectural approach designed to keep sensitive data protected while still enabling AI to operate at scale.
Unification & Compliant AI Compute Without Moving The Data
The idea is simpler than it sounds. Rather than consolidating sensitive data into one place, or shipping it to an external model, a unified analytics and AI layer connects to data where it already resides.
Individual teams keep their own access-controlled environments. A central compute layer connects to those sources through federated queries, so the analysis travels to the data instead of the data traveling to the analysis. Foundation models run inside a security-assessed environment rather than over an open connection to an outside provider. The retrieval that powers an AI assistant happens in that same federated way, which means a model can reason over claims data without that data ever physically leaving its compliant home.
We partnered on a build like this for a Fortune 25 managed care organization, using a federated data science platform with a centralized compute layer, a cross-cloud gateway to reach proprietary models, and catalog-level lineage tracking across otherwise separate sources. It scaled well precisely because it was designed to be reused, not rebuilt from scratch for every team. This pattern was pressure-tested in production and held up under real compliance scrutiny–and that is the bar that matters.
Where The Execution Framework Earns Its Keep
A walled garden on its own is just infrastructure. What enables organizations to scale is treating it as a reusable template within a structured, phased execution model–one built to address the practical questions teams encounter as they move from pilot to production: how to organize the work, sequence implementation, automate processes, and maintain governance along the way. The first phase focuses on defining the strategy for a specific use case, including the AI governance, risk, and adoption plans required to deploy AI tools. The second phase is where many organizations fall short. It centers on standardization and automation, creating reusable capabilities that eliminate the need to start from scratch with every new initiative. Once a compliant compute environment has already cleared security, identity, and access reviews, individual teams can build on that foundation rather than recreating it. Claims teams, for example, can leverage a centrally governed model registry and approved controls that have already been vetted by platform, security, and compliance stakeholders, dramatically accelerating deployment while maintaining oversight.
The effect on time to value is significant. Instead of spending months navigating governance reviews and infrastructure requirements, teams can begin with a foundation that is already in place. Because the most complex and time-consuming work have been completed once and designed for reuse, new initiatives move from concept to deployment much faster. In the claims example, teams can use approved models within the governed environment, apply them to data that remains within their own tenant, and build applications inside compliance-approved guardrails from day one.
The third phase focuses on enablement, ensuring teams can adopt and scale AI regardless of their starting point. Support is tailored to each team’s capabilities and needs. Teams with strong engineering resources may require only architectural guidance and periodic reviews, while more business-oriented teams may benefit from hands-on implementation support and shared production responsibilities. The goal is the same in either case: to ensure that a lack of specialized AI expertise does not become a barrier to progress.
The Compounding Return
The payoff extends well beyond first use case. Once a team has stood up an agentic workflow inside that federated environment, the compliant foundation is laid. The next agent, and the one after that, becomes far easier because the orchestration, the models, and the guardrails are already in place. The first use case is the hardest because it establishes the foundation. After that, progress accelerates and the returns begin to compound.
For the policyholder waiting after an accident, none of this infrastructure is visible, and that is exactly the point. What matters is that the first meaningful conversation happens in days rather than weeks, supported by accurate information that has been reviewed and validated by a human adjuster. Claims are resolved faster, payments arrive sooner, and uncertainty is reduced. For carriers, the benefits are equally tangible: lower manual costs, greater operational efficiency, and stronger customer relationships. That was the case repeatedly made throughout the Digital Insurance Summit. The industry already understands the strategic opportunity AI presents. The real challenge, and the real differentiator, is execution.
Be the first to comment