The Trust Deficit in Data

• Russell Sommers: “Statistics don’t lie, but liars use statistics.” The danger isn’t usually intentional fraud — it’s confirmation bias, pulling data to prove what you already believe instead of using it to find out what’s actually happening.
• The same dynamic that drives the industry’s “we can’t because” pattern gets amplified everywhere by social media and deepfakes, where trust itself becomes the thing that gets exploited.

AI as Enabler and Complication, at the Same Time

• Spero Zacharias framed the core tension of the episode: AI can finally solve for the data handcuffs the industry has lived with for decades — legacy systems, self-reported data nobody can verify, information trapped in unstructured text — but it introduces a new set of verification problems in the process.
• Intelligent document processing is finally unlocking the 80–90% of insurance data that arrives unstructured, but when AI gets the interpretation wrong, human critical thinking becomes more essential, not less: assume the recommendation is wrong until it’s been scrutinized.
• The panel drew a hard line between AI making a recommendation a human reviews and agentic AI taking an autonomous action — the stakes for verification change completely once AI is deciding, not just suggesting.

A Governance Model Built for Fields, in a World That Has Outgrown Them

• Traditional data governance assumes fixed fields and defined structures. The panel challenged whether that model still fits a world where AI-processed data doesn’t arrive pre-labeled and well-defined.
• One reframe from the conversation: shift governance thinking from which fields are required to which risk signals need to be captured — much of the data that matters for a claim or submission never reaches a structured field at all.
• The closed claim review, historically a batch process done after the fact, is being rethought as continuous, real-time monitoring — catching a bad AI recommendation while it’s happening, not two years and thousands of transactions later.

Data Ownership Has to Have a Name Attached

• The clearest agreement on the panel: data accuracy has to be business-owned, even when a technical or data function builds and maintains the systems around it.
• A chief data officer can build the infrastructure, but without a named owner accountable for a specific data element’s quality, governance models fail regardless of how well-designed they look on paper.
• A real example from the conversation: a claims system that used the same word, “death,” for three different drop-down values for 15 years, with no institutional memory of what distinguished them or whether staff picked the right one.

Standards as an Enabler, Not a Floor

• Spero Zacharias, drawing on his ACORD board chairmanship: standards exist to anchor an industry to something shared, not to give companies a minimum bar to clear and stop.
• State regulation, NAIC guidance, and ACORD standards are described repeatedly in the conversation as minimum requirements — the panel’s shared challenge to the industry is to build past the floor rather than manage down to it.
• Russell Sommers connected this directly to compliance: state privacy laws, the NAIC’s responsible AI principles, Colorado’s AI law, New York’s Circular Letter 7, and a new NAIC exposure draft on third-party AI use all stack into what he called an “unbelievably complex web” to track.

Cybersecurity’s Human Element

• Spero Zacharias pointed to the Scattered Spider social engineering attacks, where publicly available data was used to get two-factor authentication credentials reset at multiple insurance companies — a reminder that data trust failures aren’t only technical.
• The most recent Verizon Data Breach Investigations Report shows AI-enabled vulnerability exploitation rising sharply, even as the panel argued AI-assisted development should make software more secure over time as vulnerabilities surface faster.

An Audience Question on Industry-Wide Collaboration

• A question from the audience raised a challenge the panel embraced: with real-time data connectivity and API integration now possible, the industry has an opening to stop re-keying the same data across broker, insurer, and reinsurer systems, and instead capture it accurately once.
• The panel pointed to the London market’s Solutions Group as a model North America could learn from in building that kind of shared, collaborative data infrastructure.