Making Risk Flow

The episode delivers solid, operationally relevant insights about geospatial data in underwriting - confidence scoring, first-floor elevation applications, roof condition specifics tied to loss ratios, and the scale/accuracy tradeoff. However, substantial portions consist of company positioning, general industry commentary, and soft closing remarks that add little new substance. The most dense section is the discussion of probabilistic vs. deterministic models and their real-world implications.

“there is 20% of the time or more that it is a hallucination and the answer looks great and people want to trust it because it sounds good, it's convincing, but it could be total nonsense”

“when you look at those probabilistic ANSWERS, there is 20% of the time or more that it is a hallucination”

While the specific execution is differentiated (Geox's deterministic model with confidence scores, the Australia flooding/building-lifting anecdote), the core frameworks are standard in InsurTech: computer vision, geospatial intelligence, data-driven underwriting, and buy-vs-build economics. The Australian anecdote is genuinely novel, but most other points recycle common vendor narratives about scale, accuracy, and automation.

“in Australia, after a flooding event, they will actually lift up the building. They will take that claim and mitigate the flood risk by lifting the entire building and adding to that first floor elevation”

“We are turning that probabilistic answer into a deterministic answer. We are providing confidence”

Guests are highly relevant: Jacob Grob brings 20 years of insurance data sales experience and deep market knowledge (CoreLogic, Florida expertise, competitive landscape). Isaac Lavy is CEO of Geox with demonstrable scale credentials (130M quarterly updates, 100% coverage in two mature markets). Both are practitioners in their domain, not pure thought-leaders. However, as company leadership discussing their own product, there is inherent bias and limited external validation perspective.

“I've been in the insurance space selling data to insurance companies for 20 years. Twelve of that was at CoreLogic”

“So in the last quarter uh, we ran update on 130 million building footprints to compare to what is right now in the market”

The episode includes concrete data points: 130M quarterly updates vs. competitors' 90M annually, 20% LLM hallucination rate, 3x frequency and 1.8x severity for damaged roofs, 20-40% discounts for excellent roofs, two-hour underwriting timelines, and specific peril examples (gable walls facing coastlines, first-floor elevation in flood zones, defensible space in wildfire zones). However, many claims lack attribution or independent verification, and the cost/ROI comparison to in-house solutions is vague.

“in the last quarter uh, we ran update on 130 million building footprints”

“the frequency of property that claimed the insurance carrier that Gox says uh, this property had damaged roof condition is three times or higher. The severity is around 1.8 to compare to the average policy”

The host (Jay Carding) asks competent setup questions and encourages both speakers to elaborate, but rarely pushes back on claims, challenges assumptions, or explores tension. Follow-ups are generally soft and affirming ("Yeah, and immediately you can kind of see the benefits"). There is no productive disagreement, no hard questions about limitations, false negatives, or competitive threats. The conversation reads more like a structured vendor showcase than investigative dialogue.

“Yeah, and immediately you can kind of see the benefits there”

“Yeah. And I think there's a kind of key theme that's touched on there”