Insurance Benefits More From Traditional AI Than From GenAI. Almost Nobody Understands the Difference.

David Gritz and Tony Lew met at a consulting workshop in early 2019, both running independent consulting practices at the time. David brought product management depth from working with tech companies and an exit from a prior insurtech venture connected to SVIA on the West Coast; Tony brought product strategy — figuring out what to build in the first place. They decided to partner, working across several startups before settling on insurance as the vertical that stuck. David’s pitch to Tony, who had only modest insurance exposure at the time: it sounds boring, but it’s genuinely complex, and — unusually for a vertical this dense — the industry is actually doing good in the world, helping people and companies manage risk and prepare for what’s ahead.

Five years later, InsurTechNY runs a genuinely wide portfolio: an annual flagship conference (this was their fifth), a growth-stage accelerator, an early-stage pitch competition, MGA Lab (an accelerator specifically focused on helping startups build MGA businesses), a fund, and roughly 20 events a year total — including InsurTech Slopes, a ski and snowboard gathering David runs, now held twice annually (Verbier, Switzerland and Big Sky, Montana).

In Episode 113 of InsurTechTalk, David, Tony, and I covered why traditional machine learning — not generative AI — is where insurance captures the most real value, how analytics differs meaningfully from raw statistics in underwriting, and what to expect from the AI- and analytics-themed InsurTechNY Spring Conference.

About David Gritz and Tony Lew

David Gritz and Tony Lew are co-founders of InsurTechNY, a New York-based insurtech ecosystem organization running a flagship annual conference, multiple accelerator programs (including MGA Lab), an early-stage startup competition, a venture fund, and roughly 20 events a year — including InsurTech Slopes, a ski/snowboard networking event now held in both Utah/Colorado and Switzerland/Montana. Both founders previously ran independent consulting practices before partnering in 2019, with David’s product management background and prior insurtech exit complementing Tony’s product strategy focus.

Sorting Skill Levels at InsurTech Slopes — Applied to Startup Cohorts

David opened with a genuinely useful anecdote from InsurTech Slopes that ended up framing the whole conversation about evaluating founders. The first year, attendees self-reported their ski ability across three levels — beginner, intermediate, advanced — and the “advanced” group turned out to be wildly overconfident: roughly 12 people claimed advanced when only a handful of the actual attendees belonged there. Sorting them naturally happened on the mountain itself, as the group split organically based on who could actually keep pace on a chosen run — producing a real, observed skill hierarchy rather than a self-reported one.

The lesson David and the team applied going forward: four tiers, not three, including a new top category (“All Mountain,” for anyone comfortable on literally any terrain including woods and chutes) — because self-assessment reliably overstates actual capability, whether on a mountain or, by extension, in a startup pitch deck.

Why Insurance Benefits More From Traditional AI Than GenAI

This was the single sharpest and most useful technical distinction in the conversation, and it’s a genuinely underappreciated point in most industry conversations about AI adoption.

David’s framing, crediting analyst Adrian Jones for articulating the distinction clearly: generative AI is, at its core, the ultimate form of autocomplete — genuinely valuable for creative industries (content, media, marketing) where inventing plausible new material is the point. Insurance, by contrast, mostly needs to understand data it already has, not generate new content — which is exactly where traditional machine learning does the heavier lifting.

Concrete examples David gave:

Machine vision on property imagery — taking a satellite image and identifying hazards, or assessing the impact zone of a flood claim, is traditional AI interpreting existing information, not generative AI inventing anything
Document review in claims — using machine learning to assess whether a workers’ comp claim is likely to expand or become litigated is classification, not generation. GenAI’s role there is more modest: summarizing a case file to speed up human review, while the actual risk determination comes from the underlying classification model

The practical implication for buyers evaluating insurtech vendors: understand which category a given “AI” claim actually falls into before assuming it addresses the same problem GenAI headlines suggest.

Analytics vs. Statistics: The Credit Score Example

I pushed David to draw out the distinction between analytics and raw statistics, since the terms get used almost interchangeably in vendor pitches.

His framing: analytics is a derivative field of statistics, but branches into distinct categories — quantitative analytics (understanding data you already have, more precisely) and qualitative analytics (surveys, classification, understanding what customer information actually means). Most of the real underwriting value sits in quantitative analytics, at two levels:

Classification — a tool like relativity6, which helps identify a company’s correct risk classification, directly improves an underwriter’s ability to bucket a risk correctly and price it against the right comparison group
Deeper pattern discovery — auto insurance is David’s clearest example of where this matters. Credit score has long been used as a proxy for auto risk because of a real statistical correlation with claims performance — but credit score doesn’t actually cause accidents; it’s a rough proxy standing in for something more proximate, like driving experience or responsiveness in different road conditions. His illustrative hypothesis: Texas drivers who relocated from the Northeast may perform meaningfully better in icy conditions because of prior experience handling slippery roads — a genuinely testable, non-obvious pattern that only shows up if you have the analytical tooling to interrogate a large enough dataset for it

The distinction that matters: correlation-based proxies like credit score are a blunt substitute for the real causal factors; better analytics tooling lets carriers get closer to the actual proximate driver of risk, rather than relying on an imprecise stand-in.

What’s on the Conference Agenda

InsurTechNY’s fifth annual Spring Conference (March 20-21 in New York) is built around AI and advanced analytics across the full insurance value chain — distribution, underwriting, policy servicing, core systems, and claims — with dedicated tracks for both P&C and Life & Health.

David’s framing of why insurance is unusually well-positioned to benefit from better analytics compared to other industries: unlike airlines or manufacturers, which are locked into large fixed physical assets that take years to change, an insurance company (regulatory constraints aside) can update its underwriting algorithm and see the effect on its next batch of policies almost immediately — a much faster feedback loop for translating better data understanding into real business results.

A panel David is specifically anticipating: an “established insurtech CEO” panel featuring the CEOs of Hippo, Root, and Cowbell Cyber, discussing their respective paths to profitability — Hippo navigating natural catastrophe exposure, Cowbell managing persistent and evolving cyber risk, and Root working to make an auto book profitable in a category where even Progressive’s own economics are famously thin.

Tony’s read on that broader “insurtech 1.0” cohort — companies that went public or seriously considered it and have since had to prove out profitability — is that the common thread across the group has been genuine operational discipline: slowing growth deliberately, cutting cost, and in several cases (particularly in life insurance) returning to agent-based distribution after initially pursuing a direct-to-digital model. His read isn’t that direct digital failed outright, but that many companies found the more durable model is a hybrid — supplementing digital acquisition with agent distribution rather than replacing it entirely.

Key Takeaways

Insurance derives more practical value from traditional machine learning (classification, computer vision, document review) than from generative AI, which is better suited to creative industries generating new content
Analytics tooling matters most where it can replace blunt statistical proxies (like credit score for auto risk) with something closer to the actual causal driver of risk
Insurance carriers have a genuine structural advantage over asset-heavy industries: analytics improvements can translate into pricing and underwriting changes almost immediately, rather than waiting years for fixed-asset cycles to turn over
Self-reported skill or capability — on a ski mountain or in a pitch deck — reliably overstates reality; the only real test is direct observation under conditions that can’t be faked
The “insurtech 1.0” public cohort’s path to profitability has generally required real operational discipline: slower growth, tighter cost control, and in several life insurance cases, a deliberate return to hybrid agent-plus-digital distribution rather than pure direct-to-consumer
A five-year-old ecosystem organization can sustain roughly 20 events a year (conference, multiple accelerators, a fund, niche networking events) by finding genuinely differentiated formats rather than competing head-on with larger established conferences