He Taught a Computer to Read Textbooks Before He Taught It to Read Insurance Policies

Ron Glozman came into insurance from an unlikely direction: teenage competitive judo, a serious neck injury, then years spent as a certified snowboard instructor teaching everyone from six-year-olds to 70-year-olds, breaking his back, both wrists, and a rib along the way. His actual entry point into the industry, though, was academic. As a computer science and business student in Canada, frustrated by the inefficiency of reading thousand-page textbooks for exams that tested maybe 1% of that content, Ron built an app in 2013 to “teach a computer how to read” — summarizing textbook content so he could study in four hours and still keep his scholarship. He put it online at a friend’s suggestion, and it went viral within two weeks — used in 33 countries and 44% of the world’s top Ivy League and equivalent schools, eventually named one of the 50 best student apps of all time.

Students, of course, have limited discretionary income, and 99 cents times a million users doesn’t stretch far in an expensive city like Toronto. In 2016, while presenting his natural language processing work at a machine learning conference, an audience member — a senior executive at one of the five largest insurance brokerages in the world — reached out, having never previously worked in or even thought about insurance himself. After roughly six meetings in eight weeks, the brokerage asked a direct question: could Ron teach a computer to read policies and binders to help reduce errors and omissions? At 19 or 20 years old, he said yes, received the largest check he’d seen to that point, dropped out of school, and Chisel AI was born.

About Ron Glozman

Ron Glozman is CEO and Founder of Chisel AI, an insurance-focused AI company that uses natural language processing to extract and structure data from unstructured insurance documents — policies, binders, and related paperwork — reducing errors and omissions and speeding up policy review. Chisel AI has won the Zurich Innovation World Championship, ACORD’s InsurTech Challenge, and appeared on DIA’s list of top innovative insurance solutions.

Nothing Is Broken — There Are Just 15 to 21 Touch Points

Asked what else is broken in insurance that Chisel AI could fix, Ron pushed back gently on the framing itself: he doesn’t think “broken” is the right word. Instead, he described identifying somewhere between 15 and 21 distinct touch points across the insurance value chain where AI could meaningfully help — with Chisel AI realistically planning to build out only seven to nine of them over five years, simply because building each one properly takes real time and money.

Two examples he gave in detail: quote comparison tools for commercial lines (already common in personal auto and home insurance, but largely absent for commercial), and coverage suggestion — an Amazon-style recommendation engine applied to insurance limits. He cited a statistic he finds striking: roughly 50% of people in the US are underinsured, because as an asset like a home appreciates in value, policies often just auto-renew at the original limit until it no longer covers a full rebuild. Flagging that gap benefits everyone in the chain — the policyholder gets adequate coverage, the broker earns slightly more commission on the higher premium, and the carrier avoids the messy scenario of a claim that can’t be fully paid out.

On claims prevention specifically, he described a conversation with a company that had condensed a 50-question underwriting questionnaire down to just three questions, after determining which three actually predicted restaurant fire risk — a change that improved both customer experience and data quality simultaneously.

AI 1.0 Is Data Extraction; AI 2.0 Is What Comes Next

Ron laid out a framework for where insurance AI actually stands today. Stage one — where he believes the industry is overwhelmingly focused right now, Chisel AI included — is extracting structured data from unstructured sources: emails, PDFs, Word and Excel files, data lakes, and legacy silos. He was blunt about the underlying reality of machine learning work generally: roughly 95% of it is data wrangling — cleaning, filtering, deduplication — not the more glamorous predictive modeling people associate with AI. Stage two, which he sees as still emerging, uses that now-structured data to generate real insight: predicting the best-fit consumer for a policy, or other forms of predictive analytics. Most companies, in his assessment, are either working on stage one directly or have partnered with a company like Chisel AI to solve it before attempting stage two.

He credited the open-source AI ecosystem — Hugging Face, spaCy, and foundational models like BERT from companies such as Google and Meta — with making this kind of specialized, “last mile” work economically viable for smaller companies for the first time. Before these tools existed cheaply and accessibly, meaningful AI work required the budget and compute of a major conglomerate; Chisel AI’s value, in his framing, is fine-tuning that foundational work into something insurance-specific — the equivalent of solving the “last mile” copper-wiring bottleneck in telecommunications.

He also named several other insurance AI companies he respects: Pinpoint Predictive, where he sits on the advisory board; Atidot, in the life insurance space; and Relativity6, a fellow Los Angeles-based classification-focused AI company.

Getting Carrier Data Is Its Own Discipline

Asked how Chisel AI solves the perennial problem of extracting real data from insurance companies, Ron called it close to a billion-dollar question. His practical playbook: have security and compliance answers ready before you need them (he described fielding roughly 400-question security assessments from prospective enterprise partners), and be as easy as possible for the customer to work with — rather than requesting narrowly specific data an overwhelmed client would need to manually assemble, Chisel AI increasingly accepts a broader data dump and does its own cataloging to identify what’s relevant, shifting the burden onto Chisel AI’s own team rather than the client. He also noted a cultural shift in how protective companies have become about data generally, even data that isn’t unique or proprietary — including a recurring, slightly absurd reluctance to share standardized ACORD forms, despite ACORD’s entire purpose being industry-wide standardization.

Live Partnerships: Exceedance and Zurich

Ron shared two active partnerships. The more recent is with Exceedance, a firm known for manual policy-checking services, now integrating Chisel AI on the back end to automate that work. The longer-running one is with Zurich: after Chisel AI won Zurich’s Innovation World Championship in 2019, the two companies ran a pilot that has since moved into full production, with Chisel AI’s policy-check solution used to deliver contract certainty across Zurich’s construction line of business throughout North America — with further line-of-business expansion in active discussion.

Advice: The Untethered Soul and The Surrender Experiment

For his closing recommendation, Ron pointed to two books by Michael A. Singer, The Untethered Soul and The Surrender Experiment — both centered on living without over-identifying with day-to-day emotion, neither fearing the future nor reliving the past. He found them especially valuable during the isolation of the pandemic, and noted that while Singer doesn’t write directly about running a business, the life philosophy underlying his own success comes through clearly.

Key Takeaways

• Ron’s path into insurance ran through a viral study app built to summarize textbooks with NLP — a cold outreach from a senior executive at a top-five global brokerage turned that same technology toward reading insurance policies and binders to reduce errors and omissions
• Rather than calling insurance “broken,” Ron frames the opportunity as 15 to 21 distinct touch points across the value chain, with Chisel AI deliberately building out only seven to nine of them over five years given the real cost of building each properly
• Roughly 50% of US policyholders are underinsured because coverage limits don’t keep pace with appreciating asset values — a gap Ron sees as a clear, underused opportunity for AI-driven coverage suggestion tools
• Ron’s framework for insurance AI’s maturity: stage one is unstructured-to-structured data extraction (where most of the industry, including Chisel AI, is focused today, and where roughly 95% of ML work is unglamorous data cleaning); stage two is generating real predictive insight from that now-structured data
• Open-source foundational AI models (Hugging Face, spaCy, BERT) have made specialized “last mile” AI work economically viable for smaller, vertical-focused companies — work that previously required conglomerate-scale budgets and compute
• Getting real data out of insurance carriers requires having security and compliance answers ready in advance, and increasingly means accepting a broad data dump and doing the cataloging work in-house rather than asking clients for narrowly pre-filtered data