How to build an AI underwriting model: a claims adjuster’s step-by-step

I’ve lost count of the times underwriting teams ask me, “Can we just plug in a model and get the same loss ratio we had last year?” The answer is no. But you can get close if you treat the model as a second pair of eyes—not a replacement for judgment. Over the past three years I’ve reviewed a dozen carrier pilots, sat in on underwriting committee debates, and rebuilt one model from scratch when the original team forgot to include catastrophe load. What follows is the process I now run every time. It’s not glamorous, it’s not “cutting-edge,” and it won’t make your CEO’s deck look flashy. But it works.

We deliberately prioritized the *data* over the *algorithm* from the very beginning. The design principle we anchored ourselves to was: *"Garbage In, Gospel Out."* We knew that even the most sophisticated, AI-fueled algorithm would crumble under poor or irrelevant data, so we spent months defining strict criteria for what constituted "good" data—cleanliness, relevance, timeliness, and structure. Filtering and validating inputs wasn’t glamorous work, but it was critical. We rejected the seductive allure of bleeding-edge models early on because we’d seen too many projects fail from trying to retrofit elegant solutions onto messy data. The constraint that shaped this was time: we couldn’t brute-force quality later, so we baked it into the foundation. Choosing simplicity here—robust data over flashy algorithms—set us up for reliable downstream performance.

While the article correctly emphasizes data quality, the exponential pace of AI transformation means most carriers are severely underestimating how quickly these processes will accelerate. We're witnessing a Cambrian explosion in insurance AI where capabilities double every 3-6 months, not years. Just as digital photography disrupted film in under 5 years (Kodak, once worth $31B, filed for bankruptcy in 2012 after ignoring digital trends), AI underwriting will follow Moore's Law compounding effects. The first generation of models described here will seem primitive within 18 months as foundation models trained on trillions of underwriting decisions become available. Carriers not building modular pipelines today will find themselves locked out of tomorrow's AI ecosystem where models can ingest entire underwriting manuals, adjusters' notes, and cat models simultaneously with human-level interpretation.

Consider the parallel with how GPS transformed navigation: Between 2008-2013, smartphone navigation adoption grew from 0% to over 70% in many markets, effectively obliterating standalone GPS device sales within just 5 years. Similarly, AI underwriting will move from pilot to core competency faster than expected because the fundamental economic driver - better risk selection = higher profits - creates irresistible competitive pressure. The 28x ROI mentioned in the Florida pilot case study is just the beginning. When foundation models achieve human+ level underwriting performance, the market will tip overnight. Carriers still debating whether to build their first model risk finding themselves in the same position as Blockbuster debating streaming services in 2005.

Start by forgetting the model. Start with the loss ratio you want to explain. Pick a single book—say, Florida homeowners with a combined ratio north of 110%. Pull every policy, every claim, every adjuster note, every reinspection photo, and every premium audit document since 2018. Export to CSV; don’t wait for a data lake.