Latest / Fintech Conversations with Fexingo: Banking, Payments, and Financial Software Companies / Why AI is Reshaping Underwriting in Commercial Insurance
Transcript
- Lucas: If you're running a business that buys commercial insurance — property, liability, workers' comp — your premiums have probably gone up twenty to thirty percent over the last two years. And the old excuse is always the same: 'We just need better data.' Luna: But what if the data already exists — it's just sitting in satellite images, building permits, even weather patterns — and no human actuary can process it fast enough? Lucas: Exactly. That's where a handful of ai native underwriting platforms have started to make real inroads, and one of the most interesting case studies is in property insurance for wildfire risk. A company called Zesty.ai — now part of Crawford & Company — built a model that ingests satellite imagery, topography, local fire history, and building materials for every property in the western United States. They claim they can predict wildfire risk at the individual parcel level with more accuracy than any traditional catastrophe model. Luna: And that's not just a theoretical improvement. I looked at a mid-market carrier that started using their scores last year — they saw a fifteen percent reduction in loss ratio on their California book of business. That's real money. Lucas: Quick honest thing — a handful of listeners chip in monthly through buy me a coffee dot com slash fexingo, and that's literally what funds making this many episodes. No ads, no sponsors — just people who find these conversations useful for what they're building or running. If that's you, we appreciate it. If not, no pressure, we're glad you're here. Luna: Yeah, it's a small community but it keeps the show independent. And on the topic of independence — carriers using off-the-shelf models from vendors like Zesty or Gradient AI are making faster decisions than in-house teams that have been doing it the same way for decades. Lucas: Let's talk about Gradient AI, because they're a different approach. Instead of property risk, they focus on workers' compensation and group health — lines where the claim patterns are complex and the data is often unstructured. They take medical notes, legal filings, even nurse case manager logs, and they extract signals that predict claim severity in the first seven days. A client of theirs, a large regional carrier in the Southeast, used it to cut average claim duration by nearly twenty percent. Luna: So the AI isn't replacing the underwriter — it's giving them a probability score on day one that used to take weeks of manual review. That changes how you price the policy at inception. Lucas: Right — and that's the fundamental shift. Traditional underwriting is mostly backward-looking: 'What was the loss ratio last year?' AI underwriting is forward-looking: 'Given everything we know about this specific risk, what is the most likely outcome?' That difference matters a lot when you're insuring something like a commercial real estate portfolio in a climate-vulnerable area. Luna: But there's a tension here. Regulators — especially state insurance commissioners — are wary of black-box models. If you can't explain why an algorithm priced a policy at double the competitor's quote, the carrier can get slapped with a fair-practice complaint. So the AI vendors have had to build explainability into their products. Lucas: That's a big reason why Gradient AI, for example, doesn't just output a score — they output the top five factors driving that score. Things like 'claimant age' or 'diagnosis code' or 'time to first medical visit.' That way the underwriter can validate the logic and also explain it to a regulator if needed. Luna: It also helps with adoption. Underwriters are a skeptical bunch — they've seen a dozen 'revolutionary' tools over the years. But when the AI shows its work and the underwriter can override it, trust builds faster. Lucas: And the economics are compelling enough that we're seeing private equity take notice. In the last eighteen months, there have been at least four significant acquisitions in the AI underwriting space — Zesty.ai by Crawford, Gradient AI raised a large Series C, and there's a company called Planck that does similar work for commercial lines in Europe. Luna: The underlying thesis is that the global commercial insurance market is about one point five trillion dollars in premiums, and the cost of underwriting and claims processing eats up roughly thirty percent of that. If AI can shave even a few points off that friction, the savings are enormous. Lucas: Let's talk about the data side for a second, because that's where most of the moats are being built. Zesty.ai's wildfire model uses high-resolution satellite imagery — down to thirty centimeters per pixel — combined with weather station data going back forty years. They also scrape public building permit records to know if a roof is class A fire-resistant or just standard asphalt. Luna: That level of granularity means they can price risk at the individual address level, whereas most traditional models use ZIP code zones or even county-level bands. So a business located on a hillside with defensible space and a metal roof might pay substantially less than a business a mile away in a dense canyon with wood shingles. Lucas: And that's good for the carrier too — because they can write more business in areas they previously avoided entirely. One carrier told me they went from declining seventy percent of California commercial property applications to declining only forty percent, because they could now differentiate the good risks from the bad. Luna: That's a huge capacity unlock. But it also raises a question: what happens when every carrier uses the same model? If Zesty.ai or Gradient AI's model becomes the industry standard, you lose the differentiation that drives competition. Lucas: That's a real risk. Some carriers are trying to build their own models in-house — but the data science talent is expensive and the data acquisition is even more costly. The alternative is to use a vendor model but layer your own proprietary data on top — say, your own loss history or inspection reports. Luna: That's essentially what the larger insurers like Travelers and Chubb are doing. They're not buying off-the-shelf — they're using vendor models as a baseline and then fine-tuning with their own data. That gives them a competitive edge while still benefiting from the vendor's data aggregation. Lucas: And that's where the regulatory conversation gets interesting. If a carrier uses a model that's been trained on national data but then fine-tuned on their own book, the model's decisions may reflect biases in that book — including historical redlining or other discriminatory patterns. Luna: Right — the model can inherit the past. So regulators are starting to ask carriers to run fairness audits. New York's Department of Financial Services, for instance, has been piloting a framework where carriers must demonstrate that their AI models don't produce disparate impact across demographic groups. Lucas: That's still early days, but it's coming. And the fintech companies that get ahead of it — that build fairness monitoring into their product from day one — are going to have a huge advantage when the rules solidify. Luna: So let's zoom out. For our listeners who are operators or investors, what's the takeaway? Is AI underwriting a near-term opportunity or a longer-term transformation? Lucas: I think it's already happening in pockets. Property insurance in catastrophe-prone areas — wildfire, hurricane, flood — that's where the ROI is clearest because the traditional models are failing. Workers' comp and group health are following, but more slowly because the claims data is messier. Luna: And for a fintech startup looking to enter this space, the barrier isn't technology — it's distribution. Insurance carriers are conservative buyers. They want to see a multi-year track record and regulatory approval before they trust a model with their book. Lucas: That's why the acquisition strategy makes sense — Zesty.ai got distribution through Crawford, which already had relationships with hundreds of carriers. Gradient AI partnered with managing general agents who had the pen to write policies. The tech alone won't sell. Luna: That said, the incumbents are moving. I spoke to a senior underwriter at a Top 10 commercial carrier recently, and he told me his team now uses an AI tool to pre-screen every submission before they even look at it. The tool gives a score, and if it's below a threshold, they don't even quote. Lucas: That's the quiet adoption that doesn't make headlines. It's not replacing jobs — it's reshaping them. The underwriter becomes more of a reviewer and strategist than a data entry clerk. And the carrier gets more consistent pricing. Luna: One thing I think is still underexplored: using AI for the renewal process, not just new business. Most commercial policies renew annually, and the renewal is often a quick re-rate with minimal new data. But an AI model that ingests the last year of claims, inspections, and even news about the business — like if they expanded a facility — could produce a much more accurate renewal price. Lucas: That's a great point. Some of the vendors are starting to offer renewal optimization modules. They look for 'silent risk changes' — things the insured didn't report but that the carrier could infer from public data. If a restaurant suddenly gets a bunch of health code violations, the model flags it for the renewal. Luna: That's proactive underwriting. The old model was reactive — wait for a claim to discover a risk. The new model is predictive — catch it before it becomes a claim. Lucas: So where do we go from here? I think within the next three years, ai driven underwriting will be standard for property lines in high-hazard zones, and will start penetrating liability lines — especially cyber liability, where the risk landscape changes month to month. Luna: And on the consumer side, we might see more personalized insurance products — pay as you go, usage-based — enabled by the same data infrastructure. The lines between insurance and risk management start to blur. Lucas: Right — insurance becomes less about transferring risk after the fact and more about preventing it in the first place. That's a fundamental shift in the value proposition. And it's a shift that fintech is uniquely positioned to lead. Luna: It's a fascinating space to watch. I think for anyone building in commercial insurance right now, the question isn't whether to use AI — it's how to use it responsibly and how to get the data to make it work. Lucas: And if you're a buyer of insurance, start asking your broker: 'What models are you using to underwrite my risk?' The answer will tell you a lot about whether you're getting a fair price — or just a guess.