We Just Had Our Best Year in Two Decades. That’s Exactly Why I’m Worried.

I’ve been in this industry long enough to know that a great year can also be a warning sign more than a celebration.

2025 gave P&C carriers their best combined ratio in almost two decades — down in the low-to-mid 90s, a level most of us haven’t seen in nearly twenty years. And for the first half of 2026, underwriting income nearly tripled from the same period in 2025, improving the combined ratio by four percentage points to 92.5. However, this improvement is deceiving. As AM Best notes, catastrophe losses accounted for 6.2 points of the six-month 2026 combined ratio, down from an estimated 10.8 points in the first half of 2025, when the industry was impacted heavily by the California wildfires. Also driving the net underwriting gain was a 3% increase in net earned premiums and a 5% decline in losses incurred and loss-adjustment expenses.

The improvements do not address foundational underwriting capabilities and underlying challenges. With AM Best, Fitch, and S&P all projecting further rate softening, the combined ratio could drift back up toward 97-99% or higher if the second half brings large catastrophes.

What’s left is the actual test on who is addressing foundational underwriting capabilities: which carriers are pricing risk and paying claims with real precision, and which ones are riding a good market and price increases.

The edge most carriers already own and aren’t using

Every carrier running a modern policy administration system is sitting on years, sometimes decades, of real transaction history — quotes, policies, claims, billing, renewals. That’s not generic “industry data” available for anyone to buy. It’s the specific, provable record of who you underwrote, what you priced, what you paid out, and who stayed or left, and why. No competitor can replicate it, no matter how similar their book looks from the outside.

The reason that edge so often sits unused isn’t a lack of ambition. It’s that turning data into a working model, the traditional way, is too slow. They don’t have a fast and efficient way to turn that data into a decision. Why?

They have built machine learning models in the traditional way.

Why the traditional way is too slow for what’s coming

Strip away the terminology and a typical machine learning project runs through the same sequence, by hand, every time: understand the business problem, track down and collate the right data, clean and prepare it, explore it, implement applicable modeling techniques, pick the best performer, deploy and monitor it — and then keep retraining it, indefinitely, because its accuracy starts decaying the moment the world it was trained on shifts.

None of that is unique to insurance. But insurance has a specific version of the problem that I think doesn’t get talked about enough: our risk conditions move faster than our model-build cycles do.

If it takes six to nine months to get a fraud model or a pricing model from idea to production — which is a realistic timeline once governance and sign-off get added at the end as they usually are — the fraud patterns and loss trends underneath that model have often shifted twice before it ever goes live. You end up deploying a model that’s already slightly out of date, in a market that’s about to get more competitive, not less.

That is a decisioning gap — between accessing the data they already own and the speed at which they can act on it — and is the real opportunity in front of this industry right now.

Not another AI headline. A faster, more reliable and governed path from what you already know to what you do next.

Where we need to go

We’ve spent this year working on exactly that problem — building a faster, more governed way to turn a carrier’s own data into models that are trained specifically on that carrier’s own book, not an industry-average dataset sold to everyone.

We need to change the whole conversation: the regulatory ground under AI in insurance shifted meaningfully in 2026, and it makes “move fast, add governance later” a much riskier bet than it used to be.

We need to bring machine learning models, generative AI, agentic AI, and analytics together into one connected intelligence layer for the business, including underwriters, instead of separate initiatives competing for attention.

The regulatory reality most AI pitches skip over

As of mid-2026, at least 25 states plus the District of Columbia have adopted the NAIC’s Model Bulletin on the use of artificial intelligence systems by insurers. That’s more than half the country operating under a formal expectation that carriers document, govern, and monitor the AI they use across the business from underwriting to claims and pricing decisions.

What makes this different from the last few years of AI hype around adoption is that the NAIC now has an AI Systems Evaluation Tool in an active multistate pilot this year, designed to give examiners a standardized way to review an insurer’s AI governance during a market conduct exam. And the detail that matters most: regulators have said directly that they’ll look at vendor relationships and capabilities when they examine a carrier’s AI use.

Outsourcing the model doesn’t outsource the insurer’s accountability.

This reframes the whole conversation. Governance isn’t a compliance checkbox you sign off on once a model works. It has to be part of how the model gets built in the first place.

What Majesco built

Every model we build moves through the same governed lifecycle for every use case: a named business owner who signs off on the business impact, a data science owner responsible for the model itself, and a documented approval gate before anything reaches production. Audit trails run the whole way through, not bolted on at the end.

Three key concepts on how we build it matter the most:

The model is trained on your data, not an industry average. Every model is built on the carrier’s own book — their claims, their losses, their customers — and gets refined as that book evolves. A generic model trained on pooled, anonymized, industry-wide data will always be a step removed from your actual risk. This model approach is not.

Your data doesn’t go anywhere. The model gets built where the data already lives, inside the systems you already run. No new export pipeline carrying your claims history out to a third party, no new vendor holding a copy of your book, no new attack surface for your security team to defend. If you’re a CISO reading this, this detail should matter more to you than any accuracy benchmark.

It’s fast enough to actually learn. A standardized, governed pipeline gets models to production in weeks rather than months, without skipping validation or sign-off to get there. That matters less for the speed and more for what it enables: deploy, watch what happens, learn from it, and reshape the model — a continuous loop, not a one-time project. A model that took six months to build and can’t be touched again for a year isn’t really learning your business. It’s a snapshot in time.

This is not a vision. It is real and credible.

It is not our first production AI system. Majesco has been actively building AI capabilities from the original launch of ChatGPT over 3+ years ago. Majesco Copilot and Agentic AI agents are live, already running, already handling real work for real customers. What is described here is the next chapter of that same trajectory, not a first attempt. 

The bigger bet — A connected experience

We don’t think ML models, generative AI, agentic AI, and analytics should exist as separate initiatives fighting for the same business user, like an underwriter’s attention. We believe these AI capabilities must come together into one connected experience, built into the workflow itself rather than living in a dashboard for someone to check periodically.

The way we are applying AI in the workflow is simple and should stay simple: first give the underwriter clear insight into what’s happening, then a specific recommendation, then automate the parts that are safe enough to automate under real guardrails, and route anything uncertain to a person rather than guessing.

Insight before automation, always — not because automation isn’t valuable, but because trust has to be earned in that order, especially under the regulatory environment we operate in.

That’s the connective tissue we are creating at Majesco in our Intelligent Core solutions. The ML Models Factory is the engine. Where it shows up — inside the underwriter’s work, not next to it — is the rest of the story. You’ll see the first real look at that this fall in our Fall Product Release and at our Frontier ’26 Conference in New Orleans.

If you’re a customer, you and your data are on this journey today.

If you’re evaluating what “AI-powered” actually means in an insurance core platform, ask any vendor the same question: Is the intelligence they’re offering built from your book of business, governed the way regulators now expect, and fast enough to keep learning — or is it a generic model with a good demo?

We would love to have a conversation directly. Reach out if you want to talk before Frontier ’26.Sources: AM Best, “First Look: Six-Month 2026 US Property/Casualty Financial Results” (Sept 2026); Fitch Ratings 2026 U.S. P&C Insurance Outlook; S&P Global Market Intelligence; NAIC Model Bulletin state adoption tracking (Quarles Law Firm; aipmo.co Q2 2026 s