Seven Years of "We're Running a Pilot"
I've been hosting the InsurTech Geek Podcast for seven years. 177 episodes. And for most of that run — call it 2019 through 2024 — I could predict exactly what any carrier or TPA was going to say when I asked about AI.
"We're running a pilot."
It was the answer for five years. Pilot on underwriting. Pilot on claims. Pilot on fraud detection. Lab projects that lived inside innovation teams that reported to a CDO who reported to a CEO who was not yet convinced. The tech press spent that entire stretch writing breathless pieces about how insurance was "on the verge" of an AI revolution. And I kept interviewing smart people who were, very responsibly, running pilots.
2025 was different.
What Changed Wasn't the Technology
The models didn't suddenly get 10x better in January 2025. GPT-4 was already out. Claude was already out. The multimodal stuff was already there. What changed was that enough people inside enough carriers had run enough pilots that the business case stopped being theoretical.
In December, my co-host Rob Galbraith and I recorded our year-in-review for episode 170. Here's what I said:
"A year ago when we did our recap, we were still talking about lab projects. Lab, lab, labs, labs, labs. Pilots, pilots, pilots, pilots, pilots. Now it's full tilt implementation where we're doing med record summarization on millions of med records. We're doing claim level summaries. We're doing policyholder level summaries. We're doing far more accurate predictive modeling to estimate the total incurred prediction of a claim."
Millions. Not hundreds. Not a controlled cohort of 500 claims for a proof of concept. Millions of medical records, in production, every month.
That's the crossing-the-chasm moment. That's what it looks like when an industry moves.
Why Insurance Took This Long — and Why That's Actually Fine
The tech press got the timeline wrong because they applied a SaaS adoption curve to a regulated industry. Insurance isn't like that. You can't ship fast and break things when "things" means coverage decisions and claim payments and reserve calculations that regulators audit.
Every jurisdiction has rules. Every state has a department of insurance. The regulatory requirement that a licensed human makes the final decision on a claim — that didn't go away in 2025 and it won't go away in 2026. What changed is the interpretation of what "final decision" means. A licensed adjuster can review a claim summary generated by AI and render a decision. They don't have to personally read 800 pages of medical records to do it.
That distinction sounds small. It isn't. It's the entire ballgame.
At JBK, we watched this shift in real time. For years, the question from clients was "can we pilot this?" In 2025, the question became "how do we stop doing manual processing?" That's a completely different conversation. One is about curiosity. The other is about operations.
What Production Scale Actually Looks Like
Claim-level summaries. Policyholder-level summaries. Total incurred predictions that are materially more accurate than what adjusters were doing manually. These aren't demo features. They're running on actual books of business.
The companies doing this aren't startups. They're carriers and TPAs with millions of policies, legacy systems, and compliance teams. That's the tell. When the incumbents move something into production at enterprise scale, the technology has cleared a bar that VC-backed pilots never had to clear.
You had to survive the skeptical actuary. The legal review. The state filing. The IT security audit. The vendor due diligence process that takes nine months. If you made it through all of that and you're still running in prod, you're real.
What Comes Next
The honest answer is I don't know exactly, and anyone who tells you they do is guessing. What I do know is that the laggards are now three to five years behind the leaders, and in insurance, that gap tends to compound. The carriers using AI to summarize records faster are also closing claims faster, which means they're learning from outcomes faster, which means their models are getting better faster.
The pilots were never the point. They were the price of admission to a conversation that was always going to end in production.
We're just finally there.