Two Thousand People Trained
Every insurance executive reading this has sat through the meeting. A nice deck on the wall, a chart going up and to the right, and somebody announcing that two thousand people finished the AI training. Everybody claps. Jay Bell has one question for that room, and he asked it when we sat down live at ITC 2026 for the podcast. What can those two thousand people do today that they couldn't do before?
Jay is COO and Head of Business Transformation for Global Construction and Surety at Zurich North America, and his point is that the question isn't hard to answer if you think about it before the rollout. Set up the measurement. Calibrate it. Get a baseline. Then train people. Almost nobody does that step, so the chart ends up measuring activity and everybody in the room calls it impact. He put it plainly. "I've started to get more relentlessly focused on the difference between activity and impact."
Same Spot, Same Effort, New Tool
Jay also pointed at where the industry keeps aiming its efficiency work. Claims intake. Submission intake. The same two places carriers have poured money into for twenty years. Now leaders want to know where the return on all that AI spend went, and his answer is a fair question back. Why would anyone expect a different result from the same effort pointed at the same spot, when the only new ingredient is the tool?
His own approach started with no mandate and no money. Zurich stood up a volunteer group called DAILY, short for Driving AI Literacy, and it passed fourteen hundred members in about three months. One member of his team, Miles Wemken, had mostly used AI to polish emails. Then the team needed time studies done, and Miles put about sixty hours of his own time into building an agent in Copilot that's now used across Zurich, including in some international markets. That's a capability, and you won't find it anywhere on a training completion chart.
The Other Jay Said the Same Thing About Cyber
Two weeks before Jay Bell's episode, we published another conversation recorded live at ITC, this one with Jay Vinda, Global CISO and Cyber Risk Engineering Lead at Mosaic Insurance. Different Jay, different job, and almost the same argument.
Asked what keeps him up at night, Jay didn't reach for nation states or some exotic zero-day. His answer was the basics, for a reason most security buyers would rather not hear. Most of the controls that actually protect a company are process controls. They don't need a big technology budget, but they take a lot of effort to get to the point where they stop something, and the temptation is to skip the effort and buy a tool. His example was data classification. Companies keep trying to hand that problem to somebody else, it gets very expensive, and it still doesn't get solved. Put the time in yourself, he said, and you can solve it fairly quickly. He had a line for the expo floor too. "There's so many tools out there that say they're AI-powered cybersecurity or next-generation AI cybersecurity, and it's just a lot of fluff."
And he owns the outcome in a way I wish more executives would. The industry spends a lot of time teaching employees what a bad email looks like, and Jay worries that training has quietly become a liability shift. If you trained them, the click is their fault, and he doesn't buy that for a second. "If Mosaic gets taken down because someone clicked on a dodgy link, that's my fault." That matters more every month, because generative AI erased the old tells. The broken grammar and the strange formatting are gone, in English and in every other language. Often the first message has no link and no urgency at all. It looks like an email meant for somebody else, you reply to correct the sender, and a friendly conversation starts from there.
Why This Sounds Familiar
I've spent twenty-five years building software for insurance companies, and in most of the claims organizations I work with, a lot of the problems can be solved without AI at all. I said so on the show with Jay Bell. My favorite book on Lean is 2 Second Lean, which argues that shaving two seconds off every small task adds up to real results. Nobody issues a press release for two seconds, and you only get them by doing the work.
The bootstrapper in me recognizes this too. The whole argument of Be Your Own VC is that people who have had to break even every year without the safety net of outside money behave differently. They can't buy their way out of a problem, so they learn to work their way out of it, and they know what every dollar did. The two Jays are describing that same discipline inside big organizations. Zurich didn't buy literacy. Mosaic can't buy a classified data set. Somebody had to do it.
Cyber is also the chapter in Geek Out, my forthcoming book on seven years of the podcast, where I argue the insurtech thesis actually paid off. The cyber MGAs that came through the hard market were the ones that could prove their monitoring and underwriting discipline produced better loss outcomes, and the mature version of the product looks like a security service with an insurance policy wrapped around it. Jay Vinda's job is close to that idea with a title on it. He uses his security knowledge to sharpen how Mosaic's underwriters assess each risk, and he works with insureds on their own security strategy. He does for Mosaic's insureds what he does for Mosaic.
What I'd Ask Before the Next Rollout
If you're about to roll out an AI tool, or buy a security product with AI in the name, write down one capability first. Something specific that a person or a control will be able to do in ninety days that it can't do today. Measure where it stands before you start. Then measure it again on the same scale when the ninety days are up. If the number doesn't move, the tool didn't change anything, no matter what the adoption chart says.
Jay Bell asked the question better than I can. What can they do now that they couldn't before? Ask it before the purchase order goes out, and ask it again ninety days later with the baseline sitting on the table in front of you.