Doug Marquis, Chief Technology Officer at Zywave, says AI is moving quickly from experimentation to production in insurance — but the real challenge is not simply connecting a large language model to a workflow.
In a highly regulated industry like insurance, AI systems need domain-specific intelligence, observability, governance, and the ability to explain how they reached an answer.
In this TechVoices interview, Marquis explains how Zywave Apex uses insurance-native AI agents, MCP servers, data enrichment, and generative AI to help brokers identify prospects, personalize outreach, prepare proposals, and grow their book of business. In short, AI supports the “super broker” who can handle—and profit from—a far larger client list.
Core Takeaways
AI for Regulated Industries: Zywave’s experience building AI for insurance shows that regulated sectors require much more than generic LLM capabilities. AI systems must understand state-by-state rules, industry nuances, customer context, and compliance risks before they can be trusted in production.
Observability Is Essential: Marquis emphasized that production AI needs visibility into how an agent moves from step A to step B to step C. Without evals, metrics, tool-use tracking, and explainability, companies cannot confidently answer customer questions or verify AI-generated recommendations.
Insurance-Native MCP Matters: Zywave Apex uses an insurance-native MCP server to connect AI models with policy information, loss data, company data, contact data, benchmarking data, and other industry-specific intelligence. Marquis argues this makes the AI far more useful than a generic LLM trained on broad internet data.
Brokers Become “Super Brokers”: Zywave’s goal is not to replace insurance brokers, but to automate the time-consuming research, prospecting, campaign creation, and proposal-building work that slows them down. Marquis says this could allow brokers to manage books of business that are 10X or even 20X larger than today.
Key Quotes
Production AI Requires Data, Visibility, and Domain Precision
“We didn’t get it right right out of the gate a couple years ago. It took three or four times longer than what I thought it was going to take. We tried to throw some traditional engineers at it, and there’s a large learning curve there. Eventually we got our footing, brought in skills and outside help, and trained people. But what we learned along the way is multifaceted.”
“Understanding the data and making sure that data is in place is a core pillar. But also the observability and the visibility into what’s happening. In a regulated space, we’ve moved to this non-deterministic capability, and that’s different than traditional software. In insurance, it’s not good to have things be non-deterministic or end up in a scenario where they’re not exactly right.”
Insurance AI Cannot Ignore State-by-State Regulation
“One of the things we learned in our first gen AI project was that the state regulations are different in every state. An LLM has a hard time understanding whether the user is in Montana or California, and therefore if it’s producing an AI-generated response that might be particular to Montana and not California, you have a problem there because it doesn’t know who the user is.”
“Within insurance, there are a lot of nuances to the regulations and what’s out there in the industry in all different facets. We’ve got to be really careful about how we create these agents so that we don’t produce something that’s not going to be accurate. That typically ends up in legal issues within the space.”
Why Observability Matters Across the Whole Company
“It really is a whole company transformation because that visibility is important for the rest of the business to do its job. If we don’t have visibility within our R&D department, it’s also really hard for our support department to answer questions if a customer calls in about, ‘Why did I get this answer? Why did I get that answer? Why did you suggest this?’”
“We looked at a number of third-party products. We built a lot of stuff. We spent a lot of time understanding that and building in things like evals, making sure that we could understand whether it was working properly or not, thinking about the metrics we need to track and whether it’s choosing the right tool at the right time.”
AI Agents Could Make Brokers 10X More Productive
“What the agents do within that process is everything that a salesperson would do. They look and say, how do I identify an ideal customer? Who can I best sell to? Where have I had success in the past? From there, understand what customers I can specifically go after within a certain industry. From there, who are the contacts at that particular company that I want to contact?”
“The insurance agent, the insurance broker is really in that consultative mode versus doing a lot of the legwork that leads up to the point in time where they want to be out there in face-to-face customer meetings and ultimately closing business. Our idea is, we’re not getting rid of the broker here. They’re incredibly important, but they can handle a book that is 10X the size as this today.”
