The future of insurance is contextual, conversational and customer-first – thanks to AI

The insurance industry has long been regarded as a highly regulated, slow-moving monolith. But look closely, and you’ll see that we’re entering a new phase of significant change and overdue optimization – one that puts the consumer experience front-and-center.

I’ve long anticipated this transition. My experience in recent years working on the core machine learning (ML) team at Google taught me that the true value of artificial intelligence is not in hardware efficiency but in the radical personalization it enables. I learned a valuable lesson earlier in my career: Even the most tech-savvy tools can fail if they don’t solve a human problem. It’s a belief that’s been reinforced after recently joining The Zebra team as chief AI officer, a position that empowers me to better employ AI and ML to maximize the synergy between product and technology and make for a more efficient and customer-centric insurance shopping model.

From keywords to context

For years, consumers have used basic SEO keywords to find insurance policies while carriers relied on simple ML to price these policies. But that paradigm is changing quickly. It won’t be long before we see automation processes occurring across several different areas, particularly on the customer-facing side, where we are attempting to collect as much context as we can.

Today, customers provide deep, intent-based context, with a shift toward hyper-personalized, highly detailed queries. The next logical step is for Large Language Models (LLMs) to parse this massive amount of customer intent data and map it against complex and unstructured policy data. This represents both a major challenge and a fantastic opportunity: successfully and efficiently organizing our internal insurance data so that it can be presented back to the consumer. The goal here is to advance to a stage where organizing policy data (the provider side) and collecting information (the customer side) blend to create a perfectly personalized middle ground.

I’ve observed this incredible shift firsthand. Customers used to find The Zebra after using a basic two-word Google keyword search like “car insurance.” But courtesy of LLMs, their queries today are much more specific and customized: “I need car insurance for my 2022 Honda Civic in Austin, Texas, and I park it outside.” Providing this rich data upfront and directly to our advisors changes the whole game, allowing us to bypass tedious intake and start nurturing genuine human relationships.

I love the automation side of AI – especially its ability to eliminate manual data entry. Still, handling sensitive customer information means that AI tooling can quickly throw gasoline on a growing fire. That’s the reason I’m committed to creating a controlled environment. At The Zebra, we’ve spent a decade researching how humans sell policies, and I want to ensure our models don’t “reward hack.” A good illustration of this is an unsupervised model believing that hanging up on every customer ensures it will never technically lose a deal. At our company, it’s my job to build the guardrails that prevent these bizarre optimizations.

Exciting developments underfoot

We’ve entered a sea change moment, signaling a shift toward agentic coding and dynamic UI and away from the hard-coded, static websites that earlier dominated the insurance space. At present, most insurance online portals are linear, with the sequence of questions typically predetermined by a developer. In the coming phase, however, the LLMs will define the questions and the order in which they are asked, and UI components will be chosen or generated ad hoc, depending on the consumer’s particular context. That’s real progress.

Even more than the interface, I’m energized about forthcoming improvements in the agent-to-agent ecosystem. I envision a future where, for example, an online car-buying service like CarEdge or AutoCompanion learns key info about a given user, including their vehicle preferences, budget and safety priorities; then, at the moment the user clicks “buy,” that entire world of context is communicated directly from that website and its human or AI agent to an insurance agent. This layer of interaction eliminates the need for the customer to repeat redundant details while also ensuring that they are instantly offered the best personalized insurance policy.

Peering forward, I’m especially excited about this “agent-to-agent” future. Years ago, I built recommendations, search engines and pricings for a used car e-commerce platform, employing basic filters like “Toyota under $30,000.” But nowadays, people hunt online with so much more context and use conversational AI for nuanced intent, with focused queries like “I need a car for a family of three that fits a bulky stroller.” Extracting that deep context from a vehicle-purchasing AI agent directly into our insurance workflow is a fantastic opportunity to eliminate friction.

Another trend that has me stoked? AI’s increasing ability to help regional insurers structure their data. Smaller insurers, which know their local markets more than national players, have traditionally struggled to aggregate their numbers for digital marketplaces. But by better leveraging AI agents, we can now provide info to customers that was previously difficult to obtain, introducing them to insurance companies that are a better fit.

What’s behind my continued fascination with AI agents? Maybe it has something to do with our corporate culture at The Zebra, which emphasizes creativity, curiosity and fun – as exemplified by our obsession with Legos. Walk into our offices or attend a virtual meeting and you’ll see everyone clicking bricks together. We even commemorate exceptional company or career milestones by gifting Lego sets. These all-ages toys serve as a physical manifestation of a “Tinker mindset.” Legos have taught me that complex architectures are simply collections of tiny but well-defined parts you can remix inventively. When we launched Zebra Labs, it dawned on me that “vibe coding” with AI is akin to playing with digital Legos: Snapping together prompts, models and APIs to construct a security triage agent calls for the same kind of foundational logic needed to assemble a Lego Technic Porsche 911. To me, this mindset makes AI development feel like an organic extension of how we already work.

Navigating challenges ahead

It’s natural to be optimistic about what’s just beyond the horizon. Still, the industry has to be careful about “AI slop” – using AI to build things too quickly without truly knowing where and how it fits. This is an especially slippery slope when it comes to licensing. Case in point: If you ask an off-the-shelf LLM which policy to purchase for a Porsche (the real, non-Lego kind), it could suggest one that violates licensing laws.

You also have to think ahead about hidden hazards when operating in a regulated financial space. In my previous job at Google, I worked on improving the efficiency of Waymo models; at The Zebra’s Austin headquarters, I constantly see these autonomous vehicles, which have taught me quite a bit about edge cases. A Waymo can spot a city stop sign completely obscured by overgrown leaves thanks to its LiDAR and internal maps, safely stopping the car and logging the danger. At our company, strict compliance rules represent those hidden stop signs. Imagine a generic LLM confidently advising a user to drop comprehensive coverage to save a few dollars; doing so means it is acting as an unlicensed advisor – a massive legal liability. The lesson here is that you can’t just give AI the keys, ask it to be professional and simply hope for the best.

To prevent this, I’m creating a digital sensor suite using strict evaluation platforms. This way, before an AI agent can interact with the customer, it has to survive a simulated gauntlet graded on three metrics: factuality (is the information correct?), compliance (did it avoid making a regulated recommendation?) and accuracy (did it hallucinate any features?).

I was recently in Nashville for the Insurance Innovators conference. The audience wanted to know where I’m placing my business bets for The Zebra. My answer was simple: “personalization agents.” I’m investing heavily in resources that evaluate colossal volumes of policy data that can supercharge our human advisors. By extracting the specific data points they require exactly when they need them, I can help eliminate administrative friction and ensure they receive the perfect coverage for their unique risk profiles.

But it’s important to prioritize jobs as we make this transition to more grounded LLMs. This sector isn’t ripe for disruption simply because you can decrease and automate costs by subtracting people from the equation. This next phase I foresee should produce a more synergistic partnership between AI and human beings – especially licensed advisors who make sure we apply hyper-personalization and avoid the kinds of mistakes AI is known for. Yet we also want to unlock an experience so precise that relying solely on human decision-making will eventually feel like an unnecessary risk.

This customer-focused and increasingly bespoke insurance experience I’m dreaming of requires moving beyond simple task automation to a point where data is efficiently structured, and both human and AI agents communicate effectively across platforms. We’re almost there, and everyone – from the policyholder to the underwriter – stands to benefit.

This article is published as part of the Foundry Expert Contributor Network.
Want to join?