The Three Levels of an AI-Native Insurance Core

I dropped this article last month: Insurance AI is Full of Hammers Looking for Nails. And honestly, when it comes to what insurers actually want from AI, the gap between their needs is about as wide as it gets.

On one hand, they just want a chatty bot to explain HR policies, or an IDP to scrape text off dusty paper documents. On the other, they’re looking for AI Underwriting, AI Claims, or maybe even some all-powerful, do-it-all agentic platform that magically handles absolutely everything.

So, what happened? It’s almost comical watching a massive wave of self-proclaimed “AI-Native” vendors pop up overnight, each one claiming to do it all. OCR companies magically morphed into IDP wizards; RPA companies rebranded as AI Agentic Workflow gurus; Rule Engines became AI Orchestration; and SI companies? They’re now FDEs. It seems everyone and their mother is building an AI Middle Platform, and everyone is throwing the word "Ontology" around like confetti. So, what’s the ultimate AI makeover for Core System companies? We went from being Project companies to SaaS companies, and now? We’re the shiny new AI-Native Core. 

Alright, let’s cut the fluff and look at the actual results. Today, I want to break down my definition of AI-Native Core capabilities from a functional perspective. I’ve sliced it into three layers.

Most of today’s AI needs are stuck in Layer 1, a lucky few make it to Layer 2. But the real game-changer—the stuff that will completely rewrite an insurer's tech capabilities, cost structure, and long-term competitiveness—will live in Layer 3. "AI-Native" is a brilliant buzzword, so let’s try to find some common ground, shall we?

Level 1: Agents that run the business

This is the level everyone is already building. Agents that run insurance workflows on top of the core at run-time: intake a claim, assess it, adjudicate it, service a policy, answer a customer. They read and write business data through the core's tools, and they either assist a human or take the routine cases end to end.

It is the level a legacy system can pretend to support…for a while. Point an AI agent at a few APIs (often very expensive to expose), or worse, at a UI screen, and you get a POC. What you don't get is scale, because the moment you want agents to work across products and workflows, in real time and high concurrency, with a full audit trail, you are asking the legacy core underneath for things it cannot give you.

Level 2: Agents that change the business

The second level is where AI agents work at design-time, not run-time. They don't process a claim; they configure the products, calculations, rules and workflows that the claim runs against. Today this is the work of business analysts and IT. It is where speed-to-market actually comes from. An insurer that can iterate its products in days rather than quarters can quickly respond to a competitor, a regulator, or an interest rate move.

The distinction is this: Level 1 helps you run the business, Levels 2 and 3 (more on this below) help you change it. A lot of durable competitive advantage is in the second kind, as it drives agility. Instead of hard-coding requirements (as often the case in legacy systems) or clicking through configuration screens (the no-code world Peak3 originally built for), you simply upload product specifications in your own format or conversationally “talk to the core” and tell it what needs to be set up or changed.

This is only available if your products, calculations, rules, and workflows live as configuration the core can expose and an AI agent can safely read and rewrite. If that logic is trapped inside application screens, as it is in many legacy cores, there is nothing for a Level 2 agent to hold on to.

Level 3: Agents that extend the core system

The third level is the deepest. AI agents that change the core itself: writing new plug-in code, building a new module, patching a vulnerability. This is software development done by multi-agent teams that span requirements, design, build and test, with human engineers moving up into review and judgment.

It needs a microservices core with a real extension framework, a codified knowledge base, and a software-development platform disciplined enough to keep AI-written code traceable and safe.

Peak3’s vision of a self-evolving core system

Put the levels together at scale and you get our vision: a core that evolves itself. When a new requirement appears, say a government launches a new tax-advantaged savings product, the system first asks whether it can be met by configuration at Level 2, and only drops to Level 3 to write new code, then extensively test and validate it, when configuration isn't enough. First copilot, then human-in-the-loop, then human-on-the-loop, then genuine autonomy and self-evolution. We are now on this journey.

What "AI-native" actually means

What is important: The scope of work moves from humans to agents, not from the core to agents. Agentic AI doesn't make the core less important. It raises the bar for the core.

Being AI-native is not putting agents on top of your legacy core. Rather, each level has its structural demands. Level 1 needs clean, comprehensive APIs, MCP support, real-time performance and high concurrency. Level 2 needs products, calculations, rules and workflows held as configuration, exposed through an authoring interface an AI agent can manage safely. Level 3 needs true microservices, an extension framework and the knowledge base and tooling to develop against them. In fact, API-first microservices architecture powers each of the three levels.

A monolith built two decades ago can offer none of this cleanly. MCP and a CLI are only ever a surface expression of what lives underneath. You cannot expose a capability your substrate doesn't actually have.

This is why I am relaxed, and honestly rather excited, as the CEO of a modern core system company. The better the core, the more the AI agents on top of it are worth. We built Graphene as full microservices from day one. Every capability can be exposed by API, MCP or CLI, configured by an AI agent, or extended by one.

Agents are the visible tip of the iceberg. But the AI-native core, the knowledge base, the AI software development lifecycle (AI-DLC) platform, and so on, are the critical mass that enables the tip to float above the surface.

Alright, I know what some of you are thinking: "Why on earth do I need to rip out and replace my core system? Level 3 isn't exactly a must-have."

You could argue that slapping a PaaS layer on top of an AI Orchestration layer solves all your problems. Or maybe you just rip out all the business logic and treat your legacy system as nothing more than a dumb database. Oh, absolutely! If you aren't planning to play the long game in the AI era, "Hollowing Out the Core" might work for a while. PaaS + Middle Platforms might sound like a decent shortcut.

But in the age of AI? That’s probably not the ultimate winning move. I’ll be unpacking exactly why in my next few articles.