AI doesn't change the case for building in-house

Can you build your own core system with AI? Probably. Should you? No.

In my last short article, I explained why agentic insurance operations will continue to rely on an insurance core system (albeit an AI-native one and not the legacy that is mostly still in place). 

But I have now heard from multiple insurance groups that they were considering or even experimenting with building a new core system from scratch with AI. As AI has transformed software development, can’t an insurer now build its own core and drop solution providers like Peak3? Have I spent the last eight years building something with no future?

Let’s agree on the premise. AI has dramatically accelerated software development. It has fundamentally changed how my teams work.

But building a core is not only about writing code. It’s a complex engineering and expertise problem: knowing what to build, why, and how. There is an enormous amount of ingrained domain knowledge in it. And it isn’t just the core application. It’s security. It’s the CI/CD, DevOps and testing layer underneath. And it’s security again. 

Now, insurers know insurance. Accordingly, the argument is: Insurers have the domain knowledge, AI can handle the engineering. But even with AI, the effort is still high. Then there is the entire CI/CD, DevOps, and cloud expertise, as well as the continued maintenance to keep it all evergreen. It is not as if a core can now be built for 1% of what it used to cost.

Some numbers for scale. We have cumulatively invested way over three million hours of R&D into our Graphene core. Over the last years, AI has made us faster: our incremental effort for new feature development has dropped by roughly half in most cases. 

Assume you, building from scratch with AI, could do it far more efficiently than we did. Assume an insurer could build the equivalent at 10% of the time we’ve put in (and I consider 10% to be a big stretch). That is still an extraordinary amount of effort, to build and then to maintain, year after year.

And for what? Where is the competitive advantage in building something you could buy in the market that has been validated and proven? Every hour spent rebuilding a core is an hour not spent on the things that actually set you apart from competitors.

That’s the real question worth asking. Not “can we build it?”. Increasingly, with AI, you can build a version of almost anything, with lots of expensive tokens and your most experienced team dedicated to it. 

And the team is a big challenge. It requires different talents than what is in place for most insurers. In-house insurance IT teams have historically focused on maintaining legacy solutions and not building on the latest and greatest tech stack with AI. Building a new core system requires a different DNA, culture, talent pool, business model than running an insurance company. Building software is not the core business of any insurance company.

So, when I'm asked "can we self-build our core with AI?" and "should we?", my answer is an uncertain "probably" to the first and a clear "no" to the second. You may call me biased as the CEO of a core system vendor, but I believe our interests and the insurers’ are genuinely aligned here.

That said, "don't build it yourself" is not the same as "accept the status quo." Today too many insurers are locked into heavy dependency on vendors and system integrators for every implementation, change and fix. The real opportunity isn't building a core from scratch. Instead buy the validated foundation, while you own the configuration, the changes and the pace of your own roadmap.

That is exactly what we are building towards. We are piloting and scaling the tooling to make it happen: an AI SDLC platform, tailored knowledge bases, agentic configuration solutions, and more. These are all designed to put insurers increasingly in control of their own destiny while keeping the strengths of a commercial, off-the-shelf core. 

More on that soon. What I'm already confident of: implementation costs and timelines are going to drop dramatically. 

There's also one more approach I still want to unpack: keeping the legacy system in place and wrapping an agentic middle platform around it. I'll get into why this falls short in a future post. 

-Bill Song, Peak3 Co-Founder and CEO