AI made the build v. buy decision harder, not easier

For decades, the insurance software pendulum has swung between buy and build. It’s never been an easy decision, but at times it’s been clear there wasn’t a good solution available to buy (think about the late 1990s), or that an entire segment of the industry was buying packaged solutions (think CSC in the 1980s and 1990s, or Guidewire and Duck Creek for P&C carriers in the 2010s). But much of the time, sitting inside the carrier IT departments, the decision is quietly being litigated across the industry: to build or buy.

What’s different this year is the AI factor. Particularly in life and annuities, conversations that previously ended in deadlock or uncertainty are now bolstered by acolytes of agentic AI. Engineers and executives alike point to how fast an agent can scaffold a working system. AI has made it look a lot more enticing to build, promising cheaper and faster execution. And that’s shifted the thinking of the entire industry. The choices have often been “build, buy, or do nothing,” and the answer has most often been “do nothing” (AKA “kick the can down the road”).

Today’s frontier models and agentic coding platforms can build amazing software in a mind-bogglingly short time. In the early stages, progress feels dramatic and promising. Prototypes often look and feel almost production-ready. But a convincing first draft isn’t a production-ready system for a heavily regulated industry, and the gap between those things is what the AI-accelerated projections ignore. Lower cost estimates lean on the part of the job that got cheaper, writing code, while ignoring the part that never did: expertise about what the system needs to do.

There is no template for AI to emulate, no public model an LLM can learn from. Every carrier’s system encodes years of accumulated, complex judgment calls, like how underwriting, claims, billing and compliance actually interact, which regulations apply where, which edge cases matter and why, and even when human involvement is required. None of that has been documented for an LLM to pull from. 

The detour of trying to build a policy administration system using agentic coding has a real cost beyond the significant cost of AI tools. It’s time your engineers and their colleagues on the business side could be focused on building solutions that make sense to build, the ones you can’t find in the market or that don’t benefit from sharing R&D budget across multiple carriers. Projects that aren’t just undifferentiated heavy lifting but instead create value for the customer and distributor (such as customer and agent portals). 

AI is a valuable tool when pointed at the right problems. But it doesn’t magically solve the problem of needing expertise to build a production-grade policy admin system. 

So what works? The same thing that always worked: put resources on underwriting complex cases, marketing to an increasingly underinsured population, and pointing new tech on the edge, while leaning in to proven expertise for the core. That allocation, not another year of build-versus-buy debate, is what closes the coverage gap. Make sure your vendor is using AI to build better, faster systems with features you need so that you can take advantage of the AI revolution without having to build it yourself.

Visit www.bestow.com to learn more.