Data is the moat. The model isn't.

Once a week, someone asks me which LLMs I recommend or use at Peak3. It's a fair question. It's also usually the wrong one. 

In my opinion, foundation models are commoditising. The gaps between the best proprietary models and the best open-weight/source models are narrowing with every release. Token prices are collapsing. The capability we will have access to in six to twelve months will almost certainly exceed what we have today, at a fraction of the cost. This is a tailwind, not a differentiator. It's a tailwind that reaches everyone at once. 

Where you do win is on the data you feed those models. And how you embed the model into your organization through change management. 

Here is a simple test: Take any AI application your team is building, conversational FNOL agent, underwriting co-pilot, claims fraud, waste and abuse (FWA) assessor — and ask yourself: if you gave this exact same agent to your closest competitor, could they easily replicate the output? If the answer is yes, you don't have a moat. If the answer is no, it's because your data (curated, structured, connected, governed) is doing the work. And because you have embedded it into your organisation, driving change, adoption and continuous enhancements. 

Change management is a tricky subject and organisation-dependent. But data foundations are mostly universal. I see three data capabilities that separate insurers who will compound AI advantage from those who won't. 

• Unified customer view. An FWA AI assessor working against a fragmented, siloed claims view operates at a fraction of its potential. The same AI agent, on a consolidated customer record that ties policies, prior claims, third-party data and interaction history together, will deliver its full potential. Unglamorous data-fabric investments provide the foundation for every agent that runs on top. 

• Proprietary knowledge bases for RAG. Your underwriting and claims manuals, your underwriting and claims adjudication precedents, your internal playbooks and SOPs, all of them vectorised, retrievable, kept current. This is critical institutional memory rendered machine-usable. Competitors can't borrow it. 

• Rigorous data lineage and governance. Every AI-generated decision carries metadata: model, prompt, retrieval source, timestamp. Regulators and auditors may eventually ask. Laws in certain jurisdictions already require it, most others will likely follow. Retrofitting is painful and expensive. 

Most insurers' AI strategies are over-indexed on building shiny new things and talking about the latest advancements of LLMs. They are under-indexed on the underlying data fabric (which a modern core platform can provide). I understand why, but don't consider it a sustainable approach. 

My advice: Build with an LLM-agnostic approach so that you can simply choose the model that works best (I wrote about it a few days ago). And importantly, invest in the underlying data and core layer to ensure that your AI agents can actually become a competitive asset.

-Bill Song, Peak3 Co-Founder and CEO