When insurers talk about knowledge graphs, they mean fraud detection. That badly undersells what a graph is for. Insurance is one of the most connected businesses that exists — people to policies to claims to third parties to reinsurers and back to people — and relational databases are structurally bad at exactly that. A knowledge graph isn't a fraud tool; it's the substrate that lets AI reason over your business in the shape the business actually has.
In a relational model, a relationship is a JOIN — computed on demand, and expensive when you chase it several hops deep. "Find everyone connected to this claimant within 3 hops through shared bank accounts, addresses, or devices" is a self-join nightmare in SQL and a one-liner in a graph query language:
MATCH (c:Claimant {id: $id})-[*1..3]-(connected)
RETURN connected
The point isn't syntax sugar. It's that when relationships are first-class, a whole category of hard questions becomes tractable — and a surprising number of them turn out to be the same question.
Today most insurers answer these with spreadsheets and tribal knowledge because their data is trapped in row-oriented silos where the connections are implicit.
Here's the part that's easy to miss. An LLM agent reasoning over your business can only follow connections that are represented. Point an agent at flat tables and it sees rows; it cannot "walk" from a claimant to a shared repair shop to another claimant, because that path exists only as an expensive join nobody materialized. Give it a graph — or a retrieval layer over a graph — and the agent can traverse the same relationships a human investigator would.
The graph is what lets the AI think relationally instead of just retrieving flat facts. It's the difference between an agent that answers "what's on this policy?" and one that answers "what else should we be worried about, given everything connected to this policy?"
You don't rip out your systems. The usual path:
Full write-up, including which insurance problems collapse into graph problems:
Knowledge Graphs Beyond Fraud →
From IntelliBooks' series on the data foundation under insurance AI.
Anyone running a graph as a serving layer for LLM agents? Curious how you're handling entity resolution upstream.