Bringing a Regional Insurer Into the AI Era: Start With the Gateway, Not the Chatbot

This is a thought experiment, not a client story. If I were brought in to take a regional insurance carrier into the AI era, the board conversation would start the way it always does: someone wants a customer-facing chatbot, because that is what AI looks like from the outside. I would spend my first week arguing against it, and my first 90 days building something less visible and far more valuable.

Your employees already deployed AI without you

Here is what is actually happening inside that carrier today. Underwriters are pasting policy language into personal ChatGPT accounts. Claims adjusters are summarizing demand letters with whatever free tool they found. Someone in compliance is drafting responses with an AI assistant nobody approved. The security team calls this shadow AI and wants to block it.

I would call it the most useful dataset in the company. Shadow AI is a demand signal. Every unauthorized prompt is an employee telling you, precisely and for free, which part of their job needs leverage. Blocking it destroys the signal and pushes usage further underground. The move is to capture it instead.

The gateway is the first 90 days

So the first real deliverable is a model gateway: one governed route to approved models, with role-based access, logging, and cost reporting per team. Underwriting gets models cleared for policy language. Claims gets document summarization with retention rules that survive a regulator’s visit. Everyone gets something better than the free tool they were hiding, which is the only governance strategy that actually works. Governance that says yes safely beats governance that says no and gets ignored.

Then you read the logs. Within a quarter you have a ranked list of real AI use cases, weighted by actual daily usage rather than by whoever spoke loudest in the steering committee. At an insurer, I would expect claims document intake, policy comparison, and fraud triage to surface near the top. Those become the candidates for proper product investment, with evals and failure design, because now you know the demand is real before you spend a dollar building.

The chatbot can wait

The customer chatbot might still get built, but a year from now, informed by thousands of logged internal interactions that reveal what customers actually ask through the people who serve them. Meanwhile the measurable wins land early: shadow AI displacement you can track, per-team cost visibility, and a workforce that stopped hiding its tools.

The unglamorous truth of enterprise AI is that the first system worth building is the one that watches how your people already work. Products built on that evidence tend to survive contact with reality. Products built from a board slide usually do not.