Our recent gathering in Chicago put insurance technology leaders and practitioners in front of the same question from a dozen angles: Where does agentic AI create value, and what stops it? The honest answer that kept surfacing is that the technology is the easy part. The hard parts are data, governance, and people.
Your models are ready. Your documents are not.
The most consistent warning of the day was that pilots succeed on clean, curated data and then stall in production. A 300-page commercial submission arrives in a dozen formats, the agents run, and the outputs look plausible but unreliable. Underwriters who have read those submissions for years spot the cracks and walk away. The models are good enough, but the document and data layer beneath them usually is not. Every failed pilot also spends political capital you do not get back and teaches your underwriters not to trust the next one. Before you orchestrate anything, make the unstructured layer trustworthy, complete, and accurate.
Distrust the demo.
Every vendor demo is polished and run on data the vendor chose, which tells you almost nothing. What separates a real product from a good story is how it behaves at the edges. What happens when it fails? What does it do with your incomplete data at production scale? A vendor who has shipped to real carriers can answer those questions and will tell you honestly what the product cannot do. Push to the point of failure, run a short MVP on your own data, and treat the vendor as a partner, not a one-time purchase.
Governance now means governing reasoning.
Agentic systems are nondeterministic, so the model governance rubrics you already have do not fit. Explainability and field-level provenance matter more than a simple audit log, and you cannot bolt them on once you are in production. Separate your AI reasoning governance from your statistical model governance, build the audit trail and source citations in from the start, and watch agent spend closely, because a runaway loop over a weekend can cost real money with no ripcord. Regulators are circling; a voluntary NAIC evaluation effort already signals the questions coming.
Start at the front door and buy to build.
The recurring first use case is submission intake, where a large share of in-appetite business never gets quoted because it reaches an underwriter too late. Start there. Keep humans in the loop and reimagine the workflow one discrete step at a time rather than replatforming everything at once. Buy the commodity layers, the foundation models, and your core systems. Own your differentiator: the workflow, orchestration, and logic that make your underwriting yours. AI-ready, composable APIs and emerging standards such as MCP are the connective tissue that make this work over legacy systems.
It is a people problem.
Agentic AI runs on knowledge, much of which is undocumented and walking out the door with retiring staff. Surface it now. Be transparent about the why, and frame AI as a way to grow without adding headcount you cannot hire, not as a quiet path to cuts. Let the agents reason and let humans decide.
Pick one narrow, high-friction workflow, fix the data beneath it, decide how you will explain and audit it, and run it as a partnership with a clear business case. That is how the art of the possible becomes production.
Keep the conversation going.
Join us next week on June 16, when we bring this same agentic AI lens to our Regional Life Insurance, Annuities, and Group Benefits Forum on Agentic AI in Des Moines. And for our P/C clients, save the date for our pre-ITC Insurance Technology Forum in Las Vegas on September 29. We hope to see you there!
