Why AI Still Struggles to Deliver in Insurance

AI has reached the insurance industry — in presentations, roadmaps, and pilots. Use cases abound: underwriting copilots, claims assistants, document parsing, fraud scoring, predictive pricing.

And yet, real-world impact remains limited.

According to Roots.ai’s State of AI Adoption in Insurance (2025), 92% of insurers have launched GenAI initiatives, but only 22% have deployed them at scale with measurable results. These figures come from the U.S. — where the market is typically more advanced than Europe. In many EU countries, especially outside the top 3–5 groups, the adoption gap is likely even wider.

So why does AI still struggle to deliver on its promise?

The issue isn’t models — it’s the environment they’re built on

The challenges don’t come from the models themselves. Foundation models are improving rapidly. Tooling is mature. Open-source frameworks abound.

But AI performance depends on the environment around it — and most insurance systems aren’t ready for it.

Here are the 5 main barriers we see in the field:

1. Poor data quality

Data is fragmented, unstructured, unverifiable, or missing altogether. No model can perform reliably in that context — especially in regulated domains.

2. Legacy systems

Insurance IT often relies on siloed, batch-based infrastructure with little interoperability. Event-level traceability, real-time inference, and modular orchestration remain out of reach.

3. ROI misalignment

Too many projects are launched in isolation — disconnected from business goals like quote conversion, claims cycle time, or combined ratio improvement. Without impact, there’s no traction.

4. Regulatory constraints

AI that can’t explain its decisions won’t survive an audit. With Solvency II, the EU AI Act, and EIOPA guidance, explainability and control are no longer optional.

5. Cultural silos

Successful AI requires collaboration across underwriting, IT, compliance, legal, and operations. But projects often stall due to unclear ownership or resistance to adoption.

So what now?

The right question isn’t: “How do we experiment with AI?”
It’s: “How do we create an environment where AI can succeed?”

At Korint, we’ve built two complementary components to support that shift:

• Korint AI Enabler — our infrastructure layer that ensures insurers have structured, traceable, event-based data ready for AI use cases (copilots, scoring, automation).
• Korint AI Builder — our orchestration interface for configuring business-grade AI agents that are explainable, supervised, and connected to the insurance stack (documents, APIs, workflows).

Together, they make AI usable, explainable, and deployable in real-life insurance operations — not just in demos.

Conclusion

AI can bring meaningful transformation to insurance — but not without the right foundations.

It’s not a model problem. It’s a systems problem, a data problem, and an integration problem.

Once those are addressed, copilots don’t just work.
They deliver.

‍