Should insurers train their own LLMs?

Many insurers evaluating underwriting AI solutions may start to consider their LLM or model strategy: should we train our own models from scratch, fine-tune, or take a different approach entirely?

At this stage in time, the best answer is self-healing AI. This approach focuses on building context and execution harnesses around the frontier foundation models, delivering better results at a fraction of the cost. Plus, it comes without the technical debt and expense that is impossible to avoid in model training.

Why do custom LLM strategies miss the mark?

Any effective AI solution for commercial insurance requires three distinct layers:

1. The model. The latest frontier models have extraordinary general reasoning and generative output capabilities. While they may lack insurance-specific domain expertise, they provide a solid foundational layer that you won't be able to cost effectively replicate or rebuild in-house. Insurers can't (and shouldn't) go toe-to-toe with OpenAI and Anthropic when it comes to the model layer right now.

2. The context layer. Your proprietary intelligence. This is your proprietary intelligence that the foundation layer operates on to deliver vertical-specific results leveraging your build up domain expertise. Nailing the context layer is the determining factor in which insurers must move quickest to generate the best results from AI adoption.

3. The agentic layer. This is where you orchestrate your AI tooling, validate outputs, capture feedback, and generate actual ROI from AI.

The value for insurers lives in getting layers two and three correct, not in the model itself. This is why training strategies are almost always destined to fail.

Fine tuning vs. self-healing AI for insurance

Should you fine-tune an existing model for insurance?

Fine-tuning modifies a pre-trained model's internal parameters to specialize it for insurance tasks. This sounds appealing but introduces multiple strategic risks:

Resource demands: Fine-tuning still requires ML engineering expertise, carefully curated training datasets, GPU compute budgets, and continuous retraining cycles.

Tech debt: Frontier AI labs release more capable foundation models every 3-6 months. A fine-tuned model that took six months and significant investment to build may be instantly eclipsed by the next generation of base models that handle the same tasks out of the box—without any fine-tuning.

Model lock-in: Fine-tuning ties you to a specific model version. When better models emerge, your investment doesn't transfer. You're building on a moving target in one of the fastest-evolving technology landscapes in history.

How self-healing AI works in practice for insurance underwriting

Self-healing AI operates differently. Instead of changing the model itself, self-healing systems add context and feedback harnesses around the model to catch and correct errors and let the model course correct in real time while its operating.

Consider this workflow. An AI agent drafts a policy endorsement clause for a complex commercial risk. Before returning the result, the agent layer validates its work against your underwriting guidelines, policy templates, and regulatory requirements. When it detects errors or inconsistencies, it automatically feeds that information back into its reasoning process and refines the output.

The foundation model hasn't been retrained. The model weights haven't changed. But the system has improved through richer context and tighter validation feedback loops.

Self-healing vs. fine-tuning

What are the benefits of self-healing AI in insurance?

1. Continuous improvement without continuous investment. When agents make errors, the system feeds corrections back as context. Your AI gets smarter through operational use, not through expensive retraining cycles that divert resources from core underwriting work.

2. Future-proof infrastructure. When the next generation of foundation models is released, your agents benefit immediately. The domain context, validation logic, and feedback loops you've built carry forward. No reinvestment required.

3. Domain experts in control. Underwriters and actuaries shape how the system improves—through rules, feedback, and validation logic—without needing ML expertise. The people who understand risk best are the ones refining the AI.

Self-healing gives insurers the benefits of fine-tuning, without the need to hire scarce talent, invest resources, and without the threat of tech debt. It adapts AI to your business, at the pace your underwriters move.

Evaluating AI vendors: questions to ask about LLM training strategy

When evaluating AI solutions for insurance underwriting or pricing, ask:

1. Does your approach require training or fine-tuning custom models? If yes, understand the ongoing cost, talent, and maintenance implications.

2. Can your system benefit from advances in foundation models without re-investment? Self-healing architectures improve automatically when new base models are released, unlike fine-tuning or similar approaches.

3. Who controls improvement cycles, ML engineers or domain experts? The best systems put underwriters and actuaries in control of refinement.

4. How quickly can the system adapt to new underwriting rules or market conditions? Model retraining takes weeks to months. Context and agent updates as quickly as your underwriters work.

The future of insurance AI isn't about who trains the biggest models. It's about who builds the smartest feedback loops on top of rapidly improving foundation models. They are the ones best set to win.

Book a demo → to explore how hx's self-healing agents deliver continuously improving underwriting intelligence for commercial insurance.