The Insurance Safety Harness: five layers that make any LLM safe to deploy in insurance

Powerful language models are now good enough to talk to customers about insurance. The hard part is no longer capability, it is control. Here is how we make any LLM safe to deploy in a regulated market, by fitting it with a safety harness.

The model is the easy part

Foundation models have become remarkably capable. They can read a policy document, summarise cover and answer a customer in fluent, natural language. For a moment, it can feel like the whole problem is solved.

It is not. Insurance is a regulated financial service, and in a regulated setting one wrong sentence can cause real harm. Telling a customer their teenager is covered to drive when they are not, or that a home policy includes flood cover when it does not, is not a quirky model error. It is a mis-sold product, a rejected claim and a potential breach.

So the question that matters is not whether the model can answer. It is what happens on the small number of turns where it gets something wrong. A single model, however capable, is a single point of failure, and you cannot prompt your way to a guarantee.

Borrowing an idea from safety engineering: defence in depth

Industries that cannot tolerate failure, such as aviation, nuclear and medicine, never rely on any one component being perfect. They assume every component will eventually fail, and design independent layers so that no single failure ever reaches the person at the end. It is called defence in depth.

We take the same approach to deploying LLMs. Rather than betting everything on the model, we run it inside the Insurance Safety Harness: five independent layers of protection, each catching a different class of failure. Crucially, the harness is model-agnostic. It sits around whichever foundation model is best for the task, so we are never locked to one provider and can adopt better models as they arrive.

The five layers

1. Live guardrails

Real-time policy enforcement runs alongside every conversation. Off-topic prompts, unsafe language, regulatory red lines and out-of-scope advice are caught before they reach the customer, not after. Guardrails are configured per agent, per product and per market (motor in the UK plays by different rules to home in Ireland), and add only milliseconds to each turn.

2. Prompt-injection detection

Customers, and the browsers, emails and PDFs they bring with them, sometimes carry payloads designed to override the agent’s instructions. We detect these attempts, including indirect injection hidden inside attachments, and strip them before they ever reach the model. Every attempt is logged for security review.

3. Post-call evaluation

Every conversation is automatically scored against business and compliance criteria: accuracy, tone, regulatory adherence and customer outcome. No sampling and no QA backlog: this is full coverage across every channel, so you can drill from a score straight to the transcript and the exact moment that triggered a flag.

4. Testing and evaluation harness

Before a single change reaches production, it runs through a full suite of simulated conversations and regression tests. A new prompt, a new tool or a new product line is replayed against hundreds of edge-case and adversarial scenarios, including real situations that were flagged in the past. If behaviour drifts, it never ships, and rollback is one click.

5. Vulnerability and complaint flagging

Some moments need a person, not a bot. The harness continuously listens for signs of vulnerability, such as financial hardship, health issues, bereavement and distress, as well as complaints and dissatisfaction. The moment one appears, the conversation is handed to your team with the full transcript, the reason it flagged and a recommended next step.

Why it matters

Because the safety harness is independent of the model, it changes how confidently we can move. We can take advantage of the best general-purpose models available and still give insurers and regulators something a raw model can never offer on its own: every answer checked, every conversation scored, every difficult moment escalated, and a clear audit trail behind all of it.

This is the foundation beneath Insurance Companion, our conversational agent that helps customers understand cover, compare policies and make confident decisions, all within regulatory boundaries. If you are exploring how to bring safe, auditable GenAI into your customer journeys, we would love to show you the harness in action.

See how Open Agentic and the five layers of safety work ›