Encora: Grounded by Design – Making AI Reliable for the London Market

The promise of large language models in insurance is compelling: faster policy analysis, smarter underwriting support, accelerated claims handling. Yet for many MGAs who have moved beyond the pilot stage, a familiar obstacle has emerged. The model drafts a summary of coverage terms, then gets them wrong. Confidently, fluently, wrong.

This is the hallucination problem, and in a distribution environment where precise coverage language carries legal weight, it is a risk to be engineered out, not tolerated.

Hallucinations fall into two categories. The first is factuality errors i.e. models inventing information that doesn’t exist. The second, more dangerous for insurance, is faithfulness errors: the model is given the right source document (like a policy wording, a binder, or a claims protocol) but contradicts it in the output. This happens because LLMs are pattern-completion engines, not retrieval engines. They infer answers rather than look them up, filling gaps with something plausible rather than something accurate.

For MGAs operating in the London Market, where policy wording precision and Lloyd’s compliance are non-negotiable, this carries real consequences. The Moffatt v. Air Canada ruling established the principle clearly: an AI agent’s output is a representation made by the company. “It’s only a draft” is not a legal defence.

Hallucination is an architecturally solvable problem. The leading approach combines Agentic RAG with a continuous evaluation layer. Unlike standard retrieval pipelines, Agentic RAG introduces a reasoning loop — plan, retrieve, critique, refine, generate — reducing hallucinations by up to 78% compared to standard RAG baselines. A live evaluation layer then scores every output for groundedness before it reaches the user, blocking responses that cannot be verified against source documents.

The result is a system that cites its sources, scores its own confidence, and routes uncertain outputs to human review rather than guessing — eliminating the hidden verification labour that currently undermines AI ROI across the market.

The era of accepting hallucinations as the cost of doing business is over. For MGAs looking to move from cautious pilots to trusted production deployments, the path is clear, and the expertise to navigate it is available. The era of accepting hallucinations as the cost of doing business is over. For MGAs looking to move from cautious pilots to trusted production deployments, the path is clear. If your organisation is navigating these questions, we are always happy to share what we have learned.

Find Out More: mgaa.co.uk/members/encora