Across the London market, claims handlers are using AI tools in different ways. Some use them to summarise lengthy documents and draft correspondence, while others are beginning to incorporate them into coverage analysis, triage and reserving.
In many cases, adoption is moving faster than understanding. This gap is what this article series aims to address. The first article explores how these tools developed, explains key terminology and introduces some of the risks associated with their use
Where did it come from?
AI has been used in insurance for many years. Actuarial, pricing and claims teams have long used machine learning to identify patterns in historical data, flag anomalies and assess risk. Claims fraud detection, for example, has relied on systems trained to recognise patterns associated with potentially fraudulent claims. Many traditional systems were designed primarily for specific tasks using structured data, such as spreadsheets.
In 2017, researchers at Google published a paper introducing the Transformer architecture. Its attention-based approach made it possible for models to process relationships between words across a passage more effectively. This development helped lay the foundations for language models capable of working with unstructured text, including loss adjusters’ reports, claims submissions and manuscript wording clauses.
The next major milestone was wider public access. In November 2022, OpenAI launched ChatGPT, making conversational generative AI accessible through a straightforward interface. Its rapid adoption brought the potential applications and implications of AI into much wider discussion.
The language of AI: what you need to know
The adoption of AI has introduced a considerable amount of jargon. The explanations below cover terms you are likely to encounter in presentations, team discussions and boardrooms.
The foundations
Model: The underlying system trained to perform a particular task. When your firm licences an AI-powered product, deploys a claims tool or uses ChatGPT, it is interacting with one or more AI models. Different products may use models developed internally or supplied by third parties.
Large language model (LLM): A type of AI model trained on large amounts of data to understand and generate language. Many newer models can also work with images, audio and structured data. An LLM generates text by predicting the next token, which may be a word or part of a word, based on the context it has received. Models vary in their capabilities, design and limitations. Training and operating advanced AI models can also require substantial computing infrastructure, including specialist chips, electricity and cooling resources, creating risks and insurance considerations of their own.
How it learns
Training: The process through which a model develops its capabilities by learning patterns from data. As an analogy, imagine a junior claims handler studying a vast collection of claims files, policy wordings, coverage opinions, legal textbooks and loss adjusters’ reports to understand how insurance concepts and language relate to one another. Training is not the same as storing a searchable copy of every document, although models can sometimes reproduce information encountered during training. The model learns statistical patterns that help it generate responses to new instructions.
How you interact with it
Prompt: The instruction or question you give an AI tool. If the model were a claims colleague, the prompt would be your brief: what you need, why you need it, the format required and any constraints. Clear and specific prompts generally make it easier to obtain useful outputs.
Generative AI (GenAI): AI capable of creating content, including draft letters, summaries, code and analytical narratives. This differs from many traditional machine-learning applications, which were primarily designed to classify, predict or detect patterns rather than generate content.
Agentic AI: AI systems designed to plan and carry out a sequence of actions, potentially using other software and tools. In a claims setting, a system could be configured to open a new claim, review attachments, update a claims platform and prepare correspondence for review. Depending on its design, it may perform several steps without human approval at each stage. This raises important questions about governance, permissions and oversight, which the series will explore further.
The risks
Hallucination: An AI-generated statement that is false or unsupported but presented as though it were accurate. Examples include an invented legal precedent, an incorrect policy clause or a fabricated fact. These errors can have serious professional consequences, so AI-generated information must be checked against reliable sources, particularly where it informs claims decisions or external communications.
Context window: The amount of information a model can consider within a particular interaction. When documents or conversations exceed that limit, a tool may truncate, exclude or otherwise fail to use some material. This can result in incomplete analysis that is not immediately obvious. Context windows are expanding, but the issue remains important when handling lengthy claims files. Some systems divide documents into smaller sections, or chunks, although this can introduce challenges in maintaining context across sections.
Retrieval-augmented generation (RAG): An approach that allows an AI system to retrieve information from specific sources, such as an organisation’s claims files, policy wordings or handling guidelines, and use it when generating a response. Well-designed retrieval can make outputs more relevant and easier to verify, particularly where the system provides references to source documents. It can reduce some errors, but it does not eliminate the need for human review.
Key takeaways
- AI is not a single technology. Copilot, the AI component of a claims system and a public chatbot may use different models, data sources and controls. Their capabilities and limitations should be assessed separately.
- AI-generated answers can sound convincing even when they are wrong. Material facts, policy interpretations and other consequential outputs require appropriate verification and professional oversight.
- Understanding how to prompt effectively, manage source material and recognise limitations such as context windows will become increasingly important as agentic AI develops.
Up next: Artificial intelligence and the claims profession: asking the right questions, explores effective prompting in a claims context and why the quality of an instruction shapes the usefulness of the output.