AI (artificial intelligence) has gone from science fiction to the top of board papers in a very short space of time. In the London market, the adoption and use of AI tools have become key strategic and investment considerations for many managing agents.
But where did AI come from, why is it suddenly everywhere and how does it actually work? This short article will:
- explore its history;
- demystify some of the jargon;
- and explain some of its limitations, as the first part of the LMA NexGen Claims Leadership Group series on AI & Claims.
Where did it come from?
AI and insurance are, in many ways, nothing new. Actuarial and pricing teams have used forms of AI, including “machine learning”, for many years. This is where a computer is given historical data and asked to identify statistical patterns. In fraud detection, for example, a system might identify characteristics associated with previous fraudulent claims and automatically flag new claims that share similar characteristics. These systems were relatively rigid and typically relied on data being organised and structured in a particular way.
In 2017, a major technical breakthrough called the “Transformer” changed what AI systems could do with language. Rather than processing text in a more limited way, Transformer technology enabled models to consider the relationships between words across sentences, paragraphs and documents. This helped pave the way for tools that could move beyond structured data and work with information contained in documents, such as loss adjuster reports and policy wordings.
The next major shift came with ChatGPT, released to the public by OpenAI in 2022. The key difference was the accessibility of the technology. ChatGPT allowed users to ask questions and give instructions in plain English, without needing to understand computer code or know how to write programming instructions.
This meant that, in theory, almost anyone could use AI to help with complex tasks, contributing to the rapid growth in the number of AI models, tools and applications we see today.
A guide to decoding AI jargon
Complex technical jargon is unfortunately synonymous with AI. Below, we break down some of the key terms you are likely to encounter and explain what they mean:
The model
Think of the model as the AI’s “brain”, or the underlying engine made up of highly complex mathematical algorithms and code. When people talk about “using AI”, they are really talking about interacting with a particular model.
Large Language Model (LLM)
This is a type of AI model trained on very large amounts of text. An LLM is essentially a vastly more powerful version of the predictive text on your phone. It uses maths to predict the most likely next word in a sequence based on the words that came before it. Models have names such as GPT 4 or Gemini Pro, and they differ depending on how they have been developed and trained. LLMs run in enormous data centres containing thousands of specialised chips, which require significant amounts of computing power, electricity and water to produce outputs.
Training
This is the process through which a model learns patterns from enormous amounts of data. Imagine giving a junior claims handler a million past claims files, policy wordings and textbooks, and asking them to read everything until they understand patterns in the English language and insurance principles, ready to answer questions. That process is “training”.
Prompt
The prompt is simply the instruction or question you type into the AI – e.g. if you are asking ChatGPT something, you are “prompting” it. If the model is an adjuster, the prompt is the email you send them telling them exactly what you need done (e.g., “Summarise this bordereau and highlight any claims over £50,000”). The big shift with ChatGPT was being able to prompt in simple English.
Generative AI (GenAI)
For a long time, AI could only analyse things (e.g., “Is this claim fraudulent? Yes/No”). Generative AI can actually create new things from scratch, like drafting a letter, writing computer code, or creating a summary. GenAI is what is most commonly thought of these days as “AI”, and it being able to create things is the major technology shift we are all experiencing.
Agentic AI
This is the difference between an AI that just gives you advice and an AI that actually does a job itself. A standard LLM waits for you to prompt it, gives you a draft email, and then you must copy, paste, and send it. An “Agent” can be told to monitor an inbox, open new claims, read the attachments, draft the response, update the claims system, and send the email entirely on its own. Agentic AI is touted as the “next step” in AI, where Agents will be able to perform tasks automatically.
Context Window
A “context window” is the short-term memory limit of an AI tool. Once you give it more information than its context window can hold, it will forget things and let things slip. For example, if it has a context window of 50 pages and you give it a 150-page report, it will not be able to assess it all at once when giving a response, so the answer will be incomplete. To get around this, some systems use “chunking”, which breaks a file down into smaller sections and processes them systematically.
Hallucinations
A hallucination is when an AI model produces something false, fabricated, or nonsensical, but presents it with confidence as if it was a fact. Sometimes the error can be harmless (for example, saying England won the 2022 World Cup), or sometimes it can be dangerous (inventing a fake legal precedent and saying it supports your coverage position).
RAG (Retrieval Augmented Generation)
RAG is what allows an AI to go beyond its own training data to understand internal databases, uploaded documents, and proprietary information, meaning a company does not have to train a model from scratch. It is like giving the AI an open-book exam, allowing it to search for what it needs in your database (Retrieval), and then assess it and answer based on what it has retrieved (Augmented Generation). Good quality RAG should drastically reduce hallucinations by anchoring answers in retrieved documents and information.
Limitations and bottlenecks for AI
Given what we know about what AI can do, you may wonder why it is not more widely adopted across the London market. There are still some significant hurdles to overcome, and addressing these is a major part of the decision making around AI adoption taking place across the market today.
Hallucinations and accuracy
AI models are designed to provide an answer but that does not necessarily mean it will be the right one. Accuracy remains a key challenge when adopting AI, particularly when tools are deployed at scale and small errors can quickly become significant. For a claims team, this could mean policy information being cited incorrectly or key facts being missed when assessing a claim. Improving accuracy remains an ongoing challenge for both AI companies and end users.
Data security and privacy
This is a major consideration for managing agents. Confidential information or personal data cannot simply be entered into a public system such as ChatGPT, so any AI tool adopted by a company needs to operate securely within its existing IT environment. This can be slow, complex and expensive to implement, particularly when multiple internal stakeholders are involved in the approval process.
Legacy systems
The London market also relies on a large number of legacy systems, many of which were created long before AI was envisaged. In most cases, AI tools need to be adapted to work with these existing systems rather than the other way around which can limit how effectively the technology can be used.
Cost
Every time an AI tool is prompted, there is a cost associated with the computing power required to process the request. Every piece of data processed and every output generated has a price attached to it. While this may amount to only fractions of a penny for an individual interaction, costs can add up quickly when an entire organisation is using a tool at scale.
AI is no longer a futuristic concept. It is something we will all increasingly need to understand as part of our roles. While the technology offers significant potential for claims professionals, there are still important hurdles to overcome. Throughout this article series, we will share thoughts and insights on how AI can be used effectively in claims. Next up, we will look at ARTICLE NAME.