InsTech’s Henry Gale speaks to Bas De Goei, Insurance Industry Leader at Instabase, about the advent of large language models (LLMs) and how insurers can automate commercial submissions and claims processes.
Henry: Could you introduce Instabase?
Bas: Instabase is a software platform that enables organizations to automate manual processes by leveraging the latest innovations in AI. Instabase helps enterprises, such as large insurers and brokers, tackle a variety of unstructured documents, such as policy documents, claim documents and commercial submissions. Instabase’s technology processes those documents and extracts relevant data to help companies run analytics or manage workflows.
Henry: What is your role and what does it involve?
Bas: I am the Insurance Industry Leader at Instabase. My role is to replicate in the insurance industry the success Instabase has had in the banking sector. The company is now fully focused on both banking and insurance. I joined Instabase nearly two years ago after working at AXA for ten years, most recently leading technology, partnerships and investments at AXA Next.
Henry: Instabase uses deep learning technology, could you explain what that is?
Bas: Historically the first approaches to extracting data from documents used rules-based methods, which could compare documents to a template and extract data based on the template. If documents did not fit the specific template, these approaches would break down.
Machine learning approaches, developed some years ago, involved training models on different examples of a document to recognise variability in the template. But these require many thousands of sample documents to produce a high-quality model.
Deep learning models, such as those developed by Instabase and many market players, have already been pre-trained on millions of documents of all kinds. To use them for a specific type of document, such as an insurance submission in one line of business, you only need dozens of sample documents (rather than the thousands that would be required for traditional machine learning).
Henry: Why is deep learning technology relevant to insurers?
Bas: We believe underwriters currently spend up to 40% of their time on administrative tasks. Many of these tasks, such as reading submission emails, rejecting submissions that are out of appetite, routing submissions to the correct underwriting team and inputting information from emails and broker presentations onto an underwriting workbench, can be intelligently automated using deep learning models.
For commercial claims, insurers also receive highly variable documents such as medical records and litigation letters, that are currently being classified and the data inputted into claims management systems manually. For simple claims, significant parts of the claims process can be automated, and for complex claims, the adjuster can receive all the necessary information more quickly with intelligent automation.
Henry: How do you define intelligent automation?
Bas: Instabase’s approach to intelligent automation involves a combination of several complex deep learning models to enable machines to truly understand unstructured documents.
This starts with optical character recognition models that convert printed or handwritten text into a machine-readable document. Next, classification models understand what type of document it is, for example whether a document is a submission, a claims form or a loss run.
Once the document is classified, another deep learning model extracts data from the document. For example, each submission might have the company address in a different place. Deep learning models can read a document almost like a human, understanding what an address looks like and finding it in the document.
Next, data needs to be validated to check it is correct. For example, if a form has information about a customer, a machine can check whether the customer is in the company’s customer database.
Finally, the data needs to be refined into a clean format. There are different conventions for writing dates internationally and different units are used for measurements. These are hard for machines to understand and require another deep learning model. Once the data is validated and cleaned, it can be automatically linked via API with a relevant system. Instabase integrates with policy administration systems such as Guidewire and underwriting workbenches such as Unqork.
Henry: What trends do you see in the adoption of intelligent automation in the insurance industry?
Bas: The interest in these technologies in insurance is very high. Instabase is in discussion with most of the largest brokers and insurers in the industry. Instabase is working with AXA, which is implementing intelligent automation for commercial submissions and claims, as well as other companies which we cannot disclose.
I love the phrase that artificial intelligence (AI) will not replace you, but someone who can use AI effectively will. Insurance companies can adopt AI technologies pragmatically in their work streams today and the most forward-thinking companies are doing this already.
Henry: What is your advice to insurance companies who want to be more forward-thinking in their adoption of AI?
Bas: Companies should start with obtaining a good understanding of their current processes. In my experience with insurance companies, existing processes produce good outcomes but the processes themselves are not very efficient. Analysing the steps in a process that leads you to a high-quality outcome allows companies to identify the parts which can be automated.
For example, suppose an insurer receives submission documents to a shared email inbox. In that case someone reads the email, understands that it is a submission and forwards it to an underwriting assistant. The underwriting assistant sorts through the documents and populates the relevant information into an underwriting workbench, but then a more senior underwriter realises there is some missing information, which they need to request from the broker.
Understanding how that process works allows you to identify the steps where people are looking at documents, classifying them and extracting information. It can also highlight areas where information needs to be requested but the missing information is not noticed until a late stage in the process. In all these areas, intelligent automation can make the process more efficient.
Once the processes are mapped out, the next steps are to identify the business lines with the most to gain from automation quickly and reach out to technology companies offering solutions.
Henry: Large language models (LLMs) are an area of AI that has been popularised recently by chatbots such as ChatGPT. Can you explain what LLMs are?
Bas: LLMs are one of the latest innovations in deep learning. They are ‘large’ language models because they are deep learning models trained upon a significantly larger number of documents or texts than previous models. Because they are trained on more documents, they can understand the relationships within and between texts better than other models.
As a result, where a deep learning model requires dozens of samples to train it to understand an insurance submission, an LLM requires zero samples. Without any training, you can input a submission and it should immediately understand the risks to be covered, where the business is registered, the requested limits and other relevant information.
LLMs can also be used for ‘generative AI’. ChatGPT and other AI chatbots use LLMs to generate responses to questions. Similarly, LLMs can be used to generate business documents. For example, a large language model could extract information out of a broker submission email and its attached documents, summarise the email’s content and populate a policy document based on an insurer’s guidelines. Insurance use cases like this for LLM technologies will likely emerge in the coming months.
Henry: How else could LLMs change how insurers work?
Bas: Currently, most insurers do not have a full insight into which endorsements or modified terms and conditions have been made in which policies. Large portfolios of policies sit in PDF-based policy documents. LLMs could help actuaries understand and manage risk across a large portfolio more efficiently.
For example, an actuary could ask a LLM chatbot how many policies a portfolio has with a pandemic clause that is not up to date with the latest standard. Then those policies could be set aside and the clause changed upon renewal.
Henry: How is Instabase involved in LLM technology?
Bas: Instabase’s technology is modular, which means that we can easily switch in and out new technologies when they are developed. For example, if Google or Microsoft develops an improved optical character recognition model, we can start using that for our customers. Instabase was one of the first companies to integrate deep learning technology and aims to harness the power of the latest technologies in a practical, safe, enterprise-level way. We are actively integrating LLM technology in our product offering over the coming year.
Henry: Why has Instabase joined InsTech as a corporate member?
Bas: Over ten years working in insurance I have come to know that the insurance industry relies on relationships of trust. Instabase is looking to connect with insurers, claims management companies and brokers. Participating in InsTech’s community is a fantastic way to establish those relationships and build trust within the insurance community.
Henry: What sort of companies would Instabase like to connect with?
Bas: Most participants in the insurance market have workflows involving highly variable documents and complex workflows. We welcome any conversation with companies looking to explore how intelligent automation can improve their processes. You can get in touch with Instabase through InsTech, or reach out to Instabase directly to schedule a demonstration here.