With more than 25 years of experience in Speciality, Marine and other lines of the insurance industry, we understand insurance and the London Market well. Passionate about learning the most up-to-date trends and ideas, we bring them to you!

Read our most recent article with Paul Butler, Chief Technology Officer at Hiscox London Market, below to find out his views on how AI is affecting the insurance sector and what is being done to utilise this technology.

Paul Butler has been pioneering the use of artificial intelligence (AI) to transform underwriting processes. With remarkable advances in machine learning, Butler is well versed in implementing large language models to both insurance as a whole, and specifically, the underwriting process.

In a recent interview featured on our IT Insights InsurTalk series, Butler shared his insights on the implementation of AI-augmented underwriting, the challenges faced, and the lessons learned at Hiscox. Sounds interesting? Read on to learn about how AI is reshaping underwriting in a major Lloyd’s of London carrier.

Key takeaways:

• Hiscox London Market has integrated AI into its underwriting processes, significantly reducing data processing times and increasing accuracy.
• Collaboration with Google Cloud has enabled further advancements, such as using large language models (LLMs) to enhance underwriting speed and efficiency.
• While AI offers substantial benefits, Hiscox maintains a focus on augmenting, rather than replacing, human underwriters due to the complexity of specialised insurance areas.
• Challenges include managing AI model reliability and developing effective prompt engineering techniques.
• The future of AI in underwriting at Hiscox involves continued innovation while ensuring robust governance and oversight.

The current state of AI in underwriting

Hiscox London Market has been at the forefront of integrating AI into underwriting, with four years of development already put into testing and implementing AI into their underwriting processes. The initial focus was on using AI to automate the processing of submission emails and attachments.

Paul mentioned that before AI, “(…) we had third party operational people in India, about 20 people. And it would take them about two or three days to cleanse that data, to transform and geocode that data and get it back to us. The models are doing this in seconds“, Butler explained.

Key Developments at Hiscox Include:

• Neural Networks for Data Processing: Hiscox’s AI models handle submission data, transforming and geocoding it with a 98-99% accuracy rate, an improvement over the 95% accuracy previously achieved by human teams. This model “strips that data out, transforms it, geocodes it, and makes it available for the underwriters to start using“, according to Butler.
• Collaboration with Google Cloud: Hiscox developed a lead algorithmic underwriting proof of concept using Google’s large language models (LLMs) demonstrating the potential to improve their underwriting processes. “we’ve had some great results in using those as well” noted Butler.

The continued need for human expertise in underwriting

Despite these advancements, Hiscox remains cautious about fully automating underwriting decisions. “We’re calling it augmented underwriting for a reason“, Butler emphasised. “We’re not looking to replace the underwriters” said Paul, particularly in specialist areas like sabotage and terrorism insurance.

Their focus remains on equipping underwriters with tools that enhance their ability to make quick, informed decisions, saving precious time in the process. “We’re trying to do is minimise that time, put all that information that we’ve extracted and packaged together in front of the underwriter for them to make the decision“.

Paul Butler stressed that “I don’t think we’re ever probably going to get to a point where we are going to trust large language models” and that Hiscox is assessing the risks of using these models, and where the risks are minimal and the benefits appear high, this is the ideal point to implement LLM technologies into the underwriting process.

Hiscox’s collaboration with Google to minimise the risks and decision-making time in the underwriting time has so far yielded impressive results, with Paul explaining that “(…) we’ve got a three-day process down to a matter of minutes“, by accumulating the data, modelling and pricing it, and understanding their exposure in order to be able to make good, accurate final decisions on the model.

The evolution required to make underwriting decisions

Hiscox has learned that while AI can significantly aid the underwriting process, there are limitations to its use. Large language models “do hallucinate, they do have come up with incorrect answers sometimes and they’re very confident that they’ve got the correct answers“, Butler warned.

To address these risks, Butler explained a key learning gained at Hiscox:

To achieve high-quality, reliable results, it’s necessary not only to experiment with the prompts, but also the “personas” given to the model, the temperature, and the feedback provided.

Hiscox has developed robust governance frameworks and controls to ensure model accuracy and reliability. “We’ve spent a lot of time focusing on more recently is the governance, frameworks, the controls we’re going to need to have in place, [thinking about] how are we going to sort of monitor these models, make sure that they’re not drifting, make sure that they are accurate and they’re not hallucinating or coming up with the wrong answers“.

However, Butler remains sceptical about AI fully replacing human judgement in complex decision-making scenarios in the future. “They’re just too generic to replace highly skilled people”, he explained, adding that while there may be some highly tuned specific models that prove useful, “I don’t think those types of models are ever going to replace human decision making“.

The competitive advantages of Generative AI

As Paul Butler explained in detail, AI integration at Hiscox has yielded several competitive advantages:

• Speed to Market: The “actual real value” to Hiscox’s business has come from “getting the turnaround from a submission coming in from the broker [and] getting the quote back to the broker from three days to a matter of minutes“, according to Butler.
• Applications Beyond Underwriting: AI is also being explored for its use in other areas, such as claims processing. A pertinent example described by Butler was, “we could use satellite imagery to quickly assess if the damage to a property is genuine, see the ‘before’ and ‘after’ shots, and then pay the claim quickly“, saving both time and money, and improving customer satisfaction.

Lessons learnt from implementing AI

Butler shared several lessons from Hiscox’s journey with AI:

• Prompt Engineering is Key: Effective prompt engineering is essential to getting accurate results from AI models, but as Butler pointed out, “getting the prompt engineering right is complicated“.
• Continuous Model Updates are Necessary: AI models evolve quickly; “a year ago we were building with Google PaLM as a large language model. That’s been deprecated now. We then built it with Gemini Version 1. And now we’re retuning and reprompting for Gemini 1.5” explained Butler, emphasising the need to keep up with the constant LLM development curve with a highly skilled and dedicated team.

“Accuracy scoring” framework for underwriting

To ensure the reliability of AI models, Hiscox has developed an “accuracy score” framework. “For our neural network models, [accuracy scoring] is straightforward” according to Butler, but for LLMs, Hiscox relies on feedback from underwriters to refine accuracy. To do this, they utilise a data dashboard displaying the results of their LLM information, and with continuous feedback from underwriters, who highlight any inaccuracies in the information extracted by the AI, the data is scored on its reliability. Having this continuous feedback loop helps in fine-tuning the models and prompts, ensuring consistent performance.

Preparing for AI integration and augmentation

Successfully integrating AI requires more than technological expertise – it demands a supportive culture that “(…) is willing to experiment and invest in that experimentation“, and is to “accept failure” before achieving success. At Hiscox, fostering such a culture has been key to their AI initiatives, with “stepping outside their comfort zone” a key requirement to achieving their end goals.

Picturing the future of AI-augmented underwriting

Looking ahead, Hiscox plans to build on the foundations already in place. With a solid grounding in a range of skills such as data science, machine learning, data engineering and so on, Hiscox aims to bring people together based on their capabilities to build a flexible structure aimed at efficiently solving specialised tasks to streamline the AI-assisted underwriting process effectively: “(…) for us as a London Market carrier, underwriting is our biggest value stream“, reported Paul, so building excellent underwriting tools is a high priority.

While these changes are promising, Hiscox are not seeking to radically shake up their current underwriting structure, with Butler definitively stating “the way we’re organised, the way we’re structured, it does work. I can’t see that changing for the next few years even with all the AI stuff going on”.

Summary

By leveraging advanced technologies like neural networks and large language models, Hiscox has proved itself an industry leader in dramatically improving the efficiency and accuracy of the underwriting process, reducing manual workloads and speeding up decision-making.

However, the importance of human expertise is not to be overlooked, as AI and LLMs are unlikely to replace human skills anytime soon. Rather, they can be utilised to augment current talents, speeding up and improving the underwriting process in areas that were previously time intensive.

As AI models and organisational structures continue to be refined, the focus must remain on balancing innovation with careful governance, ensuring that the benefits of AI are fully realised while minimising risks.

If you would like to watch our full interview with Hiscox London Market CTO, Paul Butler, follow this link to head over to our IT Insights Hub and enjoy our conversation on the practical implementation of AI augmentation in full.

With over two decades of expertise in data harmonisation, digital solutions, consulting, AI, and cloud services, Future Processing is the ideal partner to help you bring your digital initiatives to life. We collaborate closely with you to understand and achieve your goals, ensuring your success in the market. If you’re interested in discovering how our industry-leading digital services can benefit your business, contact us today, and we’ll work with you to find the best solution for your unique needs.