This summer, I had the chance to work with Charlie Lilley, who joined Jaywing as part of Stubben Edge’s internship programme. We tasked him with a series of projects designed to give him exposure to different aspects of risk modelling, from decision automation and claims risk to researching opportunities in the credit insurance market.
Over the course of his placement, Charlie used our Archetype platform to build and test models, looked into how AI can support fair and explainable decisioning, and even pitched a new product concept back to the team.
I caught up with him to find out what drew him to Stubben Edge and Jaywing, what he’s learned along the way, and how he sees AI shaping the future of risk modelling. Here’s what he had to say:
Q1. What drew you to the placement, and what areas of risk and modelling most interest you?
What stood out to me was Stubben Edge’s ambition to build as much of the insurance value chain in-house as possible, rather than relying heavily on third parties. It felt like a business determined to innovate. Learning about Jaywing’s expertise in advanced modelling and analytics made the opportunity even more appealing. I wanted to see how techniques I’d studied could be applied to real business problems, particularly in credit risk, and to understand how risk management underpins financial decision-making. Strengthening those skills felt essential for a future career in finance.
Q2. From your perspective, what are the opportunities for automating underwriting decisions, and what risks do firms need to watch?
The potential for underwriting automation is huge. Insurance applications are structured, the data is rich, and AI models can deliver speed, consistency and scale in ways manual processes cannot. But the risks are just as real: bias creeping into models, over-reliance on data without proper oversight, and the constant pressure of regulatory standards. The real challenge is balance: designing models that enhance efficiency without losing the fairness and reliability that human judgement provides.
Q3. How did you approach building and testing the accept/reject decision model, and what did you learn from comparing underwriter and model decisions?
The work reinforced to me how much of modelling is about the data foundation. I spent most of my time assessing the tables, cleaning the data, and identifying the strongest variables before running the build in Archetype with parameters like target outcome and train/validation/test split. I didn’t reach the stage of fully testing or comparing model outputs with underwriter decisions, but the process was still revealing. It showed how each modelling choice, from variable selection to parameter setting, shapes the outcome and that the real measure of success is whether the model aligns with the practical decisions underwriters make every day.
Q4. You also looked at modelling claims risk. How should insurers think about predicting claim probability versus claim value?
Claims risk has two very different dimensions: how often something happens and how severe it is when it does. Treating probability and value separately helps insurers understand the drivers of frequency versus cost, which are often distinct. But to get a full view of exposure, the two need to be brought back together. I see it as a two-step process: model them separately to get clear insights, then combine them to create a realistic picture of overall claims risk for pricing and capital decisions.
Q5. Based on your research, what factors most influence claims risk modelling, and how does AI/ML add value here?
The strongest predictor of future claims is past claims history. Coverage type and limits also play a major role, as complex products or high limits naturally increase potential claim size. Other important factors can sit at the policy level, such as what is insured or the duration, at the customer level, such as creditworthiness, or externally, like economic conditions or natural disasters. AI and machine learning add value by bringing all of these elements together, finding non-linear relationships, and detecting hidden patterns that traditional statistical methods might miss.
Q6. What did working with techniques like neural networks and gradient boosted models teach you about the balance between predictive power and explainability?
Building my first model with XGBoost showed me how powerful modern methods can be at capturing complex patterns in structured data, while still providing insight into which variables matter most. Even without the time to fully interpret the outputs, the exercise highlighted the trade-off: more complexity can improve accuracy but often makes results harder to explain. The most effective models are those that get this balance right, accurate enough to drive meaningful insight but explainable enough for humans to trust and act on.
Q7. You also researched credit insurance. What stood out to you about the leading players, and where do you see gaps in the market?
Credit insurance was one of the most interesting areas I worked on. I went from knowing very little about the market to understanding the role of the leading insurers, brokers, and banks. What stood out was how the leading players are already weaving AI and modelling into their products, which shows the direction the market is heading. I also found useful insight in a survey of 48 banks, which explained why they use credit insurance, the benefits they see, and concerns around regulatory changes such as Basel 3.1. In terms of gaps, there is scope for climate-focused products, SME-targeted solutions, and better use of technology to make credit insurance more accessible. I even developed a product concept around these findings, building out a pitch deck and presenting it back to the team.
Q8. How do you think AI-based decisioning could strengthen or disrupt the credit insurance sector in the next few years?
AI has the potential to reshape the way the entire market operates, not just how individual applications are assessed. On the positive side, it could allow insurers to analyse broader datasets, spot global trade and credit exposure patterns, and make cover more widely available and better priced. The risk is that only the largest players may be able to invest at scale, widening the gap with smaller firms in a market that already has significant divides. At the same time, smaller insurers may see AI as a way to differentiate themselves, and in some cases could end up investing earlier and more boldly than the big players in order to stay competitive. The key question is not whether AI will be adopted, but how to integrate it in a way that keeps the market fair and transparent.
Q9. At the end of your internship, you’re co-authoring a paper on decision automation in insurance. What are the key themes you hope it will address, and why are they important now?
I want the paper to highlight both the opportunities and the risks of decision automation. On one side, it can speed up processes, improve consistency, and allow firms to scale. On the other hand, there are issues to manage, such as potential bias, the importance of explainability, and the need to balance automation with human judgement. It is particularly exciting that Jaywing and Stubben Edge are already building and testing models that show how AI can be applied safely and effectively. If we can showcase both the potential and the safeguards, the paper will underline why this is such an important moment for the industry.
Q10. From your time at Jaywing and Stubben Edge, what’s been your biggest takeaway about how financial services firms can better apply AI in risk and decision-making?
The industry feels like it is at a turning point. Most firms are still experimenting with AI, but the shift to using it at scale in risk and decision-making is approaching fast. What I’ve seen at Jaywing and Stubben Edge is that they are not just talking about AI, they are actively building and testing models that show how it works in practice. The lesson for me is that success is not only about clever algorithms, but about clean data, explainable outputs, and embedding models into business processes. The firms that manage this will move faster, manage risk more effectively, and build greater trust with customers.
Q11. As someone at the very start of your career, what surprised you most about how risk and AI are applied in practice compared to how they’re talked about publicly?
In the media, AI is often described as replacing humans or making flawless decisions. My experience has been the opposite. At Jaywing and Stubben Edge, AI is there to support people, not replace them. For example, I worked on a model designed to improve underwriting decisions and speed up the process, but underwriters are still essential for validating results and handling complex cases. Tools like Archetype also make modelling more accessible to analysts and underwriters, proving you don’t need to be a machine learning expert to add value.
Q12. If you could fast-forward five years, how do you see yourself and your peers shaping the next generation of risk models?
As AI and machine learning mature, more complex models that are currently too difficult to implement will become feasible. My generation will probably spend less time creating brand new models and more time improving, validating, and scaling them across firms. The focus will be on ensuring that risk models are effectively embedded in decision-making to deliver real business value. Our role will be to ensure that these tools are not only technically advanced but also practical, reliable, and widely adopted.
Q13. What’s been the most enjoyable or eye-opening part of the internship so far?
The most eye-opening part has been the people. On my first day, I was thrown straight into a meeting with a leading investment bank where I was initially out of my depth, but that pushed me to learn quickly. I always thought I worked best with my head down, but I realised how valuable conversations with colleagues, managers, and clients really are. Far from slowing me down, they gave me context and motivation, making the detailed tasks more rewarding. Speaking to people across different teams and levels also meant I was exposed to many aspects of the business I wouldn’t have seen otherwise, which gave me a far more rounded and diversified experience than I expected - and certainly kept me busy!
Q14. What advice would you give to another student or graduate who’s curious about getting into risk modelling and AI?
Give it a try. If you are open to learning, willing to make mistakes, and ready to teach yourself what you don’t know, it can be incredibly rewarding. Good analytical skills and an interest in working with data help, and those skills improve quickly once you get started. Tools like Jaywing’s Archetype make modelling more accessible and faster to use, which helps break down the barrier for people new to the field. For anyone interested in finance, it is particularly valuable because risk modelling underpins everything from investment banking and asset management to credit analysis and insurance. Beyond finance, it is also a skill set that transfers into consulting, fintech, and even entrepreneurship, so it is a strong foundation for many different career paths.
And a big thanks from us
Charlie’s time with us has been a reminder of how quickly new talent can contribute fresh ideas and energy to complex topics like decision automation, claims modelling and credit insurance. What stood out most was his ability to link technical modelling choices back to practical business challenges, and to balance curiosity with a focus on explainability and fairness.
As we continue to push the boundaries of how AI and advanced analytics are applied in risk, it is encouraging to see the next generation thinking so carefully about both the opportunities and the responsibilities that come with these tools.
Thanks to Charlie for his work this summer, and for sharing his reflections in this interview. We look forward to welcoming the next cohort of interns. Watch this space!