Peter Murphy and David Burston investigate the extent to which artificial intelligence is being used within life insurers, and the level of skill within the profession
In 2021, to better understand the current use of artificial intelligence and machine learning (AI/ML) within the actuarial profession, the IFoA AI and Automation Working Party carried out a survey. This survey was conducted through a set of interviews and discussions with actuaries working at most of the UK’s largest insurers. We also carried out a less formal review of the life insurance industry’s use of these techniques.
While the response to the survey was relatively small (around 50 people), in conjunction with our wider review, we were able to obtain a reasonable picture of the current uses of AI within the life insurance industry and identify some interesting insights. Our key conclusions are that:
- For the applications of AI/ML, life insurance is lagging behind non-life.
- Within life insurance, the use of AI/ML is currently relatively narrowly focused on pricing and claims analytics.
- Key next steps will be further upskilling, better understanding of data, and improvements in data availability and quality.
Limited uptake
Just over half (27 out of 50) of survey respondents were from life and health insurance, with a significant number of replies coming from non-life actuaries.
Use of AI/ML within the life industry was limited, with just over half of life respondents stating that they had deployed any solutions, compared with around 80% of non-life respondents. The solutions deployed had mostly concentrated on predictive modelling and claims analytics. This was consistent with our interviews, which indicated that the areas in which AI/ML was most commonly used were pricing, product development, and claims analytics and fraud detection. A very small number of respondents indicated use in other traditional actuarial areas, such as reserving and capital management, asset liability management and investment; however, during the interviews we carried out across the industry, we struggled to find much evidence of implementation in these areas.
In summary, uptake was limited to areas in which there was a real commercial imperative to use AI/ML, and where good quality data was already available. Outside these areas, there needs to be further upskilling and developments in techniques.
Individual skillsets
Very few people claimed to have no understanding of AI/ML, with 94% saying they had at least a ‘high level understanding of concepts and ideas’. However, more than half of respondents had no practical experience of applying these techniques. Although 44% of respondents stated that they had the skills to build AI/ML models, this group was skewed towards non-life. Our wider review suggests that this is far from representative of the wider profession; it seems there is relatively little expertise or experience in many areas of the industry.
Figure 1: Comparing expertise between life and non-life.
Even respondents with relatively advanced skills were still trying to improve them. There were three areas in which more than 60% of respondents wanted to improve their skills, with a consistent picture across life and non-life:
- Understanding concepts, opportunities and limitations from a commercial point of view.
- Understanding features of AI/ML systems from an actuarial point of view, particularly focusing on their risk and how to review them.
- Learning relevant coding to create AI/ML models.
In other words, even those who already had a real interest in the topic felt that there was much to improve. This is echoed by members of the working party, of whom only a small number have real deep experience.
Stumbling blocks
The biggest challenges to further use of AI, according to the survey, were the knowledge gap and lack of talent. This was especially true in life insurance, where fewer actuaries had the skills in question (perhaps because there was less commercial requirement for them).
Further barriers to adoption included a lack of knowledge and understanding among decision-makers. This was related to the perceived low interpretability and explainability of models.
The availability and quality of data was another concern. While concerns over data quality were similar between life and non-life, data availability appeared to be much more of an issue for life. This might, in part, explain the difficulty of applying AI/ML to life insurance problems that would superficially seem suited to these techniques. For example, if a company wanted to use AI to carry out balance sheet optimisation, it would need to have a clearly defined quantitative capital management framework, as well as tools to estimate the balance sheet across wide ranges of different strategies, both now and into the future. While it is straightforward to require the recalculation of the balance sheet under many different investment strategies, in practice this can be difficult for all but the simplest businesses.
There was some good news from the survey, however. For example, regulation and costs were not seen as major challenges for implementation.
Upskilling
Given that the biggest challenges are the knowledge gap and lack of talent, a key question concerns how members might upskill. The profession needs to support members to build AI/ML applications, and help more senior members to better understand the concepts and challenge the use of these techniques.
AI/ML techniques should be of natural interest to actuaries. They build on our traditional skills of data analysis and prediction, and offer us a wide range of opportunities. There are many routes in. Peter’s own personal interest emerged from investigating the uses of AI for capital modelling, and then progressed through several online courses and personal research.
It is good to see that the structured upskilling avenues available to actuaries are increasing. However, most respondents had not received any formal support from their employer – an experience they had in common with most members of the working party. Of those that had received employer support, the most common sources of development were internal training courses or online courses. While it was not common within our survey, more than 500 people have completed the IFoA’s Data Science Certificate.
These results suggest that the profession needs to continue its work to provide opportunities to members, but employers must also encourage and support actuaries to develop these skills. Younger actuaries are likely to gain AI/ML experience through exams and work, so it will be particularly important to encourage qualified and more experienced actuaries to develop their skills in these areas.
Peter Murphy is an actuary at Royal London and deputy chair of the IFoA AI and Automation Working Party.
David Burston is an actuary at Milliman.
Image credit | iStock