The UK protection market is increasingly focused on underwriting efficiencies, particularly around the use of artificial intelligence (AI) solutions to summarize lengthy medical reports and support post-issue underwriting triage.
With the potential efficiency savings on offer, insurers are faced with the dilemma of whether to build in-house or turn to a third party. This paper explores the pros and cons of each option and the considerations along the way.
A lack of experienced underwriters and the growth in electric general practitioner (GP) reports is creating a resource challenge for insurers. Before the COVID-19 pandemic, GPs returned approximately 35%-40% of reports in a digital format. Now, with increased use of electronic health records (EHRs), that number is more than 60% for some insurers.
While EHRs provide richer information, they average close to 80 pages in length. This increases assessment time and decreases underwriter productivity, contributing to new business backlogs.
AI solutions can provide a succinct summary, highlighting key risks. This offers significant potential productivity gains by reducing assessment time and enabling underwriters to focus on decision-making. Some established solutions already are able to align the summary with an insurer’s philosophy and suggest transparent decision recommendations. The underwriter continues to make the final decision in all circumstances.
Over the past three years, post-issue underwriting volumes have increased and now represent close to 5% of all auto-accepted policies in the UK. Often this involves requesting EHRs, adding to the volume of information that must be manually reviewed. While the majority of reports requested at this stage do not contain evidence of misrepresentation, all reports require a manual review to confirm.
AI solutions can be deployed to triage these cases, comparing application disclosures with customers’ health information and confirming the absence of any inconsistencies, thereby removing the requirement for a manual review by an underwriter. This results in efficiency gains in time and cost.
Post-issue underwriting is a logical and low-risk stage in the process of introducing AI solutions, but focusing development only in this area can limit potential long-term efficiency gains. While pre-issue summarization may be more challenging to develop initially, once the technology is effective and can interpret medical information accurately, it is easier to “bolt on” additional services, such as post-issue triage, than it is to approach it the other way.
Beyond pre- and post-issue implementation, a wider application often remains unconsidered and potentially underestimated. Once insurers have successfully implemented a solution that can rapidly analyze EHRs at scale, the opportunity exists to review policies previously underwritten to identify new trends and insights. This might take the form of protective value analysis around non-medical limits or new insights into disease progression.
As industry veterans can attest, innovation is easy; integration is the difficult part. This often rings true when working with third-party solutions and navigating integration challenges with existing software and legacy systems, including data ownership ambiguity and security risks.
Once these technical hurdles are addressed, the business should consider how an AI solution can be safely introduced to improve efficiency while continuing to manage risks.
Traditional treaty wording around underwriting errors and reinsurance recovery will not reference AI solutions, nor where liability lies in the event that unforeseen material risks result in a related claim. Advance discussions with reinsurers should demonstrate that rigorous testing has occurred and that ongoing audit plans are in place.
Introducing AI technology in a live environment with new business cases takes careful planning. One option is to mirror the approach of a trainee underwriter, focusing on less complex products initially, such as life cover only with limited sum assured, and increasing the sum assured and complexity over time. In this scenario, insurers could consider a ceiling on the max sum assured for AI processing without human intervention.
It will take time – if it ever happens – for (re)insurers to become comfortable relying on AI summaries on high net worth (HNW) cases, where the stakes are highest with potential missing or inaccurate information and AI model hallucination.
HNW cases present unique challenges that require a nuanced underwriting approach, often with a commercial perspective that AI is unable to offer. Potential impact and efficiency gains from training AI models on electronic GP reports are diminished when assessing HNW cases. HNW cases tend to consist of a variety of unpredictable evidence sources, not just medical, which may require a manual review. Deploying an AI solution on only part of the case reduces effectiveness overall and introduces additional complexity.
AI underwriting solutions have yet to achieve business-as-usual status in the UK, but some show great potential and have proven track records. Often, the evidence comes from the US, where the market, evidence types, and medical terminology differ and require UK calibration. However, if the underlying models are effective, application in the UK should require only minor adjustments and training.
UK-based providers are offering to work with insurers during the development phase. However, without a live deployment in any territory, the process of design, development, and extensive testing through various iterations can delay efficiency gains.
Consideration of an external solution should account for any cross-border data transfers to ensure compliance with UK GDPR and ICO guidelines. While some solutions may be hosted via a UK-based cloud solution, manual quality assurance may still be carried out overseas.
From a commercial perspective, working with a third party may appear more cost effective initially and accelerate implementation. This should be weighed against significant and expensive integration challenges and ongoing fees that could escalate over time. Trade-offs in terms of IP and data-sharing will likely play a part in future model enhancements, and third-party arrangements risk reduced scope to differentiate against competitors if multiple insurers are using the same solution. Finally, before implementing reinsurer-led solutions tied to active reinsurance agreements, it is important to consider whether the arrangement does not restrict future flexibility during tender negotiations with the wider market.
Custom solutions may be developed independently by insurers using in-house data science resources or created in collaboration with external technology/AI specialists. These may be more easily integrated into existing technology architecture.
This approach helps to protect IP, differentiate offerings, and promote a commercial edge in an increasingly commoditized market. In addition, and importantly, it reduces risks associated with external transfer of sensitive data to other territories that may have different data security standards and regulatory environments. While US solutions may host their technology in UK cloud services, some quality assurance may still be conducted outside the UK.
However, developing in-house presents resource challenges. While most insurers have data scientists, they typically serve the wider business, not just underwriting, and are unable to dedicate sufficient time to large-scale projects.
As a result, this approach may lengthen development timelines and delay implementation at smaller insurers. It may be more attractive to larger insurers with extensive resources.
Regardless of the buy-vs.-build scenario, AI regulations are an essential consideration. As a global reinsurer, 91˿Ƶ is aligned across jurisdictions with the EU AI Act, NIST AI Risk Management Framework, and ISO 42001 to ensure AI is built, deployed, and used responsibly in all markets.
Large language models (LLMs) continue to rapidly develop, becoming increasingly complex and intelligent. Previously, LLMs specially trained on large amounts of medical data outperformed more generic versions at interpreting medical terminology from GP reports. However, due to rapid acceleration of model development, more generalized LLMs, such as Anthropic’s Claude, are now delivering similar results.
It is plausible that, in the near future, large-scale global AI companies with significant resources and investment could overtake medically trained LLMs provided by niche insurtech-type companies.
The market demand for AI solutions in the UK is only gaining momentum. While there may not be a “one-size-fits-all” overnight solution available currently, insurers are committed to finding solutions to help ease resource challenges and deliver efficiency gains across the value chain.
While comparisons have been made to the slow adoption and rollout of electronic GP reports in the UK over the past 10 years, the scenario with AI solutions appears different due to market pressures and global demand. AI adoption is poised to become ubiquitous in the UK over the next five years.
This is only part of the picture. The real opportunity to leverage the full capabilities of AI will occur when alternative methods of accessing health information are available. Directly accessing medical records and assessing in real time at the point of sale remains the ultimate aim and future catalyst for revolutionary change in our industry.
91˿Ƶ is a market leader in providing expert guidance relating to the adoption of AI solutions. We look forward to working with you to discuss your options and considerations in more detail. Contact us.
