By Stefan Claus and Chris Wiltshire.
AI-driven pattern detection enables judgements to be codified, supporting sharper, focused human oversight, and freeing supervisors to trade their magnifying glass for a drone – scanning the market landscape quickly and consistently.
Applying this approach to 2024 Insurance Returns data quickly provided insights, targeting attention towards firms where it matters most. Plus keeping humans firmly in the loop ensures that speed and scale doesn't come at the expense of sound judgement. Overall verdict: improving supervisory efficiency and insight that supports market resilience, fair competition and policyholders’ protection.
Why reserves matter – for resilience and competition
An insurer’s financial strength is, at heart, the excess of assets minus liabilities. For general insurers, technical provisions (TPs) – best estimate of future claims – form the largest slice of liabilities. Setting TPs too low can make a firm seem overcapitalised and allow it to undercut more prudent rivals with cheaper premiums. This risks obscuring underlying vulnerabilities, distorting competition against well-reserved peers (potentially undermining our secondary competition objective), and weakening policyholder protection in times of stress (important to our primary objective of securing appropriate policyholder protection).
Early, systematic detection of anomalies helps supervisors target attention proportionately, ensuring consistent oversight and sustaining a level playing field for firms.
The challenge: spotting trouble in a sea of data
The UK general insurance market comprises 160+ firms and, in 2024, around £160–£170 billion of technical provisions, spanning 16 classes of business and roughly 775 firm-class combinations.
Traditionally, supervisors relied on manual reviews that applied expert judgement – part science, part art – to spot trends and patterns. But that breadth of data makes consistent judgements in identifying outliers hard and time consuming. Historically such reviews have typically taken around six months to complete and would capture only the largest firms.
Solution: bringing AI into the mix
To tackle this, we’ve started using AI to help codify our initial judgements. Think of it like switching from a magnifying glass to a drone: instead of painstakingly scanning each field, we can now survey the whole landscape quickly and consistently.
AI helps us turn what used to be subjective, manual reviews into a structured and repeatable process. Trained on thousands of reserving datasets from across the market, these AI models use pattern recognition to identify adverse trends and structural anomalies – such as unexpected shifts in claims development. These models then apply similarity scoring techniques to assess how closely new data aligns with known risk patterns, enabling a ranked confidence score for each detected issue. The AI model is able to scan hundreds of reserving triangles (defined below) – spotting patterns and surfacing insights in seconds. But crucially, a human supervisor is always in the loop, ready to interpret the results, challenge the findings, and provide feedback for continuous improvement.
And the results? The analysis can now be completed within four weeks – 85% faster!
Box A – Explainer: what is a ‘Reserving Triangle’?
A reserving triangle is a way of laying out claims data to track how claims from each origin year change in each development year. It helps insurers spot patterns and estimate how much more they might need to pay out in the future.
In practice, we often look at three main types of triangles:
- Paid triangle: shows how much money has actually been paid out so far. This helps us see how quickly claims are being settled.
- RBNS (Reported But Not Settled) triangle: shows the reserves for claims that have been reported but not yet fully paid. This gives an idea of what’s still outstanding and is sometimes referred to as ‘Outstanding Claims’.
- Claim provisions triangle: provides an estimate of the total expected cost of claims, including those that haven’t even been reported yet.
By reviewing these together, we can check if enough money has been set aside and spot any shifts in claims handling or reserving philosophy (how prudent firms are being within the range of reasonable estimates).
We also look at changes in the Incurred But Not Reported (IBNR) reserves. IBNR can be calculated by taking the claim provisions and deducting the RBNS. An IBNR reserve is the amount set aside to pay claims that have occurred but that have not yet been reported to the insurer.
The term Ultimate Claims is often used in analysis, and this refers to what an insurer expects to have paid out once all the claims have been settled. A current estimate of this amount can be calculated as the amount paid to date plus the current claim provisions.
What changed when we ran this at year-end 2024
As soon as the latest reserving data arrived, we ran the AI analysis across the market, then layered materiality and our risk appetite to focus supervisory follow ups. Three effects stood out:
- Clearer oversight. Using firms’ own data, with faster analysis turnaround times, we could hold more grounded discussions, earlier.
- Sharper detection. We picked up early signs of large or latent losses, development disruptions and inflation allowances—particularly in General Liability and Motor.
- Better data quality. Flagging anomalies prompted firms to correct their regulatory submissions and helped focus supervisory time where it counts.
Figure 1: PRA year-end 2024 review process
Illustrating the benefits through real-world case studies
AI-driven trend detection can help supervisors form and test hypotheses about what’s happening in the market. Here are a few examples:
General Liability – Hypothesis: allowances for inflation within RBNS are inadequate
We saw the paid-to-outstanding ratio rising – potentially a sign of weakening RBNS.
Chart 1: General liability – incremental gross paid to outstanding
The graph of the development of the general liability claims shows that the amount that has been paid, relative to the amount that is outstanding (reserves held in respect of known claims), is increasing at the same point of development for each of the more recent origin years (2019 to 2023). Graphically, the end point for each line is higher than where all the other lines ended at the same point on the horizontal axis.
Drilling down into individual firms and reviewing other metrics suggests that, faster post-pandemic settlement appears to explain much of the shift.
This is largely corroborated by publicly available information: in the UK the average time from filing a claim to a hearing has gone down (Civil Justice Statistics); and in the US there is clear evidence of falling ‘time-to-be-seen’ for private medical appointments since 2020/21 (Frontiers). However, in Ireland and Australia there is no solid evidence showing a decline in ‘time-to-be-seen’ (Vhi Healthcare and apha.org.au) in the private sector used by insurers.
Nevertheless, weaker-looking RBNS was often offset by stronger IBNR, providing reassurance that total claim estimates continue to be reasonable.
Verdict: reject hypothesis. The explanation is plausible, but we will continue to monitor whether changing external conditions are reflected promptly and proportionately. It also highlights the need for careful consideration of country exposures and trends when selecting suitable external benchmarks to support reserving.
Box B: Explainer – reserving development graphs
Each line in the graph represents a cohort of claims originating from a specific year. The horizontal axis shows the development year – how claims have evolved over time. For example, the position in relation to claims originating from 2023, at the end of 2024 is shown as the dot on the pink line at development year two.
The vertical axis captures the trend under investigation, such as the cost of claims paid, the reserves held for future claims or, as above, a combination of the two. This format allows us to observe both how claims originating from a single year have developed and how patterns shift across different origin years.
Marine, Aviation and Transport – Hypothesis: total claim reserves are weakening
Aggregate paid-to-ultimate and incurred-to-ultimate trends suggested ultimates weren’t keeping pace with paid or incurred developments.
Chart 2: Marine Aviation Transport – accident year – paid to latest ultimate
The graph of the development of the Marine, Aviation and Transport (MAT) claims shows that the amount that has been paid, relative to the latest ultimate (the total amount expected to be paid out when all claims have been settled), is increasing at the same point of development for each of the more recent origin years (2018 to 2019 and again in 2020 to 2024). Graphically, within each trend, the end point for each line is higher than where all the other lines ended at the same point on the horizontal axis. This implies that the reserves held are lower for the more recent origin years, and would be insufficient if paid claims in these years continue at the rate they have done in past years.
At firm level the picture was more mixed, with the trend only becoming clear at the market aggregate level. Furthermore, historical experience shows past reserve strengthening in this class, with recent years pointing to rising severities and complex, claims which are taking longer to settle.
Claim payouts in the MAT class of business have been rising year-on-year, with case reserves generally keeping pace, though pre-Covid ultimate estimates sometimes proved inadequate. Since 2020, the sector has faced mounting inflationary pressures – such as higher repair costs (IUMI Hull Inflation Index 2025), steel prices, and labour shortages – driving up claim severity. At the same time, competitive pressures have softened rates (The Insurer), raising concerns about underpricing and insufficient reserving, especially as claims become more complex and long tailed. Social and economic inflation has further increased litigation and settlement costs (AON), while the value and complexity of insured vessels and cargo have surged (Ship Universe), amplifying risk. New hazards, like lithium-ion battery fires (AVSAX), and external factors such as geopolitical conflicts and supply chain disruptions, have added to the uncertainty, with late-emerging liabilities and deteriorating aviation reserves highlighting the challenges of early reserving in this evolving landscape.
Verdict: a heightened level of attention may be appropriate. Given the presence of several evolving external factors, drawing on historical data, benchmarking, and anticipated future developments can help support the estimation of ultimates and IBNR relative to emerging costs.
Motor Vehicle Liability – Hypothesis: the allowance for inflation within RBNS reserves is insufficient
Paid-to-incurred ratios have edged up at common development points, which could indicate RBNS isn’t keeping up.
Chart 3: Motor vehicle liability – accident year – paid to incurred
The graph of the development of the motor vehicle liability claims shows that the amount that has been paid, relative to amount incurred (the total amount paid out to date plus the amount that is reserved in respect of known claims), is increasing at the same point of development for each of the more recent origin years (2016 to 2019 and again in 2022 to 2024). Graphically, within each trend, the end point for each line is higher than where all the other lines ended at the same point on the horizontal axis. This implies that the reserves held in respect of known claims are lower for the more recent origin years, and would be insufficient if paid claims in these years continue at the rate they have done in past years.
Yet we also observe reductions in ultimates that are supported by the paid and incurred data. Factors such as speed limit reductions (TfL), improved vehicle technology (MITRE) have led to fewer and less severe accidents, and streamlined claims processes enable firms to settle claims more quickly and reduce case reserves.
This positive trend is offset by rising claims inflation and medical costs since 2022 (Chart 4), as well as delayed claims from Covid that may be more expensive to settle.
Chart 4: CPI 12mth: Medical products, appliances & equipment (G) 2015=100
Nevertheless, while ultimates are falling in line with paid claims, RBNS reserves appear to be declining even more rapidly, which warrants ongoing scrutiny to ensure adequacy.
Verdict: reject hypothesis. The observed pattern has plausible, non-concerning drivers, though it remains important to reconcile changes in RBNS with settlement speeds and inflation pressures.
Back-testing: early days, but already promising results
In total, we opened discussions with 15 firms that appeared to have adverse reserving trends at 2023 year-end. Following the discussions, which were similar to the evaluation of the trends explained above, 11 of these firms had underlying adverse trends, and of these, eight experienced a reserve deterioration (ie reserves at 2023 year-end were underestimated) in 2024 – showing the value of early identification.
Conclusions
Codifying certain judgements through the use of AI, while maintaining human-oversight, brings speed, scale and consistency to a task that used to be slow and idiosyncratic. It helps us spot market wide shifts sooner, prioritise proportionately, and engage earlier with firms where signals point to pressure. That reduces the risk that under reserving quietly builds, improves comparability across firms, and supports a fairer competitive environment for well capitalised insurers, aligned to the PRA’s statutory objectives.
Furthermore, back testing demonstrated that early flags often preceded later reserve strengthening – exactly the pattern supervisors aim to catch.
Looking ahead
Continuous monitoring remains essential, especially in softer market cycles, to keep pace with evolving risks and ensure that modelling choices translate into sound prudential outcomes. And one thing we won’t be doing: letting the drone fly itself. The system scans the landscape, with humans firmly in control – making the judgements that matter for policyholder protection and market integrity.