Claims Reserving

The reserves on an insurer's balance sheet are not a number; they are an opinion, an actuarial estimate of what claims already incurred will ultimately cost to settle. When those reserves are wrong by even a few percentage points, the impact on reported profit and regulatory capital is measured in hundreds of millions. Claims reserving is where actuarial judgement meets financial materiality, and the stakes of getting it wrong make actuarial modelling in reserving one of the most consequential applications of AI in insurance.

What is claims reserving?

Claims reserving is the process of estimating the ultimate cost of claims that have already occurred but are not yet fully settled. This includes reported claims where the final settlement amount is uncertain (reported but not settled, or RBNS) and claims that have occurred but have not yet been reported to the insurer (incurred but not reported, or IBNR). Together, these reserves represent one of the largest liabilities on an insurer's balance sheet and are a primary focus of regulatory and audit scrutiny.

The traditional methods are well-established. The chain ladder method projects future development from historical development patterns. The Bornhuetter-Ferguson method blends the chain ladder projection with an a priori expected loss ratio, reducing volatility in immature accident years. Both methods use development triangles, matrices that show how claims in each accident year develop over successive evaluation periods, as their primary input.

The difficulty lies in the tail. For short-tail lines like motor physical damage, claims develop and settle within months, and the reserve uncertainty is modest. For long-tail lines like employers' liability, professional indemnity, and casualty reinsurance, claims can take a decade or more to reach ultimate settlement. The further into the future the estimate extends, the wider the confidence interval. UK insurers hold an estimated 150 billion pounds in general insurance technical provisions, and a one percentage point swing in reserve adequacy across the market represents over a billion pounds in aggregate profit impact.

The landscape

IFRS 17 has increased the frequency and granularity of reserving calculations. The standard requires measurement of insurance contract liabilities at current fulfilment value, updated at each reporting date. This means reserving is no longer a quarterly or annual exercise; it is a continuous process that must reflect current conditions, current discount rates, and current claims experience. The actuarial workload has increased substantially, and the demand for more responsive reserving methods is a direct driver of AI adoption.

The PRA pays close attention to reserve adequacy, particularly for long-tail lines where adverse development can emerge years after the reserves were set. Supervisory reviews assess whether firms have adequately recognised deteriorating trends, whether reserving methodologies are appropriate for the business mix, and whether there is evidence of management optimism in reserve estimates. Firms that are found to have under-reserved face requirements to strengthen reserves, with immediate capital and profit impact.

Social inflation, particularly in US casualty lines, has driven reserve strengthening across the London market and European reinsurers. The trend of increasing jury awards, broader interpretations of liability, and the growth of litigation funding has caused actual loss development to exceed projected development on accident years from 2015 onwards. Reserving actuaries are grappling with how to incorporate a trend that is partly social, partly legal, and partly economic into models calibrated on historical data from a different environment.

How AI changes this

Claims text mining extracts predictive signals from unstructured claims data, adjuster notes, legal correspondence, medical reports, and expert assessments, that traditional reserving methods ignore. A claim where the adjuster's notes mention "litigation" or "solicitor instructed" has a materially different expected ultimate cost than one where the notes describe a straightforward settlement. NLP models that systematically extract these signals and incorporate them into individual claim reserve estimates improve accuracy at the claim level, which aggregates into better portfolio-level reserves.

Granular claims segmentation using ML identifies clusters of claims with similar development characteristics that traditional actuarial segmentation misses. Rather than reserving all employers' liability claims as a single triangle, ML can identify sub-segments, occupational disease versus workplace injury, represented versus unrepresented claimants, specific injury types, that develop at materially different rates. More homogeneous segments produce more stable development patterns and more accurate projections.

Real-time development monitoring flags emerging trends earlier than quarterly triangle analysis. Instead of waiting for the next reserving cycle to identify that bodily injury severity is increasing or that a specific claims handler is settling claims above the benchmark, AI systems monitor claims development continuously and alert the reserving actuary when patterns deviate from expectations. This enables earlier intervention in treaty pricing and capital calculations.

Inflation and economic indicator integration feeds macroeconomic data, wage growth, medical cost inflation, construction cost indices, and judicial award trends, directly into reserving models. Traditional methods incorporate inflation through explicit assumptions set by the actuary. ML models can learn the complex, time-lagged relationships between economic indicators and claims costs, producing inflation-adjusted projections that update automatically as economic conditions change.

What to know before you start

Reserving models are among the most heavily scrutinised models in insurance. The PRA's expectations, IFRS 17's requirements, and external audit standards all apply. Any AI augmentation of reserving must be explainable, auditable, and validated to a standard that satisfies actuaries, auditors, and regulators. A black-box model that produces a reserve number without an auditable trail will not be accepted. Design for transparency from the outset.

Claims data quality is the binding constraint. Adjuster notes are inconsistent. Coding of claim types varies between handlers. Reserve movements may reflect administrative actions rather than genuine revaluation. Before training ML models on claims data, invest in understanding the data's limitations. A model trained on inconsistently coded claims data will learn the inconsistency, not the underlying patterns.

The chain ladder and Bornhuetter-Ferguson methods will not be replaced; they will be augmented. Regulators and auditors expect to see traditional methods as a baseline, with AI-derived insights providing supplementary information. Position your AI reserving work as enhancing the actuary's judgement, not replacing established methodology. The transition to AI-primary reserving, if it comes, will be gradual and driven by demonstrated accuracy over multiple reserving cycles.

Start with a claims text mining pilot on a single long-tail class where you have sufficient data volume and where adjuster notes are reasonably detailed. Employers' liability or professional indemnity are common starting points. Extract features from claims text, incorporate them into individual claim reserves, and compare the resulting portfolio reserve to the traditional actuarial estimate. The divergence, and the reasons for it, will tell you whether the approach adds value for your specific book of business.