Aipedia: Actuaries

AI in capital models and validation

Aipedia, by the Apex broking team · Last reviewed 2026-10-27

AI and machine learning can sit inside internal and capital models, but they need extra governance and independent validation under TAS 100 and the FRC Models guidance. Capital figures drive regulatory decisions, so a validation failure can be expensive for the insurer and the actuaries involved.

Part of: AI in Actuaries

In short

Insurers may use machine learning for proxy models, economic scenario calibration, dependency structures and exposure data in capital models. The FRC Models guidance says additional governance may be needed for AI techniques, including test data sets, random seeds and monitoring of self-updating models. Validators must understand what they are validating, which is harder when the model is opaque. The FRC's 2025 monitoring work includes internal model validation for non-life business. A capital error can lead to claims against a consultancy, disciplinary risk for individual actuaries and PI questions about who carries the liability.

Where AI sits in capital models

An internal model is a chain of components: data, risk calibrations, dependencies, aggregation and outputs such as the solvency capital requirement. Machine learning appears at several points. Proxy models approximate slow nested stochastic calculations. Clustering reduces large policy data sets. Algorithms help calibrate dependency structures or catastrophe exposures. The Lloyd's Market Association's 2025 discussion paper lists capital modelling among the main AI uses and flags black-box models, validation and key-person risk.

The difficulty for capital work is the tail. Capital is measured at extreme percentiles where data is thin, and a machine learning component that fits well across most of the distribution may be least reliable exactly where it matters. Validation has to probe that tail directly, and the validator needs enough understanding of the component to say why it behaves as it does, not just that its outputs look close to the full model on familiar scenarios. A component that cannot be explained to the board that approves the model is a governance problem before it is a technical one.

The duties that apply

TAS 100 P5.2: model governance; P5.1 understanding; P5.4 limitations.
The FRC Models guidance says "additional levels of governance may be needed where models use AI/ML techniques" (4.9) and asks for drift and bias in self-updating models to be documented (4.16). If understanding cannot be achieved, fitness for purpose should be reconsidered (3.7 to 3.8).
TAS 200 for insurance work; IFoA APS G1 for the Chief Actuary in non-life insurance and APS L1 for life actuaries.
The IFoA data science and AI guide stresses clear governance and agreed responsibility in multidisciplinary teams (2.11, 2.29), which is common where data scientists build model components.

The FRC's Actuarial Monitoring Programme 2025 work includes internal model validation for non-life business, although no AI-specific findings have been published.

What can go wrong

Illustrative scenario: an insurer replaces part of its life capital model with a machine learning proxy. A consultancy is engaged to validate it. The validation tests fit on the same scenario set used for training and finds it excellent. In a later stress, the proxy understates losses in combinations of risks that were rare in the training set, and the insurer's capital position has to be restated. The insurer asks why validation did not test out-of-sample tail scenarios.

Other risks: a self-updating component that changes behaviour between validation cycles, a data scientist who built the component leaving with the knowledge, and unclear responsibility between the model owner, the chief actuary and the validator. See relying on third-party AI models and AI in reserving, since reserves feed capital.

How PI usually responds

A consultancy that builds or validates a capital model component would usually look to its PI for a negligence claim by the insurer client, subject to the policy terms and any liability cap agreed in the engagement. Employed chief actuaries are generally not personally insured under a PI policy of their own; claims against the employer are a matter for the employer's arrangements, and personal regulatory exposure is separate. Where one validation approach is used across several clients, the aggregation clause matters. Losses linked to capital can be large, so check limits; see PI for actuaries.

Practical controls

Validate machine learning components on independent, tail-focused test sets.
Keep a model ID card for each AI component: purpose, data, owner, limits, version.
Record random seeds and retraining events, and revalidate after material change.
Make responsibilities between model owner, chief actuary and validator explicit.
Plan for key-person risk with documentation others can follow.
Consider independent peer review; see AI in peer review under APS X2.

How Apex helps

Apex arranges PI for actuarial consultancies doing capital and validation work; a broker can check limits, liability caps and how AI-related model work is declared. See the AI in actuarial work overview.

Sources

Frequently asked

Does TAS 100 apply to machine learning parts of a capital model?

Yes. The FRC states that TAS 100 is applicable to AI and machine learning models used in actuarial practice. Principle 5 covers understanding, governance, bias and limitations, and the FRC Models guidance suggests extra governance for AI components.

What should validation of an AI proxy model include?

Testing on data the model was not trained on, especially extreme scenarios where capital is measured, checks on stability across retraining, and a clear statement of limitations. The validator should be able to understand the component well enough to challenge it, not just compare outputs.

Is a chief actuary personally liable for an AI capital model failure?

Civil claims usually target the employer or the consultancy rather than the individual, but the chief actuary has personal professional duties under the Actuaries' Code and APS G1 or APS L1, and could face disciplinary action. How any claim is insured depends on the employer's and firm's arrangements.

Check your PI covers how you use AI

Tell us about your firm and the work you do. A broker will review how your current PI treats AI-assisted work, or arrange new cover. No fee for a quote. Or call 0117 325 0027.

Start a PI quote Request a callback

Apex Insurance Brokers Limited is authorised and regulated by the Financial Conduct Authority. Registered in England and Wales, company number 07014570. This page is general information about how AI is used in the profession and how insurance may respond. It is not legal, regulatory or insurance advice on your circumstances, and it does not guarantee that cover will be available or on what terms.