Aipedia: Actuaries

Machine learning in insurance pricing models

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

Actuaries can use machine learning in insurance pricing, provided they understand the model, test it for bias and can explain its outputs under TAS 100 Principle 5. FRC research published in 2023 found general insurance pricing was the heaviest user of AI in UK actuarial work, so scrutiny may land here first.

Part of: AI in Actuaries

In short

Gradient boosting machines and similar techniques can find patterns that a generalised linear model misses, but they are harder to understand and explain. TAS 100 P5.1 requires understanding of the model, P5.3 covers material bias and P5.4 limitations. The FRC Models guidance says that if a practitioner cannot achieve understanding they should reconsider fitness for purpose, and that opaque models need sub-group testing for bias. If a model misprices a book or produces unfair outcomes, the insurer client may look to the actuary or consultancy, and actuarial PI usually responds subject to the policy terms.

How pricing actuaries use machine learning

The FRC's research on AI in UK actuarial work found general insurance pricing to be the heaviest current user of AI and machine learning. Typical uses include risk models built with gradient boosting machines alongside or instead of a generalised linear model, demand and retention models, interaction detection and automated feature selection. Some teams may use machine learning to find signals and then rebuild them in a GLM that is easier to explain; others put the machine learning model directly into the rating engine.

Explainability tools such as Shapley values, partial dependence plots and variable importance charts help show which factors drive a price. They are aids to understanding, not a substitute for it: a chart of feature importance does not show whether a factor is acting as a proxy for something it should not.

The duties that apply

TAS 100 P5.1: understand the model. The FRC Models guidance (3.7 to 3.8) says that if understanding cannot be achieved, the practitioner may be unable to meet P5.1 and should reconsider fitness for purpose.
TAS 100 P5.3: material bias. The guidance treats algorithm and data bias as "especially significant" for opaque models and expects outputs to be tested by sub-group (5.18 to 5.20).
TAS 100 P5.2 and P5.4: governance and limitations. Additional governance may be needed for AI models (4.9): random seeds, test data sets and documented open-source code.
TAS 200 applies to technical actuarial work in insurance.
The IFoA data science and AI guide asks for outputs that are "readily interpretable" (2.15) and for bias to be identified, communicated and mitigated (2.17).

What can go wrong

Illustrative scenario: a consultancy builds a gradient boosting model for a personal lines insurer. Validation looks strong on the hold-out sample. After launch, the model underprices a segment where historic data was thin, and the book grows fast in exactly that segment. Separately, a postcode-derived feature turns out to correlate with a protected characteristic, and the insurer's conduct review asks why it was not picked up.

Other failure points: overfitting to a period of unusual claims experience, retraining that changes prices without anyone noticing, and a model that cannot be explained to the insurer's pricing committee or its regulator. Insurer clients are subject to the FCA's Consumer Duty, so they are likely to ask their actuaries to show that pricing outcomes are fair. See bias and fairness testing for AI models.

How PI usually responds

A claim by an insurer client that negligent model design or validation caused losses would usually fall within actuarial PI, subject to the policy terms. The questions that commonly decide such claims are what the actuary was engaged to do, whether the work met the standard of a competent pricing actuary at the time, and how much of the loss flowed from the insurer's own commercial decisions. Mispricing losses can be large, so limit selection matters; see PI insurance for actuaries. Where the same model or code is used for several clients, check how the policy's aggregation clause would treat related claims. Fines and regulatory penalties imposed on the insurer are generally a separate matter and may not be recoverable.

Practical controls

Write down why a machine learning model is fit for purpose, and what you would do if you could not explain a driver.
Record random seeds, training and test data sets, and code versions.
Test predictions and prices by sub-group, including proxies for protected characteristics where data allows.
Monitor live performance against the validation sample, with triggers for review.
Agree in writing who owns the model after hand-over, and what the client must monitor.
If you use a vendor or open-source model, see relying on third-party AI models; for the data feeding the model, see AI experience analytics.

How Apex helps

Apex arranges PI for pricing actuaries and consultancies, and a broker can check how your policy treats model-build and AI-assisted pricing work. See the AI in actuarial work overview.

Sources

Frequently asked

Can I use a gradient boosting model for pricing under TAS 100?

Yes. TAS 100 applies to AI and machine learning models, but it does not forbid them. You need to understand the model well enough to meet P5.1, test for material bias under P5.3 and communicate its limitations. If you cannot achieve that understanding, the FRC guidance says to reconsider fitness for purpose.

Are Shapley values enough to explain a pricing model?

They help, but they are not the whole answer. Shapley values show how factors contribute to individual predictions. They do not by themselves show whether a factor is a proxy for a protected characteristic, or whether the model will behave sensibly on new business. Use them alongside sub-group testing and judgement.

Who is liable if an AI pricing model loses the insurer money?

It depends on the engagement and the facts. The insurer makes the pricing decision, but if the actuary's model design, validation or advice fell below a competent standard and caused loss, a claim may follow. Clear scope, documented validation and limitations all help, as does PI that reflects the size of the books you advise on.

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.