FCA authorised · FRN 724952 0117 325 0027Goods in transit: buy onlineQuote & buy →
Apex Insurance Brokers
Speak to a brokerGet a quote →

Aipedia: Actuaries

AI in reserving: risks for actuaries

Machine learning reserving methods can be used alongside traditional techniques, but the actuary remains responsible for the reserve and must understand, govern and explain the model. Reserve deficiencies often surface years later, which makes AI reserving errors a long-tail PI exposure.

In short

Reserving actuaries may test machine learning models that work on individual claims data rather than aggregated triangles. These can react faster to changes but are less transparent and can be unstable when retrained. TAS 100 Principle 5 applies, and the FRC Models guidance calls for extra governance for AI models, including monitoring drift in self-updating models. If a reserve proves deficient and the model cannot be explained, the actuary's position in a later claim is weaker. Actuarial PI usually responds to reserving negligence claims, subject to the policy terms, but claims often arrive long after the work.

Using AI in your work? Make sure your professional indemnity covers it. A specialist broker can check your wording and find cover that does.
Get a second opinion on your PIDoes PI cover AI mistakes?

How actuaries use AI in reserving

Traditional reserving relies on aggregated triangles and methods such as the chain ladder method and its variants, with judgement applied to development patterns and tail factors. Machine learning approaches work on claim-level data: predicting the ultimate cost of each open claim, estimating IBNR from claim features, or flagging claims likely to deteriorate. The Lloyd's Market Association's 2025 discussion paper lists reserving among the main AI uses for actuaries in that market.

In practice many teams may run machine learning alongside established methods rather than instead of them, using it as a challenger or an early warning. That is sensible: the comparison is itself a control, and differences between the two methods are a prompt to ask why. The risk grows when the machine learning output becomes the selected reserve without the same level of scrutiny the traditional method would have received.

The duties that apply

What can go wrong

Illustrative scenario: a consulting actuary introduces a claim-level machine learning model for a commercial liability book. It produces lower reserves than the chain ladder, and the client is keen to adopt it. The model was trained on a period before a change in the client's claims handling and before a rise in claims inflation, so it learned settlement patterns that no longer apply. Two years later, reserves are strengthened sharply. The client's board asks why the traditional method's higher figure was set aside.

Other problems include case reserve changes that the model treats as signal, small classes where the model has too little data, and retraining that silently changes the result between quarters. See also AI experience and claims analytics.

How PI usually responds

Reserving negligence claims would usually fall within actuarial PI, subject to the policy terms. Because PI is generally written on a claims-made basis, the policy in force when the claim is made, not when the work was done, is the one that usually responds. That makes continuous cover, the retroactive date and run-off cover important for reserving work. Insurers may look at whether the method was reasonable at the time and how the choice was documented and communicated. If you become aware that a past reserve may be wrong, the policy may require you to notify a circumstance promptly. See PI for actuaries.

Practical controls

How Apex helps

Apex arranges PI for reserving actuaries and consultancies; a broker can review continuity of cover, run-off and how AI-assisted reserving is treated. See the AI in actuarial work overview and whether PI covers AI mistakes.

Sources

Frequently asked

Can a machine learning model replace the chain ladder for reserving?

It can be used, but keeping a traditional method alongside as a check is sensible. TAS 100 requires you to understand and govern the model and communicate its limitations. Running two methods and explaining the differences is good evidence of that, and helps if the reserve is challenged later.

What does the FRC say about self-updating reserving models?

The FRC Models guidance says self-updating models should have drift and bias documented, and that additional governance may be needed for AI and machine learning models, such as recording random seeds and using test data sets. It is guidance, not a mandatory standard.

When does a reserving claim usually arise?

Often years after the work, when a deficiency emerges or the business is sold or reviewed. Because PI usually responds on a claims-made basis, you need continuous cover with a suitable retroactive date, and run-off if you stop practising.

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.