Aipedia: Actuaries

Relying on third-party AI models

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

Actuaries can use vendor and open-source AI models, but the FRC says the practitioner must still ensure the model is fit for purpose. Vendor contracts often limit what the vendor will pay, so the actuary and their PI can end up carrying the loss.

Part of: AI in Actuaries

In short

Actuarial teams may buy rather than build: licensed pricing and reserving tools, catastrophe and mortality models, open-source machine learning libraries and generative AI services. The FRC Models guidance (2.12 to 2.13) says the practitioner must ensure fitness for purpose, may rely on the provider's TAS 100 evidence, and treats open-source code as in scope. If a third-party model fails, the client will usually look to its actuary first. Recovery from the vendor depends on the contract, which commonly caps liability. Actuarial PI usually responds to the client's claim subject to the policy terms, but the gap between vendor and actuary is worth managing up front.

How actuaries rely on third-party AI

Third-party models are common in actuarial work. Consultancies may license pricing and reserving platforms with built-in machine learning, use external mortality and catastrophe models, build on open-source libraries for gradient boosting or neural networks, and use generative AI services for drafting and coding. In-house teams use the same tools, sometimes chosen and configured by IT or data teams rather than actuaries.

The advantages are real: tested code, regular updates and specialist expertise. The risks are that the actuary sees less of the model's inner workings, updates can change results without warning, and the vendor's contract was written to protect the vendor. The Lloyd's Market Association's discussion paper recommends model ID cards and version control, which matter even more when the model comes from outside.

The duties that apply

The FRC Models guidance (2.12 to 2.13): the practitioner must ensure a third-party model is fit for purpose, may rely on evidence from the provider that supports TAS 100 compliance, and should treat open-source code as within scope.
TAS 100 P5.1 to P5.4: understanding, governance, bias and limitations apply to bought-in models as well as home-built ones. If understanding cannot be achieved, the guidance says to reconsider fitness for purpose (3.7 to 3.8).
The IFoA data science and AI guide says "care should be taken when using third party generative AI tools" (2.14 to 2.15) and stresses clear governance and agreed responsibility (2.11, 2.29).
Code Principle 2: competence and care, which includes the decision to rely on a tool.

What can go wrong

Illustrative scenario: a consultancy uses a licensed machine learning reserving platform for several general insurance clients. The vendor releases an update that changes a default setting for how large claims are treated. The release notes mention it briefly; no one at the consultancy notices. Reserves for two clients are understated at the next valuation. The clients claim against the consultancy. The consultancy looks to the vendor, whose licence caps liability at a year's fees and excludes indirect loss.

Other risks: an open-source library with an undocumented change in behaviour, a vendor that cannot or will not explain its model, a generative AI service whose terms allow it to use inputs, and a vendor that goes out of business, taking its documentation with it. See machine learning pricing models and capital model validation.

Vendor terms, PI and the gap

Clients rarely sue the software vendor; they usually claim against the actuary or firm that gave the advice. A claim of that kind would usually fall within actuarial PI, subject to the policy terms. The firm, or its insurer, may then try to recover from the vendor, but recovery depends on the licence: caps, exclusions and governing law commonly limit it. That is the gap.

Points to look at when negotiating vendor terms: liability caps relative to the exposure you take on, warranties about model performance and documentation, notice of material changes, data handling and confidentiality, intellectual property indemnities, and access to evidence that supports TAS 100 compliance. Your own client terms matter too; see PI and liability caps in your terms. Some PI policies contain clauses about AI or technology; see AI exclusions and clauses. More on cover: does actuarial PI cover AI mistakes? and PI for actuaries.

Practical controls

Keep a register of third-party and open-source models, with owner, version and purpose.
Obtain and file the vendor's evidence supporting TAS 100 compliance.
Run independent checks or benchmarks against the vendor's output.
Read release notes and re-test after updates before using them in client work.
Negotiate change notification, documentation and liability terms.
Apply the same discipline to generated code; see LLM-written code and spreadsheets.

How Apex helps

Apex arranges PI for actuarial firms that rely on third-party models; a broker can review how your policy responds and where vendor terms leave you exposed. See the AI in actuarial work overview.

Sources

Frequently asked

Can I rely on a vendor's statement that its model is TAS 100 compliant?

The FRC Models guidance says you may rely on evidence from the provider, but you remain responsible for deciding the model is fit for purpose for your use. File the evidence, understand the model's limitations, and carry out your own checks proportionate to the work.

Does open-source code count as a third-party model?

Yes. The FRC Models guidance treats open-source code as within scope of its third-party considerations and suggests documenting open-source code as part of additional governance for AI models. Version control and testing after updates are sensible controls.

If a vendor model fails, who pays the client?

Usually the client claims against the actuary or firm that advised it, and that claim would usually fall within actuarial PI, subject to the policy terms. Recovery from the vendor then depends on the licence, which commonly caps or excludes liability.

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.