In just a few years, large language models (LLMs) have gone from laboratory curiosity to everyday tools in many professions. Actuaries are no exception. Before bringing them into technical work, it pays to understand what they are, what they do well and, above all, where their limits lie.
What an LLM is, and what it is not
An LLM is a statistical model trained on enormous amounts of text to predict the next word (or token) given a context. Useful abilities emerge from that seemingly simple capability: drafting, summarizing, translating, explaining concepts, generating code and answering questions. Tools such as ChatGPT, Claude and Gemini are interfaces built on models of this kind.
It is essential to be clear about what an LLM is not. It is not an exact calculator, it is not a database of verified facts and it does not reason like a credentialed actuary. It generates plausible text, not necessarily correct text. That distinction governs its entire professional use: the model assists, the actuary is accountable.
Real use cases for the actuary
Framed properly, LLMs save time on tasks where the final judgment remains human. Some uses already seen in practice:
- Drafting technical documentation. Producing first drafts of calculation reports, methodology notes, executive summaries or sections of a valuation report. The model structures and drafts; the actuary verifies figures, assumptions and regulatory compliance.
- Help with Python and R code. Writing functions, debugging errors, commenting code, porting a spreadsheet calculation to a reproducible language or explaining what a piece of legacy code does. This is one of the uses with the clearest payoff, always subject to review and testing.
- Reviewing contracts and policies. Locating clauses, comparing versions, extracting coverage terms, exclusions or claim definitions for later analysis. Useful as a first screen, never as a legal opinion.
- Client communication. Translating a technical result into language a board committee or an HR department can understand, adjusting the tone of an email or preparing an FAQ.
- Data exploration. Suggesting analytical approaches, generating code to plot distributions, proposing consistency checks or helping interpret preliminary results.
The pattern is the same in every case: the LLM speeds up the first draft or the supporting work, and the professional supplies the expert judgment, the validation and the accountability.
Risks that cannot be ignored
Professional use of these tools brings concrete risks that an actuary must manage explicitly.
- Hallucinations. The model can invent figures, regulatory references, citations or formulas in very convincing prose. Where a misapplied rate or factor has financial and regulatory consequences, this is especially dangerous.
- Data confidentiality. Entering employee census data, personal data, a client's financial information or material covered by professional secrecy into a public tool can breach confidentiality agreements and personal data protection law.
- Bias. Models reflect the biases present in their training data. In applications that touch pricing, risk selection or demographic projections, this can lead to unfair or discriminatory outcomes if left unchecked.
Safeguards and governance
The answer to these risks is not to prohibit but to govern. Some practical safeguards:
- Mandatory human verification. No numerical result, regulatory reference or technical conclusion generated by an LLM should be used without independent checking. The human in the loop principle is non-negotiable in actuarial work.
- Data policies. Define clearly which information may and may not be entered into these tools. Prefer enterprise solutions with contractual guarantees of no retraining on your data and with access controls, and anonymize whenever possible.
- Traceability. Document when an LLM was used, for which task and what a person reviewed. This is consistent with the documentation requirements of actuarial standards.
- Training. The team must understand the model's limitations as well as its capabilities. A user who trusts it blindly is more dangerous than one who does not use the tool at all.
Professional bodies are already addressing the issue. The Society of Actuaries (SOA), the Casualty Actuarial Society (CAS) and the Institute and Faculty of Actuaries (IFoA) have published material on artificial intelligence, data science and their ethical and professional implications for the actuarial profession. Their publications are worth following to keep one's judgment current.
One more tool, not a substitute
LLMs are a real change in the way we work, comparable in potential to the arrival of the spreadsheet. Used well, they free time from repetitive tasks for judgment, communication and analysis, which is where the actuary adds value. Used badly, they can introduce errors that are hard to detect and compromise confidentiality.
The conclusion is a sober one: the LLM drafts, suggests and accelerates, but the signature, the accountability and the professional judgment remain the actuary's. Adopting these tools with rigor, clear policies and systematic verification is the only sensible way to capture their benefits without inheriting their risks.