Commentary on a 2026 request for proposals from the Casualty Actuarial Society's Artificial Intelligence Working Group. Full details are on the CAS website.
The CAS Artificial Intelligence Working Group has issued a call for research examining how large language models can be deliberately adapted to support core actuarial reasoning in property-casualty insurance.
Read that carefully, because the wording is precise in a way that most AI-in-insurance commentary is not. Not extract data. Not automate tasks. Support core actuarial reasoning. And not "can LLMs do this by default" but how they can be deliberately adapted to.
That is a much harder question than it sounds, and I think it is the right one. Let me explain why, and what a serious answer would have to demonstrate.
What "Actuarial Reasoning" Actually Is
The phrase gets used loosely. Pinned down, actuarial reasoning is a specific cluster of cognitive operations, and they are not all the same kind of thing.
Selection under sparse evidence. Choosing a development pattern, a tail factor, a distributional form, when the data does not determine the answer. The actuary is not computing — they are deciding what to compute, on grounds that are defensible but not derivable.
Reconciliation of conflicting indications. Paid development says one thing, incurred says another, frequency-severity says a third. Reasoning is what happens in the gap between them.
Assumption propagation. Understanding that a change in the trend assumption affects the loss ratio, which affects the rate indication, which affects the retention assumption, which affects the mix, which affects the trend. Actuaries hold these dependency chains in mind and reason about them under change.
Recognition of when the method does not apply. The most valuable and least documented skill in the profession. Knowing that a chain ladder is inappropriate for this book, this year, because of a change nobody wrote down.
Communicating uncertainty to people who want a number. Genuinely a reasoning task, not a presentation task: deciding what range is defensible and what caveats are material.
An LLM that produces plausible actuarial-sounding text is doing none of these. It is doing the fifth badly and the other four not at all.
The Three Failure Modes That Matter
Anyone evaluating LLMs for actuarial reasoning will encounter these. They are worth naming in advance.
Fluency mistaken for competence. A language model will produce a well-structured, confident, professionally-phrased justification for a tail factor selection. The prose quality is uncorrelated with the appropriateness of the selection. This is the single greatest risk in the whole area, because actuarial work product is judged partly on its written quality, and a system that writes well will pass review it should not pass.
Arithmetic that looks right. Language models handle numerical reasoning unreliably in ways that are hard to spot, particularly multi-step calculations where an intermediate error propagates. The mitigation is architectural — the model should orchestrate calculations performed by deterministic code, never perform them itself — and any research that does not adopt this pattern is testing the wrong thing.
Absence of the negative case. Actuarial judgement is substantially about not doing things. Not applying a method, not trusting a data segment, not extrapolating. A model trained on published actuarial work has seen overwhelmingly the cases where methods were applied, not the far larger set where an experienced actuary quietly decided against. This is a training data problem with no easy fix, and it is why "the model agreed with the actuary" is weak evidence.
What Deliberate Adaptation Would Actually Involve
The CAS phrasing — deliberately adapted — implies that off-the-shelf capability is not the object of study. Four adaptation routes, in ascending order of promise for this domain:
Prompting and structured elicitation. Cheapest, and more effective than its reputation. Forcing the model to enumerate alternatives, state the evidence for each, and identify what would change its selection produces markedly better output than open-ended questions. It also produces an auditable reasoning trace, which matters more here than the answer quality.
Retrieval grounding. Connecting the model to actual reference material — standards of practice, the firm's own methodology documentation, prior analyses of the same book. This addresses hallucinated methodology, which is the most common visible failure. It does not address judgement quality.
Fine-tuning on domain corpora. Often proposed, and I am sceptical for this application. Fine-tuning on published actuarial literature teaches the model to sound more like the literature. Given that fluency is already the primary confounder, making the output more fluent without making the reasoning better is arguably counterproductive.
Tool orchestration with deterministic components. The most promising route. The language model handles the parts it is genuinely good at — interpreting the question, selecting an approach, explaining the result, flagging inconsistencies — while every numerical operation runs in code that can be inspected and tested. This plays to the actual comparative advantage and eliminates the arithmetic failure mode entirely.
How Such a System Should Be Evaluated
If I were designing the evaluation, I would build it around four tests, and I would weight the last two most heavily.
1. Method selection against expert panel. Present cases where the appropriate method is contested and compare the model's selection and reasoning against a panel. Measure agreement, but measure the quality of the stated rationale separately — a right answer for a wrong reason is a failure.
2. Sensitivity to material information. Change one relevant fact in the case — a mid-year change in claims handling, a shift in policy limits — and test whether the recommendation changes appropriately. A model that gives the same answer regardless of material context is pattern-matching, not reasoning, and this test exposes it immediately.
3. Refusal behaviour. Present cases where the correct professional answer is that the data does not support a conclusion. Does the model say so, or does it produce a confident estimate anyway? This is the single most diagnostic test available, and almost nobody runs it.
4. Robustness to leading framing. Present the same case framed to suggest a preferred answer. Does the model's recommendation move? Actuarial independence has value precisely because it resists commercial pressure. A system that shifts under leading questions cannot support that function.
The Professional Question Underneath
There is a question sitting under the technical one that the profession has not yet answered.
When an actuary signs an opinion informed by a language model's analysis, what does professional obligation require them to have done? Applicable standards require the actuary to understand the methods and models they rely on, to satisfy themselves as to appropriateness, and to document their judgement. None of that changes because a model was involved.
My reading is that the reliance chain must remain intact: the actuary must be able to explain the selection, defend it independently of the model's justification, and identify the circumstances in which it would be wrong. A model that produces a conclusion the actuary cannot independently defend has not supported their reasoning — it has substituted for it, and the signature that follows is not properly supported.
That is not an argument against the research. It is an argument for exactly the framing the CAS has chosen: support actuarial reasoning, deliberately adapted. Support means the actuary reasons better. It does not mean the actuary reasons less.
What I Hope the Research Produces
Not a benchmark score. A taxonomy: which actuarial reasoning tasks these systems genuinely improve, which they leave unchanged, and which they actively degrade by adding fluent justification to unchanged judgement quality.
That third category is the one the profession most needs mapped, and it is the one commercial vendors have the least incentive to investigate.
Key Takeaways
- The CAS AI Working Group's call targets deliberate adaptation of LLMs to support core P&C actuarial reasoning — a far harder problem than data extraction.
- Actuarial reasoning comprises selection under sparse evidence, reconciliation of conflicting indications, assumption propagation, recognising method inapplicability, and communicating uncertainty.
- The dominant risk is fluency mistaken for competence: actuarial work is partly judged on written quality, so a well-writing system passes review it should not.
- Tool orchestration with deterministic calculation components is the most promising adaptation route; fine-tuning on published literature may worsen the fluency confounder.
- Refusal behaviour and robustness to leading framing are the most diagnostic evaluation tests, and are rarely run.
Frequently Asked Questions
Can AI replace actuarial judgement? Current evidence does not support this. Language models can support specific reasoning tasks — surfacing alternatives, checking consistency, drafting explanations — but the core judgement operations of method selection under sparse evidence and recognising when a method does not apply remain human. Professional standards also require the signing actuary to understand and be able to defend the methods relied upon.
What is the main risk of using LLMs in actuarial work? Fluency being mistaken for competence. A language model produces confident, professionally-phrased justification regardless of whether the underlying selection is appropriate, and actuarial work product is partly assessed on written quality. This makes poor judgement harder to detect rather than easier.
How should LLM-assisted actuarial work be documented? The actuary should record the judgement independently of the model's output — the alternatives considered, the basis for selection, and the circumstances under which the selection would be wrong. If the actuary cannot defend the conclusion without reference to the model's justification, the reliance is not adequately supported.
Jonas Osman Abdelghafour is the founder of Quantica Risk, which builds and independently validates actuarial and AI models for insurers. This article is commentary on a publicly available CAS research call.