AI in Actuarial Science: Beyond the Buzzword, a Practitioner-First Approach

The profession is changing. How will you adapt?

For decades, the actuary’s value was inseparable from their ability to build models, spreadsheets, analyses and reports, work that often required days of manual construction.

This constraint is starting to break. Generative AI (Gen AI) is creating a new class of tools that reduce the time needed to deliver analysis from days to minutes. However, these tools are effective only when two key foundations are in place: structured, well-governed data and strong actuarial expertise. Gen AI does not create actuarial knowledge; it scales knowledge that has already been defined, structured and guided by actuaries through data and prompts.

This article describes how Gen AI tools are built in practice in an actuarial context, what it takes to make them work, and why actuaries remain at the centre of the process.

Two accelerators, one methodology

In an actuarial context, two accelerators can significantly improve work efficiency:

• interactive analysis systems (e.g., dashboards and natural language interfaces) that enable real-time exploration of actuarial results without the need to create custom Excel files, and
• reserve engines that scale calculations across an entire book of business.

Interactive analysis systems address result analytics. In current reporting, actuarial results are often confined to modelling platforms and accessible only to actuaries. As a result, requests from finance teams or senior management typically trigger a lengthy manual process, from data extraction to the production of analysis files.

The first accelerator addresses this constraint by structuring actuarial outputs in a queryable database and presenting them through dashboards or automated reports. A language model layer then lets users ask questions and get real-time responses – without creating manual files.

The architecture follows a bottom-up approach. Actuarial expertise defines what data to extract and how to interpret it. Data engineering then designs data flows and structures model outputs and parameters into a database, followed by an analytical layer for aggregation, variance decomposition, automated narratives and visualization. The AI layer is applied last to interpret user requests and generate responses; in effect, acting as an interface built on these foundations.

A reserve engine addresses actuarial calculations. Current modelling processes require significant time and resources to build, run and validate calculations product by product. The second accelerator scales this across the entire portfolio.

It begins with actuarial expertise selecting a representative product and building a calculation spreadsheet, which is then translated into generalized data structures and reusable Python code. Built on this foundation, Gen AI can extend actuarial logic to additional products, and ultimately the full portfolio. It can also generate test cases, produce documentation and help trace calculation issues, further boosting modelling efficiency.

Both accelerators apply Gen AI in different ways but follow a common principle: Gen AI is not a creator, but a multiplier. While it generates outputs based on user prompts, it depends on three other components working together: actuarial expertise to define the logic, data engineering to structure and govern the data, and RAG (retrieval-augmented generation) and prompt design to ensure outputs are relevant in an actuarial context.

RAG, prompt engineering and fine-tuning: Improving Gen AI in the actuarial context

Prompt engineering, fine-tuning and RAG are optimization methods that help generate outputs that are relevant and reliable in a professional actuarial context.

Expert prompting optimizes user inputs by providing clear, effective instructions so the model can generate more precise outputs. Fine-tuning applies supervised learning with labelled data to train large language models (LLMs) on domain-specific knowledge. RAG grounds each response in verified sources (e.g., documentation, calculation code) rather than the model’s generic knowledge, reducing hallucinations. By grounding responses in retrieved source material rather than the model’s parametric memory, RAG improves auditability: each response can be traced to the underlying document or calculation, an important requirement in a regulated actuarial context.

Each method improves output quality, but on its own addresses only part of the problem: prompt engineering shapes how inputs are framed, fine-tuning adapts the model’s behaviour and tone, and RAG grounds responses in verified data. Together, they make LLMs more reliable and cost-effective. Here is an example of user prompts and LLM responses for interactive analysis systems with different optimization methods:

No optimization

Q: Explain why reserves increase from $100 million in Q3 to $110 million in Q4.

A: A $10 million (10%) increase in reserves from Q3 to Q4 is typically driven by changes in underlying assumptions and economic conditions rather than a single factor.

Prompt engineering

Q: You are a reserving actuary. Explain the reserve movement from Q3 to Q4 with the following data. Quantify the impact of each driver and provide a concise explanation.
Reserves at Q3 changed from $100M to $110M, the discount rate changed from 3.0% to 2.6%, liability duration is 10 years, mortality assumption change impact is +$4M

A: The increase in reserves from $100M to $110M (+$10M) can be explained by the following drivers:

1.	 Mortality assumption update (+$4M)
The update to mortality assumptions increased expected future benefit payments, contributing +$4M to reserves.

2.	Discount rate decrease (3.0% → 2.6%) (+$4M)
The reduction in the discount rate increases the present value of future liabilities. Using a standard duration approximation, a 40 bps decrease in discount rates increases the PV of liabilities by $4M given a 10-year duration.

3.	Other impacts (residual: +$2M)
The remaining movement reflects other factors such as experience variances, model updates or assumption refinements not explicitly quantified.

RAG

Q: Explain why reserves increase from $100 million in Q3 to $110 million in Q4.
(Assume RAG retrieves Q4 attribution data including interest rate impact, mortality impact and residual buckets.)

A: Reserves increased by $10M from Q3 to Q4, from $100M to $110M. Based on the Q4 reserve attribution:

1.	Interest rate impact: +$4M
Driven by a decrease in discount rates over the quarter, increasing the present value of future liability cash flows.

2.	Mortality assumption update: +$4M
Reflects updated best-estimate mortality assumptions approved in Q4.

3.	Other impacts: +$2M
Primarily related to experience variance and other model updates during the quarter.

Fine-tuning

Q: Explain why reserves increase from $100 million in Q3 to $110 million in Q4.

A: Reserves increased by $10M in Q4, driven primarily by changes in economic assumptions and experience over the quarter.
The increase is likely attributable to a combination of:
1.	Economic impacts (interest rates): Lower market yields would increase the present value of future cash flows, particularly given the typically long duration of insurance liabilities. This is expected to be a key driver of the increase.

2.	Assumption updates: Revisions to best-estimate assumptions, such as mortality or lapse, may have contributed to a structural increase in liabilities based on emerging experience.

3.	Normal period movements: This includes interest accretion on the opening liability, offset by benefit payments and experience variances during the quarter.

From a practical perspective, RAG is a cost-effective starting point. Because actuaries rely on data to inform business decisions, connecting to current, verified sources with clear traceability provides greater control over outputs in an actuarial context. Prompt engineering complements this by guiding LLMs to generate responses aligned with actuarial intent. When the use cases stabilize and scale, fine-tuning can be introduced to further enhance performance.

The case for prioritizing RAG is also technical. Fine-tuning is the most resource-intensive option: it requires labelled data, significant computing capacity and retraining whenever underlying data changes. RAG strikes a better balance. With access to relevant data, it needs less context and fewer follow-ups to produce high-quality outputs, reducing token usage, avoiding lengthy retraining processes for new data and lowering operational costs at scale.

Structuring data for Gen AI: The role of data engineering

Gen AI is only as reliable as the data it can access. In a RAG architecture, the model retrieves responses from a knowledge base curated by data engineers rather than relying on its own training. As a result, data engineering becomes the most critical, and often underestimated, layer of any Gen AI system.

The core mechanism of RAG relies on two steps: chunking, which breaks knowledge into segments; and vectorization (or embedding), which converts each segment into a high-dimensional numerical vector representing its semantic meaning. Conceptually, texts with similar meaning are placed close together in this vector space.

At query time, the user’s question is embedded in the same space, and the system retrieves the closest matching chunks, typically using cosine similarity. These vectors are stored and searched in a dedicated vector database, enabling efficient retrieval even across millions of segments.

Output quality is therefore bounded by two factors: the quality of the chunks and the quality of the embedding model used to encode them. Poorly structured data or a weak embedding model leads to inaccurate retrieval, which in turn increases the risk of hallucinations, even with a state-of-the-art LLM.

To mitigate this, data engineering must guarantee consistent naming conventions, rich metadata tagging for auditability, chunking strategies aligned with actuarial logic and automated pipelines to keep the knowledge base current.

Data engineering does not simply “prepare” data for Gen AI; it defines the ceiling of what Gen AI can reliably produce.

Actuarial expertise remains indispensable, even in the era of Gen AI

In the traditional model, actuaries build formulas cell by cell, extract data manually and construct reports from scratch. Much of their time is devoted to these mechanical tasks rather than to analysis, interpretation and decision-making.

Gen AI can now scale calculations and analyses rapidly, with minimal manual effort. At first glance, this appears to replicate one of the actuary’s core strengths: building. However, actuarial judgment remains both critical and irreplaceable. In the interactive analysis systems described earlier, actuaries must still validate the accuracy and reasonableness of the AI-generated outputs. In reserve engines, particularly for innovative product features, actuaries determine how risks should be quantified.

The role of actuaries is therefore not eliminated but refocused on its core values: applying mathematics, statistics and data to model the future in ways that help individuals and organizations manage risk.

Using Gen AI to amplify actuarial expertise

Gen AI has gained rapid traction due to its powerful capabilities, but it is not a standalone solution. It depends on a robust supporting architecture, including actuarial expertise, data engineering and model optimization. Without actuarial expertise, there is no sound logic to scale. Without data engineering, there is no reliable data source. Without model optimizations, even advanced LLMs are prone to hallucination.

The actuarial profession is therefore not being replaced by AI, it is being reshaped. In an AI-enabled environment, actuaries should use Gen AI as an amplifier, accelerating processes they have already mastered while focusing on what cannot be generated without human input: expertise and judgment.

Alex Chan, FCIA, FSA, is an Actuarial Manager at EY Canada. He transferred from EY Hong Kong and brings multi-market experience in the life insurance sector, with extensive knowledge of various capital regimes and IFRS 17. His work spans traditional actuarial reporting and modelling, as well as technology initiatives.

Cem Unlubayrak is a Senior Manager in actuarial services at EY Canada, specializing in actuarial modelling and emerging applications of artificial intelligence. He has worked extensively on research and development initiatives, including the design of advanced actuarial models and data-driven tools for insurance applications.

Simon Girard, FCIA, is a Partner in EY’s insurance and actuarial practice and leads the practice in Canada. He serves client across North America and is committed to innovation and advancing the actuarial profession. He sponsors the development of advanced analytics solutions for actuarial and strategic applications for the life and annuity insurance industry. His other focus areas include accounting change (IFRS 17) and actuarial and finance transformation.

This article reflects the opinion of the authors and does not represent an official statement of the CIA.