← Back to blog

Generative AI for Actuaries

Published: Fri May 01 2026 03:00:00 GMT+0300 (Eastern European Summer Time) · Revised: Fri Jun 12 2026 03:00:00 GMT+0300 (Eastern European Summer Time)

Preface

This article is originally written in Finnish and submitted as a part of the Finnish actuarial qualification. I’ve translated it using my Long Document Translation tool, available as an API and on the chat interface. This article contains the translated article formatted to fit the web environment.

Heikki Kettunen · February 2026

Abstract

This article explores the fundamentals, applications, and challenges of Generative AI, particularly from an actuarial practice perspective. Artificial intelligence (AI) is an umbrella term for methods that aim to replicate human intelligence, such as reasoning and learning. Recently, Generative AI and its associated large language models have been in the forefront due to their ability to efficiently produce text-based content. Language models typically use Transformer neural networks to predict an appropriate response based on user input. For knowledge workers, Generative AI offers powerful tools for automating routine tasks and processing information. The technology has applications in areas such as software development, document handling, and data analytics. Successful use of Generative AI requires careful planning and an understanding of the system’s limitations. Its capabilities are directly tied to its training data, and the system itself may not necessarily recognize when it is unsuitable for a given task. Responsible use also demands attention to information security and copyright considerations. At the time of writing, the development and adoption of AI systems in companies is highly active. Actuaries’ strong technical expertise provides them with a good foundation to understand, develop, and apply these systems.

Contents

  1. Introduction
    • 1.1 Terminology
  2. Generative AI System
    • 2.1 Neural Network
      • 2.1.1 Transformer
    • 2.2 System Description
      • 2.2.1 Language Model
      • 2.2.2 User Interface
      • 2.2.3 Backend System
      • 2.2.4 Auxiliary Systems
  3. Applications
    • 3.1 Documentation
    • 3.2 Summaries
    • 3.3 Meetings
    • 3.4 Scenarios
    • 3.5 Information Retrieval
    • 3.6 Verification
    • 3.7 Software Development
    • 3.8 Data Pipelines and Automation
    • 3.9 Data Analytics
    • 3.10 Actuarial Applications
      • 3.10.1 Pricing
      • 3.10.2 Claims Liability Analytics
      • 3.10.3 Regulatory Monitoring
  4. Using Artificial Intelligence
    • 4.1 Operating Environment
    • 4.2 Regulation
      • 4.2.1 EU AI Act
    • 4.3 Risks Associated with AI Use
      • 4.3.1 Information Security
      • 4.3.2 Data Accuracy
      • 4.3.3 Reproducibility
      • 4.3.4 Copyright
  5. Summary
  6. Appendix A: Machine Learning

1 Introduction

Artificial intelligence (AI) refers to the ability of machines to perform tasks that require human-like intelligence, such as learning, reasoning, perception, language skills, or decision-making. AI serves as an umbrella term for a wide range of applications that differ in both their applications and their implementation. Recently, Generative AI has been at the forefront of interest in both research and business.

Generative AI systems produce content, such as text, images, audio, or program code, based on input. These systems leverage so-called language models that calculate the meaning of the prompt provided by the user and guide the system to generate content accordingly. ‘Generative AI for Actuaries’ is a suitably narrow title for a short article. Generative AI can find many applications in insurance companies beyond actuarial practice, such as automated customer service chats. Conversely, the adjustment of parameters for general linear models, which may be familiar from pricing, for example, can be seen as machine learning, i.e., one form of AI.

For actuaries, Generative AI offers versatile tools for data processing and automation. Transferring manual tasks to AI, such as taking notes during a meeting or filling out a report template, frees up time for other tasks. AI can serve as an assistant in computational tasks. Interpreting or translating code, documentation, and synthetic data generation can be rapidly accomplished by Generative AI.

The use of Generative AI also involves limitations and risks. The areas of expertise of a Generative AI system are essentially determined by its training data. Outside these subject areas, useful responses cannot be expected from the system. The features of the system in use should be designed to be suitable for the task, and the system’s operation in its assigned tasks should be monitored and tested. Furthermore, the correct use of the systems, as well as information security and copyright considerations, must be taken into account.

Actuaries’ technical expertise provides them with excellent opportunities to understand AI systems and to be involved in their development if necessary. The article provides a high-level overview of the operating principles of Generative AI systems, examples of potential actuarial applications, and considerations for the proper use of AI. The article cannot be considered comprehensive due to the breadth of the subject and its rapid development, but it aims to provide a brief, timely introduction for actuaries interested in the topic.

In addition to research articles, many introductory overviews have been written about AI, such as the English-language ‘A Primer on Generative AI for Actuaries’ [12], specifically aimed at actuaries. The following articles provide in-depth information on technology related to Generative AI: machine learning of neural networks [1], the architecture and operation of language models [15] [2] [11] [5], and information retrieval in AI systems [9].

1.1 Terminology

This section provides an overview of key terminology related to AI. Anthropomorphic expressions such as “AI understands,” “machine learns,” or “neural network” have become established in common usage. These can create misleading perceptions of the true characteristics of AI systems. On the other hand, these terms illustrate the dynamic way AI systems process information, but it is important to remember that they are programs based on mathematical models and application logic.

Even within this article, terms such as AI, AI system, Generative AI, and language model are sometimes used to describe the same entity from different perspectives. Specifically, ‘AI’ is used as a subject when referring to the operation of any of the aforementioned. However, these terms should not be conflated, as they have their own established meanings. AI is a general term for systems that perform tasks typically associated with human intelligence, such as reasoning, learning, or language comprehension. Generative AI refers to AI that produces new content (text, image, audio, etc.) based on input. A language model, in turn, is a neural network model trained for natural language use, which predicts the meaning or continuation of a character string. In many Generative AI applications, the language model is the means by which a response is generated from the input for the user, or instructions are provided to another part of the system.

Below are key terms related to AI and their meanings:

  • Artificial intelligence (AI): A general term for technologies that aim to perform intelligent functions. AI does not refer to a single method, but to a collection of approaches that mechanise functions such as learning, reasoning, or language processing.
  • AI system: An entity that includes an AI model, potential supporting systems, and application logic. The system’s purpose is to provide an AI-based service or functionality.
  • Generative AI (GenAI): A subfield of AI where models are capable of generating new content based on input and their training data. Examples include the generation of text, images, audio, or code. These often leverage large language models.
  • Language model: A computational model trained to process natural language. Often based on neural networks leveraging transformer operations. Large language models (LLMs) can handle complex linguistic tasks and answer questions, generate text, or perform basic reasoning.
  • AI agent: An AI system that interacts with a separate system. Separate systems may include file systems, the internet, or other agents. Agent interaction may involve information gathering, information input, or both.
  • Neural network: A computational model composed of interconnected neurons. The connections between neurons form a linear model, which is made non-linear using an activation function. The non-linearity and good scalability of the model make it possible to describe complex phenomena.
  • Machine learning (Appendix A): A method in which a computational model is fitted based on example data without explicit programming. Learning occurs by iteratively adjusting the model’s parameters so that the model’s output and selected true values converge.
  • Deep learning: A subfield of machine learning where the trained model is a multi-layered neural network (deep neural network). The potentially vast number of parameters in deep neural networks makes them suitable for describing complex data such as language or images.

2 Generative AI System

AI as a scientific concept originates from the 1950s, when a research group related to the topic was founded at Dartmouth College [3]. The field has since grown and significantly expanded its scope, largely in line with advancements in computational power. AI is projected to have various future prospects, and it is already in diverse use today.

Generative AI produces new content based on the input and the content it has been trained on. The new content is based on the trained content, such that the algorithm calculates what would most likely be the correct response based on the input. The part forming the program’s ‘AI’ is a computational model, so demanding practical applications require an AI system containing traditional application logic around them. Generative AI systems refer to application entities in which Generative AI plays an essential role.

A Generative AI system can be divided into the following parts:

  • User interface: The user interface and interactive functionalities, enabling user interaction with the system.
  • Backend system: Control of the language model and auxiliary systems. Communicates with the user interface.
  • Language model: A neural network that computes the response to the input. Alongside the language model, an input pre-processing module operates.
  • Auxiliary systems: Integration of external data sources (RAG), in-conversation memory, code execution interface, etc. Auxiliary systems form a data processing software entity controlled by the language model. Auxiliary systems can be controlled using control loops built into the backend system and trained for the language model.

2.1 Neural Network

Neural networks are computational models used in language models to calculate the meanings of words. There are several different neural network architectures, suitable for various applications. Architectures are designed according to what kind of data and how that data is processed by the neural network. Transformer architectures have become standard for language models, which we will discuss in more detail at the end of this section.

Neural networks can be considered a generalization of linear models, where non-linearity is introduced using activation functions. Thus, neural networks are capable of significantly richer representations than linear models.

ReLU (Rectified Linear Unit) is a typical activation function used in conjunction with neurons.

Definition 2.1.1 (ReLU). ReLU is a function defined as σ(x) = max(0, x).

A neuron is a fundamental component of a typical neural network. It can be defined as the value of an activation function calculated from the difference between the inner product of an input and a weight vector, and a bias. In practice, b is often 0. The number of parameters in a layer consisting of neurons can be adjusted by changing the number of neurons.

Definition 2.1.2 (Neuron). A neuron is a function f : Rⁿ → R, defined as f(x) = σ(⟨x, w⟩ − b), where inputs x ∈ Rⁿ, weights w ∈ Rⁿ, and bias b ∈ Rⁿ.

A neural network consists of layers composed of affine-linear operations followed by activation functions. Layers constructed from interconnected neurons are referred to as feedforward layers, as defined below. The number of parameters in a fully connected network is nᵢ × nₒ, where nᵢ is the number of inputs and nₒ is the number of neurons in the layer.

Definition 2.1.3 (Feedforward Layer). Let n₀ = d ∈ ℕ be the input dimension, L ≥ 2 be the number of hidden layers, and nₗ, l = 1, …, L be the dimensions of the hidden layers. Let Tₗ : Rⁿˡ⁻¹ → Rⁿˡ be an affine-linear function defined as Tₗ(x) = W⁽ˡ⁾x + b⁽ˡ⁾, where W⁽ˡ⁾ ∈ Rⁿˡˣⁿˡ⁻¹ is the weight matrix and b⁽ˡ⁾ ∈ Rⁿˡ is the bias of layer l.

A neural network is assembled by connecting layers to one another. The size of a neural network can be adjusted by adding layers or parameters to layers.

Definition 2.1.4 (Neural network). Using Definition 2.1.3 for network layers, the function Φ : Rⁿ⁰ → RⁿL, defined as Φ(x) = TL(σ(TL−1(σ(…σ(T₁(x)))…))), x ∈ Rⁿ⁰, is a deep neural network.

2.1.1 Transformer

We take the Transformer as an example of a specialized neural network architecture, since modern language models are almost without exception of this type. Transformers utilize the Attention operation, which is an operation between query, key, and value vectors. The result is a weighted sum of the value vectors, where each value is weighted by the mutual suitability of its corresponding query and key. That is, the better the query and key match, the more attention is directed to the corresponding value.

Query, key, and value vectors are formed using their corresponding inputs and parameters, i.e., weights. Large language models primarily use so-called “Self-Attention”, where query, key, and value vectors are all formed using the same input. This type of operation functions well for language models, as in calculating the meaning of each part of the input, every part of the input is considered. In the context of language models, inputs are generally matrices consisting of multiple vectors, so we generalize by defining operations with matrices.

Definition 2.1.5 (Attention). The Attention operation is defined as Attention(Q, K, V) = softmax(QKᵀ / √dₖ) V.

Definition 2.1.6 (Softmax). Softmax is defined as a multidimensional logistic function softmax(x) = exp(x) / (1ᵀexp(x)).

In a Transformer, a feedforward network is typically used after the Attention operation. We therefore define a complete Transformer layer as follows. Typically, inputs containing multiple vectors are processed as matrices.

Definition 2.1.7 (Transformer Layer). Let the input be X ∈ Rⁿˣᵈ, where n is the number of input vectors and d is the length of the input vectors. The Transformer’s parameters include the weight matrices:

  • W^Q ∈ R^(d×dₖ) (query weights)
  • W^K ∈ R^(d×dₖ) (key weights)
  • W^V ∈ R^(d×dᵥ) (value weights)

Using these, the matrices Q = XW^Q ∈ R^(n×dₖ), K = XW^K ∈ R^(n×dₖ), and V = XW^V ∈ R^(n×dᵥ) are formed. The output of a Transformer layer is σ(T(Attention(Q, K, V))), where T is a feedforward network according to Definition 2.1.3.

There are different versions of the Transformer; for example, encoder, decoder, or encoder-decoder Transformers are discussed. It should be noted that both encoders and decoders inherently utilize the same operations presented above, but encoders are trained to form a representation of the input, while decoders are trained to make predictions based on the input. Actual implementations often also include details that we omit, such as word embeddings, result normalization, residual connections between layers, and causal masking in decoders. For those more interested in the topic, a good source is the 2017 publication [15], in which the Transformer architecture was originally presented and which can be considered one of the catalysts for the broader enthusiasm for AI.

2.2 System Description

The operation of a Generative AI system based on a language model can be broadly divided into three phases: 1) input parsing and context building, 2) response generation using the model, and 3) output post-processing and formatting. These tasks are distributed among the components of the Generative AI system, which we will examine next.

In addition, control loops can be created in the system, where, upon the fulfillment of defined conditions, the response generated by the language model is fed back to the model. These loops can be used to control potential auxiliary systems.

2.2.1 Language Model

A language model is an AI model specialized in natural language processing. These models are almost without exception based on the Transformer architecture described previously. The operation of conversational language models can be illustrated with the following pseudocode. The algorithm returns a word that it predicts to follow the given character string. Responses longer than one word are generated iteratively by feeding the original input augmented with generated words, until the model produces an end token or the defined maximum response length is reached.

Algorithm 1  Language model
  Input:  Prompt
  Output: Response
  Let n ← Max response length
  for i = 1 to n do
      Generation ← Predict(Prompt + Response)
      if Generation = 'EndToken' then
          quit()
      else
          Response ← Response + Generation
      end if
  end for

A language model needs to be trained for its intended task. During the training process, data is fed to the model, and it is rewarded based on the quality of the results it yields. Training data defines the subject areas that the language model comprehends. Training large language models is a substantial undertaking that requires specialized resources. Language models from cloud service providers are pre-trained.

It is possible to fine-tune pre-trained models, i.e., to train them to better understand a specific topic or task. Language model training can be described in different stages:

  1. Pre-training: Acquisition of general language proficiency using a large text corpus. In this stage, the model learns the basic structure and vocabulary of the language.
  2. Fine-tuning: Specialization for a specific application or domain.

Fine-tuning material must include all specialized information and operating instructions that are not included in pre-training but which the AI must comprehend. For example, for an insurance industry system, fine-tuning could be performed with a corpus that includes insurance legislation, calculation models, claims decisions, and customer discussions.

Language models do not use natural language, but process the input as tokens, i.e., units split into parts of text and encoded as numerical vectors. For example, the word “koiralla” may be tokenized into two parts: “koira” and “-lla”, both separately. The goal of tokenization is to separate all concepts and language details that the language model needs to comprehend.

Tokenization first produces integers, which are then converted into corresponding one-to-one vectors. To these word-describing vectors, a vector determining the word’s position is added, which is finally fed to the language model.

Furthermore, pre-processing often includes normalization, such as converting letters to lowercase or standardizing characters. The purpose of normalization is to minimize the required vocabulary, which in turn reduces the number of parameters required in the neural network. The smaller the vocabulary needed to cover the required meaning, the more efficiently the model operates.

2.2.2 User Interface

The function of the user interface is to collect the user’s input, send it to the backend system, and present the model’s response. The user interface and the backend system typically communicate via a network connection.

The most typical user interfaces can be divided into browser-based chat services and copilot-style interfaces. Chats offer text-based and independent interaction with AI. Possible desirable additional features may include, for example, file handling and saving response content. Copilot-style interfaces are embedded in another program, where the AI monitors the content of the document being processed and the user’s inputs, and can provide real-time assistance by analyzing or generating content.

2.2.3 Backend System

The backend system relays the received input to the language model, and further relays the received response back to the user interface. Additionally, the backend system manages the auxiliary systems in use and other program logic of the AI system. A typical implementation method is a server designed to communicate with the user interface.

2.2.4 Auxiliary Systems

A language model alone is capable only of producing responses consistent with its training. Additional information sources or interfaces to other programs can be integrated into the AI backend system, for example. Auxiliary systems are often referred to as tools or functions, and AI systems that utilize them as agents. Agents can interact with other software, services, or agents, and integrate the features they offer into the AI user experience.

Auxiliary systems are separate software components or services that receive instructions or calls from the language model, perform the associated tasks, and return the results for the model’s use. They can act as bridges to external databases, API interfaces, or internal systems, such as document management or computational environments. Consequently, they significantly extend the practical application areas of the language model.

For example, if a user asks the AI for tomorrow’s weather forecast, the system operates as follows:

  • The model’s training or instruction includes information on when a user’s question can be resolved by calling an external function.
  • The language model constructs a formally structured function call, such as getWeather(location="Helsinki", date="2025-06-13").
  • The backend system recognizes the call and executes it by retrieving weather data from an online service.
  • The results are returned to the model, which formats them into the final response for the user.

Response generation based on information retrieval, i.e., Retrieval-Augmented Generation (RAG) in English, is a key example of the utility of auxiliary systems. Although large language models contain a vast amount of general knowledge based on their training data, this information does not update without retraining. This means that they do not automatically know about new events, regulations, or content that were not available at the time of training. An information retrieval system is an important feature for systems that need to stay up to date with extensive and changing information.

Each information source generally requires its own tailored search function or interface model. Possible sources include, for example:

  • Internal organizational guidelines and documentation (e.g., PDF documents)
  • Internal organizational data (e.g., Excel documents, databases)
  • Public databases or online services (e.g., Finlex, THL, Statistics Finland)
  • Real-time online content (e.g., news, weather forecasts)

3 Applications

Generative AI has a wide range of potential applications. Generative AI’s ability to interpret data flexibly and generate content makes it a versatile information processing tool as well as a key component of automated systems. This section presents, with examples, the suitability of Generative AI for the actuary’s work. Generative AI also finds broader applications in the insurance sector; customer service tasks have already been transferred in some companies to the responsibility of AI-driven chat services.

Generative AI can replace humans in routine tasks that do not require extensive understanding but are difficult to automate with other methods. AI also does not fatigue; thus, its accuracy in repetitive tasks that it performs well can be superior to human performance.

In the actuary’s work, a large amount of information is processed. AI is well-suited to generating text based on existing information. Generative AI also assists, for instance, in the creation of illustrative images for presentations. AI models that, in addition to text, understand other data types, such as images or audio, are called multimodal. Using these, it is efficient to convert the content of video or audio files into text.

Actuaries’ work often includes system administration and application development tasks. The expertise required for these tasks does not necessarily fall within every actuary’s core competencies. The general knowledge stored in language models and programming skills provide them with the ability to analyze and interpret code into plain language. This can lower the barrier to participation in tasks requiring IT expertise.

Next, tasks are presented in which Generative AI can begin to impact actuaries’ work. Not all examples necessarily utilize Generative AI in its most obvious form, but all essentially involve a component related to Generative AI, such as a text-interpreting language model.

3.1 Documentation

Documenting existing resources, such as databases, calculation models, or software, is potentially an arduous task where AI can be of significant assistance. Generative AI can quickly generate clear documentation based on specifications, instructions, or software code. Translation from one language to another is often performed better by language models than by traditional translation applications.

AI can also generate drafts of new documents. The quality of drafts is determined by the background information available to the AI and the precision of the instructions provided to it. With an AI-generated draft, it is possible to quickly obtain a foundation with the correct structure and key facts. Text produced by current language models is often identifiable by its vocabulary and sentence structures, and AI does not yet possess the ability to produce such profound conclusions that consider the broad context of matters, which are required for a meaningful whole. In text generation, AI functions best as a tool that assists the user in transforming their thoughts into text more quickly and clearly.

3.2 Summaries

Generative AI can rapidly summarize large amounts of text or answer questions related to text. Comparing two different versions of a specific document is also quickly accomplished with AI.

It is important to consider how well AI is capable of understanding the content of the text input to it. Especially if the AI’s training does not include foundational knowledge on the text’s subject, there is a risk of summaries being superficial. As stated in the preceding section, AI is not a substitute for personal understanding, but rather an efficient text processing tool.

3.3 Meetings

A multimodal model can be used for transcribing meetings. AI-based transcription tools are accurate and capable of distinguishing speakers. Based on the transcription, notes or a summary compiling the most essential points can be generated, which facilitates the drafting of meeting minutes and information sharing.

3.4 Scenarios

A typical actuarial scenario analysis is based on assumptions about potential events and their estimated impacts on a company’s operations and solvency. The impacts of realized scenarios are complex and are difficult to model mathematically without simplifying assumptions.

AI can be trained on historical data and analyses of realized scenarios and their impacts on the subjects under investigation. Thus, AI can generate analyses of corresponding scenarios with the input values provided to it. This can offer an additional perspective alongside mathematical models. It is important to remember, however, that AI does not automatically adapt its analyses to the broader context or prevailing conditions; therefore, the actuary must evaluate the suitability of the analysis produced by the AI.

3.5 Information Retrieval

Large language models contain a vast amount of general knowledge. Many models have assimilated during their training Wikipedia and the documentation of many programming and information systems. Without auxiliary systems, this information is, however, static, meaning it does not update. With the aid of RAG systems mentioned in section 2.2.4, it is possible for AI to collect and apply current information.

A system with suitable RAG capabilities can be queried to retrieve information from a database, or asked to collect the latest news related to a specific topic and summarize it. An AI system can also be an effective tutor for new subjects. For instance, it can be provided with material about which the user has a vague understanding and asked to explain the necessary background information. An AI that has access to internal organizational guidelines can quickly answer questions about practices and procedures.

AI does not necessarily respond with “I do not know” when it lacks information. Instead, it may produce more or less plausible text, which stems from its fundamental nature to predict statistically related words based on the input. Therefore, it is important for the user to evaluate the quality of the received response, and systems should be designed such that they incorporate sufficient information.

3.6 Verification

An AI system performing verifications can be taught the structure and typical content of the entity being verified. In such cases, the system can flag if, for example, a section of a report is missing, is left blank, or contains content deviating from the norm. Standard spell checking is also performed effectively with AI.

For instance, interlinked workbooks are a typical method for actuaries to calculate figures describing an insurance company’s finances and policy portfolio. Generative AI can control a mechanism that warns if defined links are not met.

With AI, it is possible to create background verification systems which, for example, regularly check data arriving in emails and issue a warning if the content does not match the usual pattern or if the information has not arrived on time.

3.7 Software Development

Simple and even slightly more complex programming tasks are often successfully completed by Generative AI without errors. Extensive programming tasks requiring specific background knowledge are challenging for AI systems at the time of writing. Conversely, translating even large ready-made programs from one programming language to another is accomplished relatively well. Also, as in document drafting, Generative AI can assist in drafting software projects by creating a framework for an extensive entity.

Background AI systems like Copilot offer functionalities such as automatic completion and commenting, as well as standard text generation, which can accelerate the work of even experienced coders. However, the greatest benefits are likely to be found in situations where the user requires assistance, for example, with an unfamiliar programming language or undocumented source code.

3.8 Data Pipelines and Automation

AI can collect data from various sources and transform it into a unified, machine-processable format. Consequently, it can function as part of data pipelines and increase the flexibility of automation systems. An example is the classification of freely described insurance events, where, with the aid of AI, events can be classified and converted into a format suitable for a unified data pipeline.

  • Input: “An annual amount of five thousand euros is reserved for a traffic accident from May until the end of 2027.”
  • Response: “Annual amount: 5000, Start date: 1.5.2025, End date: 31.12.2027”

3.9 Data Analytics

AI is capable of processing data significantly faster and on a larger scale than humans. For example, from large datasets, it can be difficult or arduous for a human to discern trends or correlations. AI can be leveraged purely for data analysis, as well as for predictions and analyses developed based on data.

With AI, it is also possible to generate synthetic data or modify existing datasets. Rapidly generated synthetic data enhances, for example, the testing of new computational models.

3.10 Actuarial Applications

The following more specialized applications are mainly based on the general-level features presented in the preceding sections. Legislation concerning insurance operations, as well as the mathematical and detail-oriented nature of the tasks, at least for now, restrict the direct use of Generative AI for actuarial tasks such as pricing and actuarial calculations. However, we will examine some possibilities for utilizing language models in actuarial processes.

3.10.1 Pricing

In the context of customer-specific pricing, large amounts of text-based material can be processed, such as data related to previous insurance events, claim department assessments, and information provided by the insured. With the aid of language models, the aforementioned materials can be classified, analyzed, or structured, thereby enhancing the adjustment of parameters for the existing pricing model.

  • Input: “A customer is seeking an insurance offer for a grocery store in central Helsinki. Within the last year, five shoplifting incidents have been reported. Attached is correspondence regarding the shoplifting incidents and planned security measure developments. Also attached are the store’s location details.”
  • Response: “Based on the shoplifting incidents, the risk profile is higher than average (5 incidents within the last 12 months). The security measure development plan (addition of camera surveillance and staff training) is justified and can be considered risk-reducing for the future. The area has a moderate flood risk.”

3.10.2 Claims Liability Analytics

Based on historical insurance event descriptions and the realized claims or decisions made for them, data can be created which enables AI to analyze insurance event descriptions. Thus, AI can assist in the generation of information used for claims liability assessment and reserving.

  • Input: “Insured diagnosed with a fracture of the upper right humerus and a rotator cuff tear. Recently operated on, rehabilitation progress is being monitored.”
  • Response: “Reasons to prepare for 12 months of fixed-term incapacity for work. As rehabilitation progresses, the degree of incapacity for work will become clearer. In similar cases, full recoveries, permanent impairments, and permanent partial incapacities for work have occurred.”

3.10.3 Regulatory Monitoring

When legislation is updated or an addition is made to it, AI can be useful in comparing the former and current versions of the law, in summarizing changes, and in preliminarily identifying potential areas of impact.

  • Input: “989/2024 Current regulation preceding amendment 989/2024 Note: amendment to Chapter 2, Section 9 b, and Chapter 6, Sections 20 d and 20 e, Chapter 8 a, Section 1, Chapter 23, Sections 9 and 31; addition to Chapter 2, Section 8 a”
  • Response:
    • New Chapter 2, Section 8 a (Legal Entity Identifier): Insurance companies must adopt a legal entity identifier in accordance with the ESAP Regulation (EU 2023/2859).
    • Amendment to Chapter 2, Section 9 b (Notification Obligation): The Financial Supervisory Authority (FIN-FSA) must notify EIOPA of the granting or withdrawal of an operating licence in ESAP format, including the company’s names, identifier, information type, and the inclusion of personal data.
    • Amendments to Chapter 6, Sections 20 d and 20 e (Disclosure Obligations): 20 d: Companies engaging in life insurance business must disclose their corporate governance principles and submit information to FIN-FSA in ESAP format. 20 e: Obligation to publish the investment strategy and arrangements with asset managers and alternative investment funds, and to send information to FIN-FSA in ESAP format.
    • Amendment to Chapter 8 a, Section 1 (Solvency Reporting): An insurance company must annually publish a report on its solvency and financial condition approved by the board, in accordance with the Solvency II Directive and ESAP.
    • Amendments to Chapter 23, Sections 9 and 31 (Liquidation and Bankruptcy Situations): ESAP-formatted notifications regarding liquidation and bankruptcy documents must be submitted to FIN-FSA.
    • Transition period: The obligations for legal entity identifiers and ESAP notifications become binding from 10 January 2030.

4 Using Artificial Intelligence

The effective and responsible use of AI must be considered when designing and using the AI system. The application of language models in connection with a specific subject requires both an understanding of the specific subject within the application’s domain and of the AI’s processing within that subject environment.

When generating Finnish content, it should be noted that the primary language of most large language models is English or Chinese. A top-tier language model primarily developed for the Finnish language does not exist. Many large language models are, however, capable of producing Finnish text at a level roughly comparable to their primary language. In information retrieval and tasks requiring complex instructions, the desired outcome is most reliably achieved using the language model’s primary language.

Language models do not understand the content they generate as humans do, but rather predict the most suitable next part of a text sequence based on the input provided. However, it is difficult for a system based on a language model to identify situations where its information is insufficient. Even when the correct meaning is absent, the language model attempts to produce a response that best fits the given input based on the data it possesses. The ultimate responsibility for the appropriateness of the responses generated by AI rests with the user.

A Generative AI system must be selected according to its intended application. Typical usage should be tested, and the relevance of AI responses evaluated. For some applications, it may be necessary to fine-tune the AI system to suit them.

4.1 Operating Environment

Factors influencing the choice of operating environment include the size of the AI model used, the location of data sources, user numbers, and information security requirements. The operating environment becomes particularly important when the system is utilized in conjunction with an organization’s internal materials, such as documents, reports, or databases.

Cloud services offer user-friendly and quickly deployable solutions that scale as needed. They are particularly suitable for situations where proprietary data sources are not already in place or when there is a desire to quickly test an AI system. However, the use of cloud services involves information security and transparency issues: the user must ascertain how the service provider processes and stores data, especially if confidential or sensitive information is concerned.

On-premise systems are customizable according to usage needs, and they do not increase information security risks. They allow for the direct integration of the AI system with internal data sources without data leaving the organization’s control. Large language models require significant computational power, thus, deployment may necessitate investments in IT hardware. Locally installed systems can function as an internal language model-based search engine, a knowledge worker assistant, or as part of an automation system.

When choosing an operating environment, it is necessary to weigh the system’s technical requirements, use cases, the level of information security, and costs. Cloud services are often suitable for rapid deployment and scaling, whereas on-premise solutions offer enhanced information security and integration capabilities outside the cloud.

4.2 Regulation

The use of AI is subject to regulation and recommendations. Regulation aims to restrict the use of AI to ethical applications. In Finland, national and EU laws apply, as well as international agreements. AI-related regulation can be expected to become more precise as the use of AI becomes more widespread.

Finland’s national legislation currently primarily addresses the use of AI in public administration [10]. As an EU member, Finland has signed the Council of Europe’s AI convention, which obliges it to ensure that AI systems used do not act contrary to human rights, democracy, or the rule of law.

Perhaps the most typical way for actuaries to utilize Generative AI, i.e., as a personal knowledge worker tool, is not regulated, as long as it is not considered to influence decision-making. However, some AI applications potentially used in insurance operations, such as the use of AI in pricing, as part of insurance decisions, or in customer service, are regulated. Future actuarial practice may also include monitoring and evaluating the regulatory compliance of AI systems.

4.2.1 EU AI Act

At the time of writing, the most comprehensive regulatory framework governing the use of AI is the Regulation published by the European Union [14]. This refers to Regulation (EU) 2024/1689 on artificial intelligence, also known as the AI Act, which was approved in June 2024.

The Regulation classifies AI systems into four different risk levels according to their intended use. The classification affects the requirements placed on AI systems. Organizations operating in Europe and utilizing AI must be aware of the Regulation’s requirements and ensure that their AI systems comply with the Regulation.

  • Prohibited AI systems: These include systems that enable social scoring (e.g., evaluation and classification of citizens’ behavior) or generate manipulative content, for instance, for the purpose of disseminating misinformation or influencing behavior. Real-time biometric identification in public spaces is also prohibited under certain circumstances.
  • High-risk AI systems: These include systems that significantly impact individuals’ rights or safety. For example, the use of AI in recruitment, creditworthiness assessment, health status analysis, or insurance decision-making may fall into this category. Such systems require risk management, transparency, information security, and human oversight.
  • Limited risk AI systems: These include systems that interact with the user, such as chatbots or Generative AI applications. In these cases, the user must be clearly informed that they are interacting with an AI.
  • Minimal risk AI systems: For example, AI-powered content recommenders or video games incorporating AI fall into this category. No specific obligations are imposed on these.

4.3 Risks Associated with AI Use

The use of AI as a data processing tool, as presented in this article, does not fall under particularly risky AI applications. However, ensuring well-justified decision-making, content quality, information security, and adherence to copyright considerations requires attention.

The following three points summarize the most important aspects for avoiding risks.

  • The use of AI should not remove important decisions from the purview of expert judgment.
  • The user must be aware of the impacts of AI use.
  • The user must understand the content generated by the AI and how it was formed.

4.3.1 Information Security

The choice of an AI system’s operating environment fundamentally affects the system’s information security. A local system that operates solely on an internal network does not introduce additional information security risks. Conversely, the inputs and responses of an AI system operating in the cloud traverse the internet, exposing the data to all associated threats [8]. Storing data on an external server exposes it to data breaches targeting that server. Regardless of service provider assurances, it is difficult for a cloud service user to verify how their data is used. Actuaries process sensitive data, such as personal information, which the use of AI should not jeopardize.

4.3.2 Data Accuracy

AI-generated responses depend on its training data, as well as any additional information provided to it. It is not easy for AI to assess the accuracy of its information outside this context. Insufficient information on a particular subject by AI may lead to so-called hallucination. Hallucination stems from the nature of the language model to predict words related to a topic, thus, even for an unfamiliar domain, it can develop some kind of response through word associations. For this reason, it is important that the subject area is sufficiently familiar to the user so that they can identify errors made by the AI.

To prevent issues related to data accuracy, when deploying AI, it is necessary to plan which subject areas it will be utilized for and ensure sufficient training in this area.

4.3.3 Reproducibility

When using a computer program, the same input should always yield the same response. Due to the complexity of language models and their natural language-based output, this requirement may, in some situations, need to be relaxed, such that responses do not need to be word-for-word identical, but essential facts should remain consistent with the same input every time. As part of software logic utilizing an automation system, AI must produce its responses in a predetermined format.

Model updates, processes containing randomness, or changing interfaces can affect results. Therefore, system version control, configuration documentation, and thorough testing are crucial in ensuring reproducibility. It is difficult for a user of an unfamiliar system to know what factors their received response is based on.

4.3.4 Copyright

The training of large language models requires an enormous amount of textual data. Much data is freely available, but on the other hand, many sources of interest to language model designers, such as fiction and scientific articles, are often protected by copyright. In the United States, numerous copyright disputes related to AI training have recently been debated [4]. For example, OpenAI and Meta have faced accusations of copyright infringement [13] [6].

In addition to service providers, users must also consider copyright considerations related to Generative AI. A language model may use quotations in its responses, in which case their accuracy must be verified. Content generated by AI used as a source must be cited. The citation should ideally include the name and version of the AI program used, the owner of the application, the generation date, and information on how the AI was used, i.e., in the case of a language model, the prompt given to it or its summary.

5 Summary

Generative AI offers new opportunities for data processing, content production, and automation. For actuaries, Generative AI systems can provide support in routine tasks, programming, documentation, and information system development. The role of actuaries can also grow in assessing, measuring, and regulating the impacts of AI systems.

The use of AI requires critical thinking and careful implementation. AI does not understand the content it produces humanly but mathematically predicts words and structures within the context it has been taught. Information security and conscious use are key factors to consider, especially in a regulated industry such as insurance.

At the time of writing, the adoption and understanding of Generative AI can be characterized as being in transition. Since ChatGPT’s first public version on 30.11.2022, language models have developed, and their use has become significantly more widespread. AI companies and research institutions are making significant investments in research and development and in the infrastructure required for AI services. Knowledge workers at both company and individual levels are experimenting with and developing ways to utilize AI systems in their work. The commercial generalization of AI into insurance operations and business more broadly is currently underway.

Actuaries have a good starting point to support the adoption of AI. This article provides a general overview of Generative AI systems and their potential applications in actuarial practice.

References

  1. Michael M. Bronstein, Joan Bruna, Taco Cohen, and Petar Veličković. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv preprint arXiv:2104.13478, 2021. Proto-book version; comprehensive survey on geometric deep learning.
  2. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877–1901, 2020.
  3. Dartmouth. Dartmouth artificial intelligence conference, 2024. Accessed: 2025-03-31.
  4. A. Feder Cooper and James Grimmelmann. The files are in the computer: Copyright, memorization, and generative AI. arXiv preprint arXiv:2404.12590, 2024. Provides a precise technical definition of memorization in LLMs and its legal implications.
  5. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, et al. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. First-generation reasoning models trained via pure reinforcement learning (RL), open-sourced.
  6. Richard Kadrey, Sarah Silverman, Ta-Nehisi Coates, et al. Richard Kadrey, Sarah Silverman, Ta-Nehisi Coates, et al. v. Meta Platforms, Inc. Complaint filed in U.S. District Court, Northern District of California, Case No. 3:23-cv-03417-VC, 2023. Alleges copyright infringement for training LLaMA using pirated books (Books3, LibGen); case allowed to proceed, later partially dismissed on fair use grounds.
  7. Heikki Kettunen. Neural networks and deep learning systems as parametric spans, 2024. Accessed: 2025-07-15.
  8. Issa M. Khalil, Abdallah Khreishah, and Muhammad Azeem. Cloud computing security: A survey. Computers, 3(1):1–35, 2014.
  9. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems, volume 33, pages 9459–9474, 2020. Accepted at NeurIPS 2020.
  10. Oikeusministeriö. 487/2023, 2023. Accessed: 2025-07-15.
  11. Mary Phuong and Marcus Hutter. Formal algorithms for transformers. Technical Report arXiv:2207.09238, DeepMind / CoRR, 2022. Mathematically precise pseudocode for transformer architectures.
  12. Stephan Mathys and Stephen Carlin. A primer on generative AI for actuaries, 2024. Accessed: 2025-07-15.
  13. Paul Tremblay and Mona Awad. Paul Tremblay and Mona Awad v. OpenAI, Inc. Complaint filed in U.S. District Court, Northern District of California, Case No. 3:23-cv-03223, 2023. Alleges OpenAI ingested plaintiffs’ copyrighted novels in training ChatGPT; court granted in part / denied in part on motions to dismiss.
  14. European Union. Regulation (EU) 2024/1689 of the European Parliament and of the Council, 2024. Accessed: 2025-03-31.
  15. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), volume 30, 2017.

Appendix A: Machine Learning

In Appendix A, we define concepts necessary for the machine learning of neural networks [7]. Familiarization with these concepts can be useful not only for developing applications but also for illustrating the basis on which machine learning programs, such as language models, produce their responses.

First, we define learning.

Definition 6.0.1 (Learning). Let X, Y, Z be measurable spaces. Let L : M(X, Y) × Z → R be a loss function. The goal of learning is to select a hypothesis set F ⊂ M(X, Y) and form a training algorithm, which is a mapping A : ∪ₘ∈ℕ Zᵐ → F using training data s = (z⁽ⁱ⁾)ᵐᵢ₌₁ ∈ Zᵐ to find a model fₛ = A(s) ∈ F whose performance is good with both training data s ∈ Zᵐ and unknown data z ∈ Zᵐ. Performance is measured by the loss function L(fₛ, z).

The task of systems utilizing machine learning is to predict the correct outcome based on the input given to them.

Definition 6.0.2 (Prediction). A prediction Z = X × Y, using training data s = ((xᵢ, yᵢ))ᵐᵢ₌₁, input x⁽ⁱ⁾ ∈ X, and true values y⁽ⁱ⁾ ∈ Y. The goal of model fₛ : X → Y is to produce a good prediction fₛ(x) of the true value y for an unknown (x, y) ∈ X × Y.

The loss function L : M(X, Y) × Z → R provides a numerical value for how well a machine learning model performs its task. For example, in a classification task, this is reflected in how often the model correctly classifies its input. In the context of conversational language models, the training objective is more complex, depending not only on factual information and grammar but also on trainer preferences. Therefore, language models are typically trained in stages: first, a vast amount of general knowledge, and then, for example, a desired conversational style. Typical examples of loss functions include mean-squared error and cross-entropy.

Definition 6.0.3 (Mean-squared error). Let y be the true observations, x the inputs, and f(x) the predictions. Then the mean-squared error is defined by the formula MSE = (1/n) Σⁿᵢ₌₁ (yᵢ − f(xᵢ))².

Definition 6.0.4 (Cross-entropy). Cross-entropy for discrete distributions q with respect to distribution p on the definition set χ is defined by the formula H(p, q) = −Σₓ∈χ p(x) log q(x).

Machine learning is realized as the iterative, controlled adjustment of parameters. In addition to the loss function value, a method is needed to calculate the parameters’ effect on the obtained loss function value. A familiar example is the gradient method.

Definition 6.0.5 (Gradient method). xₙ₊₁ = xₙ − αₙ∇F(xₙ), n ≥ 0.

Algorithmic description of the stochastic gradient method typically used with neural networks.

Definition 6.0.6 (Stochastic gradient method). Algorithmic definition:

  • Input: Differentiable function r : Rᵖ → R, sequence ηₖ ∈ (0, ∞), k ∈ [K], Rᵖ-valued random variable Θ⁽⁰⁾.
  • Output: Sequence of random variables (Θ⁽ᵏ⁾)ᴷₖ₌₁ ∈ Rᵖ.
  • for k = 1, …, K do: let D⁽ᵏ⁾ be a random variable such that E(D⁽ᵏ⁾|Θ⁽ᵏ⁻¹⁾) = ∇r(Θ⁽ᵏ⁻¹⁾); set Θ⁽ᵏ⁾ = Θ⁽ᵏ⁻¹⁾ − ηₖD⁽ᵏ⁾; end.
  • To minimize the empirical loss, r : Rᵖ → R is obtained in the form r(θ) = Rₛ(Φₐ(·, θ)).
  • Iteration can be performed in batches m ∈ ℕ, m′ ≤ m: θ⁽ᵏ⁾ = θ⁽ᵏ⁻¹⁾ − ηₖ (1/m′) Σ_{z∈S′} ∇_θ L(Φₐ(·, θ⁽ᵏ⁻¹⁾), z), where S′ is a mini-batch of size m′.

Neural network gradients are calculated using the backpropagation algorithm. The method proceeds as if in reverse, starting by calculating the gradients of the neural network’s prediction, and then the gradients of the preceding layer by utilizing the previous gradients, and so forth.

Definition 6.0.7 (Backpropagation algorithm). Define for each intermediate variable the adjoint v̄ᵢ = δL/δvᵢ. In a neural network represented by a directed graph, the input can affect the output only through connections between intermediate variables. Then, for all variables vᵢ, i ≤ l, the derivative is v̄ᵢ = Σₖ∈E v̄ₖ δvₖ/δvᵢ, where E is the set of connections leading to the next layer.