Actuaries were doing "artificial intelligence" for centuries before it had that name: modeling risk from data. Here we explore the statistical foundations, machine learning, neural networks and model governance that are redefining the profession today, through the lens of the SOA and the CAS.
Modern artificial intelligence is, at bottom, statistics at scale: learning patterns from data to predict what has not yet happened. That has always been the actuary's job. When Jacob Bernoulli proved the law of large numbers, or when Hans Bühlmann formalized credibility, they laid down principles we now recognize in machine learning: regularization, Bayesian shrinkage and the bias-variance tradeoff.
What changed over the last decade is not the concepts but scale, computing power and flexibility. Where the actuary once used a generalized linear model (GLM) with carefully chosen variables, today there are random forests, gradient boosting and neural networks able to uncover nonlinear interactions across millions of policies. The question is no longer whether the actuary uses AI but how to use it with rigor, transparency and accountability: precisely the ground where actuarial training sets the actuary apart from a pure technologist.
All AI applied to risk rests on the same backbone the actuary studies. These are the building blocks.
Random variables, distributions, maximum likelihood estimation and Bayesian inference: the language used to quantify uncertainty and to train any model.
The generalized linear model (Poisson, Gamma, logistic) is the natural bridge to machine learning: it predicts frequency, severity and probability through a link function, with full interpretability.
Bühlmann credibility is essentially shrinkage: blending the individual signal with the collective one. It is the same idea as the Ridge and Lasso penalties that regularize predictive models today.
Overfitting is the great enemy. Cross-validation, train/test splits and the bias-variance tradeoff are the criteria that separate a model that memorizes from one that generalizes.
Counting processes, Markov chains and stochastic models underpin both reserving and the sequential models of deep learning.
VaR, TVaR and loss functions define the objective a model optimizes. In AI, the loss function is the direct translation of this actuarial concept.
Algorithms that learn from data without being explicitly programmed. In insurance they are already the standard in modern pricing.
Regression and classification from labeled data.
The most widely used family in insurance thanks to its performance on tabular data.
Discovering structure without prior labels.
The Casualty Actuarial Society (CAS) has formalized this shift in its exam syllabus: the monograph "A Machine-Learning Approach to Parameter Estimation" and the e-forum "Machine Learning in Insurance" show how these algorithms are integrated into property and casualty practice with the same rigor as always.
Brain-inspired models that learn hierarchical representations. Their strength: capturing highly complex relationships and unstructured data.
A neuron is a logistic regression: a weighted sum plus an activation function. Stacking layers yields a multilayer perceptron capable of approximating almost any function.
Backpropagation and gradient descent adjust the weights by minimizing a loss function. Activations such as ReLU and the sigmoid introduce the nonlinearity that gives the model its power.
Telematics and images in auto insurance, natural language processing of claims, embeddings of categorical variables and neural networks for projecting mortality (neural extensions of the Lee-Carter model).
A single neuron is already a model: it weights the input signals, adds them up and passes them through an activation function. Adjust the weights and watch how the probability of a claim changes.
There is no break, only a continuum of flexibility. The CAS monograph "From GLMs to Comprehensive Insurance Pricing" follows exactly this path.
Transparent, accepted by regulators, interactions specified by hand.
Ridge/Lasso select variables and control overfitting.
Capture interactions and nonlinearities automatically.
Learn representations and handle unstructured data; maximum flexibility.
The actuary chooses the point on the continuum according to the problem: the more regulated and explainable the model must be, the closer to the GLM; the more complex the signal, the closer to the network. Hybrid models such as the Combined Actuarial Neural Network (CANN) combine a GLM with a network that corrects its residuals, preserving interpretability while gaining accuracy.
This is where the actuary becomes indispensable: predicting well is not enough; you have to be able to explain, justify and answer for the model.
Techniques such as SHAP, LIME and partial dependence plots open the "black box" and translate a complex model into auditable decisions.
Monitoring biases and sensitive variables (proxies) to avoid unfair prices. The CAS has published a series on fairness and discrimination in insurance models.
Independent validation, documentation, monitoring and control of the model life cycle, aligned with actuarial standards of practice (ASOP, ISAP) and emerging AI regulation.
The two major actuarial societies publish continuously on artificial intelligence and data science. These are the key references for keeping up to date.
The section and community devoted to predictive analytics, AI and the future of the profession, with its newsletter, articles and podcasts.
A newsletter with articles on AI and large language models (LLMs), agentic AI, cloud data platforms and the transformation of actuarial work.
The SOA research hub on artificial intelligence, with studies, guides and resources applied to actuarial work.
The archive of the Predictive Analytics & Futurism newsletter: machine learning, neural networks and predictive modeling, issue by issue.
The journey from the GLM to machine learning in property and casualty pricing (Chalk, Deacon, Guillén, Martinelli, 2026).
Machine learning applied to parameter estimation in insurance (Kunce & Chatterjee).
A collection of CAS papers on machine learning applied to property and casualty practice.
The CAS magazine covers AI and analytics; iCAS awards the CSPA (Certified Specialist in Predictive Analytics) credential.
Links to the official SOA (soa.org) and CAS (casact.org) portals. Some resources are members only. Verified in July 2026; sections and newsletters may be reorganized over time.
AI amplifies the actuary who masters statistics, understands the business and answers for their models. Explore the specialties where this combination creates the most value.