AI and Emerging Technologies Symposium

Alex Waite is the IFoA Council representative on the AI & Emerging Technologies new Practice Board and the IFoA President-Elect. He will explain the IFoA’s approach to AI and introduce a new manifesto for how actuaries can engage with the topic.

In this panel session, a leadership team of the IFoA AI & Emerging Technologies practice board will host an interactive exchange about the current challenges and future perspective of the actuarial profession.

Speakers: TBC

Behavioural Signatures Under Uncertainty: Profiling, Calibrating, and Governing the Risk Disposition of Large Language Models

Large language models are increasingly making decisions that matter, yet nobody audits their risk appetite before letting them loose. Recent research applying behavioural economics paradigms such as Holt–Laury risk elicitation, the Ellsberg paradox, and the St. Petersburg game reveals that LLMs carry systematic, measurable biases. Older architectures exhibit extreme risk aversion and near-total ambiguity avoidance, while newer generations trend toward risk neutrality. Crucially, risk preferences, ambiguity attitudes, and paradoxical reasoning are separable dimensions, meaning single-axis evaluations are dangerously incomplete.

This talk presents a framework for profiling, benchmarking, and governing AI risk behaviour. We show how decision-theoretic tools such as CRRA/CARA utility estimation, epsilon-contamination models, and discount factor analysis can quantify an LLM's behavioural signature before deployment. We examine where these profiles diverge from human baselines and what the gaps mean for real-world outcomes. We then explore calibration techniques, from persona interventions that dramatically shift risk posture to the danger of behavioural overshoot where modifications produce extreme rather than calibrated adjustments. Looking ahead, we outline pathways toward continuously tunable risk-reward spectrums, enabling smooth transitions from conservative to aggressive agents within multi-agent teams, where heterogeneous risk dispositions can be orchestrated by design rather than discovered by accident.

Speaker: Arman Khaledian, Zanista AI Ltd

Responsible AI in practice: how actuaries can take the lead

Artificial intelligence presents one of the most significant opportunities the actuarial profession has faced in decades. As AI reshapes decision-making across insurance, pensions, finance, and risk management, actuaries are uniquely positioned to bring rigour, trust and accountability to its development and use. Yet as adoption accelerates, so too do challenges around governance, regulation and professionalism.

In this presentation the Professionalism, Regulation and Ethics Working Party for actuaries working in data science and AI, will demonstrate how existing actuarial strengths and capabilities map directly onto key AI challenges such as explainability, bias, and regulatory compliance, and how actuaries can position themselves to make a meaningful practical contribution based on their core skillset. We'll also explore how actuarial skillsets can evolve to ensure we stay ahead of the curve.

Attendees will take away practical tools and frameworks, including tools for AI model validation, approaches to assessing bias and fairness, and guidance on applying existing actuarial standards in AI contexts. The presentation will also include case studies illustrating actuarial contributions across the AI lifecycle.

Drawing on cross-professional learning, we'll set out how the actuarial profession can strengthen AI oversight and governance—positioning actuaries as key leaders of responsible, well-regulated AI.

Speakers: Geraldine Finucane and Neptune Jin, Professionalism, Regulation and Ethics Working Party and Fei Huang, University of New South Wales

The VANGUARD framework – Value‑Aligned, Governed AI Rollout & Delivery

The challenge in moving from AI experimentation to embedded infrastructure is not model capability, but adoption in a way that delivers value, preserves professional judgement, and stands up to regulatory and ethical scrutiny.

The VANGUARD framework is a structured approach to AI adoption that treats AI an operating model and governance problem rather than a tool collection. VANGUARD is designed to help actuarial and wider insurance teams deploy AI where it genuinely adds value, sequence adoption over sensible timescales, and implement guardrails that enable innovation rather than constrain it.

Drawing on real world experience across many industries, including insurance, the framework brings together core elements illustrated with case studies:

• Value alignment: selecting AI use cases based on strategic importance and decision criticality.
• Process fit: distinguishing where AI should act as:

• • assistant 
  • reviewer or 
  • autonomous component

• Governance by design: embedding accountability, explainability, and human oversight throughout the AI lifecycle.
• Preparing staff to work as part of a human-machine team: including training and skills, motivation and mental model alignment; building a psychologically safe and accountable culture that learns and optimises. 
• Strategy and managed rollout: scaling from pilots to production with top-down leadership, and without creating hidden risks or dependency on opaque systems.

The session reframes AI governance and organisational culture as a facilitator of professional standards, not merely a compliance exercise.

Attendees will leave with: 

• a practical mental model for governing AI adoption, 
• language to engage boards and regulators,
• a clearer view of how actuarial judgement evolves, but remains essential

Speakers: Visesh Gosrani, Cydelta Ltd and Dr Matthew Foxwell, Social Machines

Beyond GLMs: A Reproducible Benchmark of Modern ML and Causal Methods for Motor Pricing on Open Data

Generalised linear models remain the regulatory and operational backbone of motor pricing, yet the gap between GLMs and modern machine-learning techniques is now well-documented in research papers  - but rarely benchmarked end-to-end on data that the audience can actually rerun. This session closes that gap.

We present a fully reproducible comparison of pricing approaches across three open, jurisdictionally distinct portfolios: the French freMTPL2 dataset (≈678,000 policies), the Swedish Wasa motorcycle portfolio from Ohlsson & Johansson (64,548 policy-years), and the Australian De Jong & Heller portfolio (67,856 policies). All three are distributed under open licences via the CASdatasets package, ensuring every result can be replicated by attendees the next morning.

For each portfolio we benchmark classical Tweedie and frequency-severity GLMs against gradient-boosted trees, GLM-boosted hybrids, and neural frequency-severity architectures, and we extend the comparison into causal territory with double-machine-learning estimates of price elasticity. Models are evaluated on out-of-time deviance, calibration, monotonicity with respect to known risk drivers, and stability under simulated covariate shift, the criteria actuaries are actually asked to defend in front of model risk committees.

We share the practical conclusions: where ML improves predictive lift, where it breaks calibration, and where causal methods change the commercial answer. Attendees will leave with a public GitHub repository and a candid view of the residual gaps between research benchmarks and production pricing.

Speaker: Luba Orlovsky, Earnix

Insurtech in Chronic Disease Management – Not One-Size-Fits-All

This session draws on emerging insights from the Insurtech in Chronic Disease Management Working Party, which explores how digital health and Insurtech are reshaping actuarial modelling, decision-making, and product design. A key finding from our work to date is that Insurtech solutions are not one-size-fits-all. Their effectiveness depends heavily on the local healthcare, regulatory, and cultural ecosystem.

This session looks beyond whether Insurtech for chronic disease management works. It explores when digital health data is useful for supporting the ongoing management of chronic conditions, and when might it be robust enough to inform underwriting, pricing or claims. Using hypertension and diabetes as case studies, the session will compare how Insurtech is being used across different ecosystems (US, China and UK) and assess the factors that still limit its wider value. It will also explore whether Insurtech can do more than improve insurer's efficiency and decision-making, and instead have the potential to narrow parts of the protection gap through earlier intervention, wider access to care, and more inclusive protection for people living with chronic conditions.

Key takeaway: Insurtech is not one-size-fits-all: its value depends on the local healthcare, regulatory and cultural ecosystem, and the real challenge is knowing when digital health data can support the ongoing management of chronic conditions, when it may be robust enough to inform underwriting and pricing decisions, and how it might also help narrow parts of the protection gap.

Speaker: Cillian Tierney, Partner Re

Self-Regularising Pricing Models: Smarter GLM Fitting Without the Tuning Burden

Generalised linear models remain the dominant tool for P&C insurance pricing, yet their standard implementation via iteratively reweighted least squares (IRLS) leaves systematic room for improvement: the weighted least squares step at the heart of each iteration is unbiased but not optimal, and penalised alternatives such as Ridge regression and Elastic Net while widely adopted in data science workflows require computationally intensive cross-validation for tuning.

This paper introduces a family of Stein-type shrinkage estimators as direct replacements for the weighted least squares step within IRLS. From a practical data science perspective, the proposed approach enables a new class of self-regularising actuarial models—models that automatically control their level of complexity without the need for repeated tuning or trial-and-error. This removes a key bottleneck in current workflows, reducing computational cost and making pricing models faster to build, update, and deploy in practice.

Speakers: Vali Asimit, Bayes Business School and Claudio Senatore, SAS Institute

AI governance was built for models that predict and recommend, with humans in the loop. Agentic AI breaks that model: systems that plan, use tools, and act autonomously introduce risks that point-in-time assessment can't catch. This talk argues for a shift from auditing models to overseeing behaviour, continuous monitoring, action-level guardrails, and accountability for autonomous decisions and outlines what a governance stack fit for agentic AI looks like in practice.

Speaker: Dr. Adriano Soares Koshiyama, Co-founder & CEO, Holistic AI; Honorary Research Fellow, UCL Computer Science

As part of the symposium, we’re excited to introduce the Artificial Intelligence and Actuarial Innovation 2026 (AI² 2026) competition.

This challenge invites individuals and teams to turn real-world actuarial and business problems into practical, innovative AI-driven solutions.

More details coming soon.

Future of Life Actuarial Modelling with AI

The session focuses on innovation in life actuarial modelling, particularly how AI could reshape the way models are developed in the future. The talk would explore both conceptual changes and practical applications of AI in actuarial workflows.

Speakers: Ka Hei Cho, Aon and Marc Fakkel

Rethinking model risk management for an AI-driven future

As life insurers step up their use of artificial intelligence (AI), both the opportunities and challenges for Model Risk Management (MRM) are evolving rapidly. Firms must be prepared to adapt their MRM approach. Firms that fail to keep pace with this changing environment risk greater exposure and may find themselves on the back foot when scrutiny arises – whether from regulators, internal stakeholders or when issues materialise. In this session, we will explore how MRM frameworks need to evolve in response to complex AI models, highlighting the distinct risks AI can introduce and explore in the context of specific common AI use cases in insurance, how model review techniques need to be adapted to mitigate these risks. We will also explore how AI itself can be leveraged to enhance analytical capability and improve efficiency in the delivery of MRM activities.

Speakers: Harvard Lee and Stephanie Leung, KPMG

When the AI Gets It Wrong, Who Answers for It? Consumer Duty and the Accountability Gap in Algorithmic Decision Making

Consumer Duty created the most significant shift in accountability standards for UK life and pensions firms in a generation. The profession welcomed it. And then it carried on deploying AI systems that create a structural accountability gap Duty was specifically designed to close.
 
When an AI system influences a decision that subsequently harms a consumer, the current governance frameworks in most firms don’t know who is accountable.

The gap is not hypothetical. AI assisted tools are already influencing drawdown recommendations, bulk annuity pricing, claims assessments, and retention modelling across the life and pensions market. In most cases the governance documentation describes how the model was validated. It does not describe who made the decision, what weight the model output carried in that decision, or what the actuary's professional responsibility is when the model produces a result they find plausible but have not independently verified.

This session maps the accountability gap precisely. It identifies three failure modes in current Consumer Duty AI governance frameworks, proposes a practical accountability architecture that firms can implement without waiting for regulatory guidance, and puts three direct questions to the audience about whether their current documentation would survive an FCA review in the event of a Consumer Duty breach linked to an AI assisted decision.

The session draws on transformation delivery experience inside regulated life and pensions environments, the governance frameworks and the emerging regulatory conversation around the EU AI Act and its interaction with UK Consumer Duty obligations.

Speaker: Gayle McDonald-Allan, SS&C

From zero to agent: a live build of an automated underwriting agentic AI model

What are AI agents, how do they work, and most importantly, how can actuaries build them? 

This hands-on session will answer those questions through a live demonstration: building an automated agentic AI underwriting system from scratch. 

The presenters will show how an agentic AI system can be designed and implemented using open Python libraries.  Focusing on practical workflow design and decision-making, this system will have real world relevance across the actuarial profession.  It will also make use of techniques that could be applied on any number of other actuarial use cases.

The presenters will both describe and share the system's code repository for all to see, and they will demonstrate, using interactive and stimulating visuals, how it operates. 

The session is intended for actuaries, actuarial students, AI enthusiasts and business leaders alike. No prior knowledge of AI will be assumed, and there is no need to have read the article series beforehand. Some familiarity with Python may be helpful, but is not essential. Attendees will leave with a practical introduction to agentic AI and a clearer view of how these techniques can be applied within actuarial work.

Speakers: Tim Hill, Zenith Actuarial and Srinivasan Iyengar, Pacific Life Re

AI-Powered Financial Explainability: Multimodal Root Cause Analysis

Life insurance companies operate in highly complex financial environments where capital and balance sheet movements are influenced by multiple interacting drivers. Traditional analysis methods rely heavily on manual interpretation, making it difficult to generate timely, consistent, and explainable insights for senior management.

This use case presents a multimodal AI framework that integrates structured financial data, historical trends, and unstructured business context to enable automated root cause analysis of capital and balance sheet movements. Leveraging Databricks as a unified data platform, the solution builds machine learning models that learn relationships between inputs (assumptions, market variables, business drivers) and outputs (capital, profitability metrics) across multiple scenarios and time periods.

The framework adopts a Retrieval-Augmented Generation (RAG) architecture with agent-based orchestration. Multiple agents dynamically retrieve relevant structured data, historical patterns, and domain-specific documents, combining heuristic rules with expert knowledge and machine learning insights.

The result is an interactive system that allows users to query financial movements (e.g., changes in capital by line of business) and receive contextual, explainable outputs. This significantly enhances decision-making by providing faster insights, reducing manual effort, and improving transparency in financial reporting.
The solution bridges the gap between actuarial complexity and business understanding, enabling scalable, AI-driven financial intelligence.

Speaker: Avnish Nainawatee

From Claims Narratives to Causal Graphs: An Agentic Pipeline for Casualty Actuarial Intelligence

Casualty claims, particularly in professional liability, encode complex chains of alleged failures, contributing factors, and cascading consequences across heterogeneous document formats: narrative reports, expert assessments, loss runs, legal filings, and photographs. Yet actuarial practice still collapses this multimodal, causally rich information into flat tabular representations, leaving the most information-dense data source in insurance largely unexploited analytically.

This session presents a production-grade pipeline combining LLM-based multimodal extraction, formal ontology definition, and directed acyclic graph (DAG) construction within a multi-agent architecture. Vision-language models and large language models jointly extract causal entities and relations across modalities, producing DAGs that encode full liability chains. A domain-specific ontology, defining entity types (actors, professional duties, alleged breaches, damages), relationship taxonomies (causal, temporal, aggravating), and structural axioms, ensures semantic consistency and interoperability. LLM-assisted ontology learning bootstraps this process from claims corpora while preserving actuarial interpretability. The agentic orchestration layer coordinates specialised services for data preparation, graph construction, and confidence validation, with human-in-the-loop checkpoints aligned to actuarial governance requirements.

We demonstrate direct applications to professional liability: enriching loss development factor estimation by conditioning on graph-identified loss mechanisms, identifying latent loss drivers across heterogeneous professions to support underwriting appetite calibration, and querying causal patterns for proactive risk prevention. We address key challenges including LLM context window constraints on large graph processing, causal abstraction complexity, cost-performance trade-offs, and the change management required within actuarial teams to operationalise these techniques at industrial scale.

Speakers: Aurelien Couloumy, Dylogy and Waswate Ayana, Convex

From Data to Value: Blockchain Tokenisation in Cyber Risk Underwriting and Reinsurance Settlement

Building on its engagement with the Financial Conduct Authority (FCA) consultation on the prudential regime for cryptoasset firms, the Blockchain, Digital Assets and Fintech Working Party of the IFoA is exploring how distributed ledger technology (DLT) can reshape both data and value flows within insurance markets.

DLT enables the tokenisation of assets and policy data, allowing both financial value and data to be represented, shared, and transacted in a secure, programmable, and auditable manner. While much of the focus has been on payments, its broader application lies in transforming how risk is assessed and settled.

This session first examines the application of tokenisation to policy and risk data, with a focus on cyber insurance. By structuring risk-relevant data as tokenised, permissioned datasets, insurers could move toward more dynamic, telematics-style underwriting, where exposures are continuously updated and pricing becomes more responsive to real-time signals.

Building on this foundation, the session then explores how similar principles can be applied to reinsurance settlement. The global reinsurance market channels hundreds of billions of dollars annually, yet settlement remains slow and fragmented. Tokenised deposits and regulated stable-value instruments could enable near-instant, programmable settlement, as reflected in emerging initiatives such as Fnality International.

The Working Party will present a practical framework linking tokenised data and tokenised value, highlighting implications for settlement finality, liquidity risk, data governance, and actuarial modelling.

This session offers a forward-looking perspective on how actuaries may operate in a future where both underwriting and settlement become increasingly real-time, data-driven, and programmable.

Speaker: Mohamed Soliman, BoE/Bermuda Monetary Authority

Foundation Models in Actuarial Practice: Exploring Use Cases and Value Creation

Foundation models are emerging as a promising new paradigm for actuarial practitioners, offering strong performance on tabular data with minimal feature engineering and hyperparameter tuning. However, the framework for deploying these models within actuarial practice is still developing, and legacy techniques remain deeply embedded in business-as-usual modelling tasks.

This session explores the application of foundation models in actuarial contexts, focusing on where they deliver meaningful value relative to established techniques. Building on recent working party research comparing a tabular foundation model with classical approaches on an insurance lapse problem, we extend the discussion to a broader range of actuarial applications.

Our findings indicate that model performance depends critically on the underlying structure of the problem. Well-understood, additive environments often favour traditional models such as generalised linear models, while more complex or less structured settings are more likely to benefit from foundation models.

Key trade-offs are considered, including interpretability, probability calibration, workflow efficiency and governance. Attendees will gain a practical framework for assessing when foundation models are likely to add value in actuarial contexts.

Speakers: Cillian Williamson, University College Cork, Scott Hawes, Barnett Waddingham and Karol Galowski, EY

A quantum walk through actuarial space

Quantum computing opens an entirely new horizon for actuaries. Despite their reputation, quantum concepts can be surprisingly intuitive when framed in actuarial contexts. This introductory session will leverage actuarial examples to demystify quantum computing. We will cover:

1. Qubits and quantum circuits
2. Quantum effects like superposition and entanglement
3. Encoding and manipulating information on a quantum computer
4. Actuarial examples like annuity and option valuations
5. Speedup through quantum algorithms
6. Implications of quantum for cybersecurity and other non-actuarial applications. 

Speaker: Muhammad Amjad, Canada Life

Machine clean: using large language models to tidy datasets

By this point, we have all interacted with the powerful technology available in Large Language Models (LLMs), and most of us will have encountered at least one workshop on how to improve your prompts to get the most out of the model.

This talk will guide the audience through more advanced techniques available when interacting with an LLM through Python. Techniques such as fine-tuning and biasing can be used to influence model output and enrich results. We also show how the user can extract the confidence (or 'logprobs') of a response and leverage this to determine which outputs are worth focusing human review. This is all done via an actual case study of cleansing a messy cause of death dataset.

Key takeaways will be that it is possible to integrate LLMs directly in your Python code, and the extra techniques that this can unlock.

Work behind this talk can be found in the article below, which we intend to expand on: https://www.theactuary.com/2026/03/16/machine-clean-using-large-language-models-tidy-datasets 

Speaker: Kay Khine Myo, Hiscox

Teaching a LLM running on your laptop to speak actuarial Excel

LLMs have proved very consequential in technical work which relies heavily on coding, as code is a byproduct of language and for certain programming language, public websites like Github or Stack Overflow provide billions of data points for a LLM to learn from. When it comes to actuarial applications, one of the largest barrier to entry is specialized software. Excel is by far the most largely used piece of software by actuaries (both in the life and non-life) and it is the logical first step when trying to integrate more closely LLMs and actuarial work. In this talk we present an open framework for teaching small open-source coding models to generate end-to-end actuarial Excel workbooks from a natural-language specification and a source Excel file with some data. The approach combines four components: a curated corpus of annotated actuarial spreadsheets, synthetic workbook generation using a frontier model, a deterministic representation pipeline that converts Excel into a purposely designed, efficient LLM-readable format, and parameter-efficient fine-tuning of compact models to generate complete Excel files. We show how this intermediate representation makes spreadsheet structure explicit, reduces hallucinated formulas, and enables automated validation, audit trails, and reproducible benchmarking. The framework is evaluated on practical actuarial tasks such as non-proportional reinsurance pricing and triangle-based reserving, and compared with prompt-only use of general-purpose LLMs and commercial spreadsheet assistants. Beyond model performance, we discuss governance, local deployment, and how an actuary-in-the-loop review process can align outputs with actuarial judgment.

Speakers: Manuel Caccone, Italian Society of Actuaries and Giacomo Maugeri, Independent Researcher

Leveraging technology to streamline and improve sustainability reporting: How do I know if my sustainability report is any good?

Sustainability reporting requirements for insurers are expanding rapidly, driven by evolving regulatory frameworks, investor scrutiny, stakeholder expectations and regional fragmentation. Compared to traditional financial reports, sustainability reporting is characterised by a high degree of uncertainty, and insurers face a balancing act between greenwashing and greenhushing, whilst ensuring they meet international frameworks and convey their intended messages. 

From climate risk disclosures to broader ESG metrics, actuarial teams are increasingly expected to support reporting and disclosure, using their skillset to interpret, quantify and communicate sustainability information across multiple overlapping standards. The result can feel less like strategic analysis and more like a compliance burden.

The Sustainability Reporting & Disclosure Working Party has spent the past year rationalising the leading global frameworks into a comprehensive database, overlaid with LLM-enabled analytics designed specifically to ease this reporting burden and improve sustainability reporting. This session will demonstrate how LLMs can augment actuarial judgement by automating the heavy lifting of reviewing documentation and frameworks, surfacing insights from unstructured data and enhancing reporting processes. We will showcase real-world use cases by comparing published sustainability reports against global frameworks. 

Attendees will leave with:

• A clearer understanding of the evolving sustainability reporting landscape for insurers
• Practical insight into how data analytics can reduce reporting complexity and operational burden
• An appreciation for how enhanced reporting can create strategic value, not just regulatory compliance.

Speakers: Lloyd Richards, Crowe UK and Cristian Calin, Zurich

Speakers: Jonathan Choi, Cybercube, Simon Margetts, Moodys, Stephanie Race, Earth Analytics Group

Including AI² 2026 winner announcement.