IFoA Workshop - AI for Actuaries · Dr Rohan Yashraj Gupta

Designed and led by Satya Sai Mudigonda, Rohan Yashraj Gupta and expert AI engineering mentors.

At 9:54 in the morning — before a single slide of theory — participants watch an AI agent answer a question every pricing actuary knows by heart: "Why is this policy priced the way it is, and can I sign it?" The agent loads the official rating table, explains each factor in plain English, and drafts a memo an actuary could put their name to. Nobody told it which tools to call, or in what order — it decided.

The rest of the day is spent earning that moment. By 5:15 PM, every participant has built that agent themselves, watched it fail, and fixed it — and understands exactly why a governed AI agent is different from a clever chatbot.

That is the promise of this workshop: not a tour of AI, but a day of building, breaking, and governing your first actuarial AI agent — grounded from the first minute in real pricing, reserving, and reporting work.

The Series

The programme runs as a day-long, hands-on event — two live sessions of three hours each — followed by a mandatory two-week group case study with mentor-led online support. Attendance alone doesn't earn completion; building something does.

The series opened in Mumbai on 15 May 2026 at the Hilton near the airport, Gurugram on 10th July 2026 at Club 4, and continues in Bangalore on 24 and 25 July 2026 at Four Points by Sheraton, Whitefield.

Why This Workshop Exists

AI is no longer adjacent to actuarial work — it is arriving inside it. Pricing teams are benchmarking GLMs against gradient-boosted trees. Reserving and reporting teams are drafting documentation with large language models. And a newer shift is already underway: from models that predict a number to agents that take a goal, call your tools, and act — the difference between an analyst who advises and a colleague who does the work.

At the same time, regulators are moving faster than most insurers' procurement teams. The EU AI Act treats life and health pricing as high-risk with human oversight as a design requirement; the NAIC Model Bulletin and Colorado's SB21-169 push governance and anti-proxy testing in the US; IRDAI's guidance in India is evolving; and the IFoA Actuaries' Code and FRC Technical Actuarial Standards apply whatever the tooling. AI fluency is quietly becoming a professional expectation, not a nice-to-have.

Yet most actuaries face the same practical gap: deep statistical training and strong GLM intuition, but limited hands-on exposure to Python, machine-learning workflows, generative AI, and the emerging world of AI agents. Generic AI courses don't close this gap, because they aren't grounded in exposure, frequency–severity structure, fairness in pricing, or the professional duty to explain and sign off on a model.

This workshop was built to close exactly that gap — by actuaries, for actuaries — with every example drawn from Life, Health, and General Insurance.

What the Community Is Asking For

Practising actuaries consistently raise the same questions, and the programme is structured around them:

"I know GLMs — how do tree models actually compare?" Not in theory, but on a realistic pricing dataset, with an honest, out-of-time leaderboard — including where the humble GLM holds its own.
"How do I explain a machine-learning price to my chief actuary, my regulator, and my customer?"Interpretability (SHAP), fairness, and governance are treated as first-class topics, not afterthoughts — one prediction, three explanations.
"Can GenAI actually help with documentation and reporting — safely?" Hands-on prompt engineering with clear guardrails on confidentiality, hallucination, and validation.
"What is an AI agent, really — and could I build one?" Every participant watches one built live, breaks it, governs it, and then builds their own in the case study.

What Makes This Different

A single, honest through-line runs from the first slide to the last: the reasoner narrates, the tools know, and the human signs. Rather than nine disconnected topics, the whole day follows one actuary's real problem — explaining a price to a chief actuary who has to sign it — carried across both sessions through a purpose-built illustrative insurer.

Three design choices set it apart from a generic AI course:

You watch an agent break — on purpose. The centrepiece is a live build where the agent confidently invents a rating factor that doesn't exist, and is then fixed with a deterministic guardrail you can unit-test. Watching an agent hallucinate and get repaired teaches more than any polished demo.
Your existing models become the agent's tools. The GLM, XGBoost, and SHAP you already trust aren't replaced by AI — they are wrapped as governed functions the agent calls. Nothing you know is thrown away; it's given a driver.
Governance is built in, not bolted on. Fairness, auditability, tool scope, and a named human signature are engineered into the build, mirroring the regulatory direction the profession is heading in.

Inside the Day

Session 1 — The Agentic Actuary: Foundations

Lectures + guided Python walkthroughs + discussion (3 hours)

The agentic shift — where actuaries and AI meet across pricing, claims, fraud, and operations, illustrated with documented industry systems (Lemonade's AI-driven quoting and claims, John Hancock's Vitality programme, Tractable's photo-based motor damage assessment) — and the one distinction the day hangs on: chatbot versus agent.
The reasoner — what large language models can and cannot do, why hallucination happens and why actuaries must care, confidentiality discipline, and structured prompt engineering — ending with every participant making their first live call to a frontier model from a Colab notebook.
The tools — the anatomy of insurance data (policies, exposure, frequency and severity); GLMs in statsmodels, then XGBoost on the same book; evaluation with MAE, RMSE, AUC, and actuarial lift charts — all built as clean functions an agent can later call.
Trust and governance — SHAP for global and local explanation, proxy discrimination, fairness as a plural, and the regulatory landscape (IFoA standards, IRDAI, NAIC's AI Model Bulletin, Colorado SB21-169, and the EU AI Act), distilled into a pre-production checklist.

Session 2 — Building Actuarial Agents

Conceptual overviews + guided demos + live coding (3 hours)

Inside the agent loop — the anatomy of an agent (reasoner, tools, memory), the ReAct pattern, function calling, and the framework landscape — plus an honest answer to "can't I just use ChatGPT or Claude?" (great for prototyping; not what a chief actuary can sign).
The live build — an actuarial AI agent, a Pricing Logic Explainer, built live in the Agno framework on Google's Gemini API: it takes rating logic as input, explains it in plain English, fails visibly when it hallucinates, and is repaired with a guardrail — then ported across lines of business in a few lines each.
AI-assisted coding — building tools faster using the generative-AI features built directly into Google Colab, so participants leave able to do this on Monday with nothing to install.
Scaling up — a capstone "analyst agent" that runs the morning's entire modelling pipeline on one instruction, retrieval-augmented generation (RAG) so agents cite your own documents, multi-agent review (an analyst that drafts and a reviewer that recomputes), and MCP for sharing tools across a team.
Case-study briefing — tracks, teams, deliverables, and evaluation.

All modelling runs on a purpose-built hypothetical insurer — ABC Insurer, with motor, health, and life datasets in INR — so every participant works with data that looks like their day job. Nothing sensitive, nothing real, everything recognisable.

The Two-Week Case Study

Participants form teams of ~5 and choose one track — and in this edition, every track ships a real, governed agent:

Model Build & Interpretation — wrap statistical and ML models as tools and let an agent run the comparison and draft the interpretation.
Documentation with GenAI — a documentation agent grounded in your own methodology (RAG) with a reviewer gate.
Build a Simple AI Agent — a working agent that automates a real actuarial task end-to-end, optionally exposed to a desktop assistant via MCP.

Final submissions include a Python notebook that runs from a clean environment, a before-and-after trace showing a guardrail catching a failure, the governance checklist answered with evidence, and a one-page executive summary written for a chief actuary — with mentor support available throughout the fortnight.

What Participants Leave With

The ability to apply AI techniques inside actuarial workflows, not alongside them.
Confidence to interpret and validate predictive models — and to defend that interpretation to a signing actuary or a regulator.
Practical fluency in using GenAI for reporting and documentation with appropriate professional scepticism.
A genuine, hands-on understanding of AI agents — and the experience of having built, broken, and governed one.
A complete take-home kit: six Colab notebooks that run end-to-end, prompt templates, agent-design templates, cheat sheets, and a facilitator-quality handbook.

Design Principles Behind the Programme

A few things distinguish this workshop from a generic AI course:

Practitioner-first. Every concept answers "what does an actuary do with this on Monday morning?"
Honest over impressive. Model comparisons report what the data shows — including where the GLM beats XGBoost. If a model has limits, they're named on the same slide.
Rigorous sourcing. Every industry claim traces to public filings, peer-reviewed papers, or regulator publications. No recycled AI hype.
All three lines of business. Worked examples rotate deliberately across Life, Health, and General Insurance, so no participant feels the material was written for someone else.
Zero setup friction. Everything runs in Google Colab on free tiers — participants are coding within minutes, not after an afternoon of installs.

Faculty

Satya Sai Mudigonda — Chairman, Sri Sathya Sai Institute of Actuaries. Tech-actuarial consultant with over three decades of experience across the US and India in AI/ML applications, analytics, and actuarial modelling.

Dr Rohan Yashraj Gupta, PhD, FIA, FIAI — GI Actuary at Accenture and Adjunct Faculty. Qualified General Insurance actuary specialising in agentic AI, InsurTech, and actuarial automation. Workshop designer and lead presenter.

Expert AI-engineering mentors — each edition is supported by a lead AI engineer as co-faculty and case-study mentor, bringing production AI-systems expertise to complement the actuarial perspective.

Delivered in collaboration with the Institute and Faculty of Actuaries (IFoA) · actuaries.org.uk · All datasets are illustrative; ABC Insurer is a hypothetical entity created for workshop.