The actuarial teams getting real value from generative AI in 2026 are not the ones that bought a chatbot licence. They are the ones that treat the LLM as a workflow engine — grounded in approved sources, wired into the model office, and reviewed exactly the way any other production process is reviewed.

By Johann van Rooyen, FASSA, CEO

Stripped of marketing, generative AI is now a credible productivity layer over the unstructured side of actuarial work: methodology papers, committee minutes, IFRS 17 disclosures, claims narratives, policy wording, the email thread that decided this quarter’s lapse assumption. Traditional actuarial software is excellent at numerical calculation and weak at context. That is the gap copilots fill.

From summarising emails to running workflows

The first wave of actuarial GenAI was simple — summarise a long paper, draft a committee note, explain a result in plain English. The interesting work has moved on. Recent research demonstrates the shift through four actuarial case studies spanning unstructured-claims feature extraction, retrieval-augmented market comparisons, vision-enabled motor damage interpretation and multi-agent reporting. The same authors who reframed LLMs as both contributors and assistants in actuarial modelling have been showing that the durable value is not the draft. It is the connection between data, expert reasoning and production workflow.

We use AI to accelerate delivery, in a robust, secure and maintainable manner — see our How we use AI page for the controls we apply to our own work.

Four patterns that earn their place

Retrieval-augmented actuarial knowledge. A generic LLM is fluent and can hallucinate. A RAG system grounded on the firm’s approved documents — methodology notes, valuation reports, committee minutes, internal standards — turns the chatbot into a governed knowledge interface. The non-negotiables: citations on every answer, paragraph-level provenance, drafts distinguished from approved documents, access rights enforced.

Structured extraction. Actuarial teams spend significant time converting text into modelling variables — cause-of-loss from claims notes, contractual options from policy wording, assumption changes from a committee pack. LLMs returning controlled fields rather than free text accelerate this. The trap is that extraction errors become hidden model errors. Sampling, reconciliation, confidence scoring and human review on material extracts are mandatory.

Code generation, not code substitution. The strongest pattern is not “ask the model for the final answer” but “ask the model for a reproducible script that can be tested, versioned and re-run”. Experience investigations, data checks, sensitivity runs. Generated code goes into version control and through review like any other code. It does not get pasted into production.

Agentic workflows. Agents can plan multi-step tasks, call tools, query databases and draft commentary. They can also act — which makes access control, logging, segregation of duties and rollback a precondition, not an afterthought. The teams who get this right keep the agent on the propose-and-draft side of the line and keep the actuary on the approve-and-decide side.

Where it has to land first

Six workflows are credible today: claims triage and reserving support; assumption management; IFRS 17 movement commentary; model documentation; regulatory horizon scanning; recurring management reporting. Each should be assessed by materiality. An internal study summariser is not the same risk as an AI process that influences pricing, underwriting or financial reporting — and the controls should not be either.

South African governance, in one paragraph

The supervisory picture is now coherent. The November 2025 FSCA / Prudential Authority report flagged explainability, disclosure, conduct, operational risk, model risk and cybersecurity as priorities. POPIA’s automated-decision-making provisions are relevant where AI materially influences underwriting, claims or benefit decisions. The IAA’s working set of papers on AI Governance, Testing and Documentation gives the actuarial-specific scaffold; the American Academy of Actuaries’ professionalism paper on generative AI is the cleanest reminder that professional standards apply regardless of how the analysis was produced.

What good actually looks like

Five properties separate a mature actuarial GenAI deployment from a clever pilot. It is grounded in approved sources rather than general internet text. It is traceable, with every output linking back to data, documents, assumptions and model runs. It is permissioned, so users only access what their role allows. It is testable, with sampled outputs, benchmarks and drift monitoring. And it is integrated into the existing actuarial process — not running in parallel as an uncontrolled shadow workflow.

Drop any one of those and the firm has built a productivity tool that the second line cannot defend.

If you want help moving from copilot experiments to a governed actuarial workflow, our Finance Modernisation practice covers the data, assumptions, lineage and orchestration that production AI workflows actually require. See also our companion piece on working with AI agents.

Sources