The most interesting AI work in actuarial right now is not happening in pricing. It is happening in the model office — and specifically in the slow, repetitive, evidence-heavy world of ALM and solvency, where the gap between what the analyst spends a quarter doing and what an agent can credibly automate is the widest in the firm.
Asset-liability management and solvency modelling are information-heavy disciplines. Scenarios, assumptions, asset data, liability models, management actions, capital requirements, governance papers and board communication, all repeating on a quarterly rhythm. An AI agent is more than a chatbot — it can retrieve documents, call tools, run code, interact with models, summarise results, check consistency and prepare draft outputs. That makes it well suited to recurring model-office work. It also makes it dangerous if deployed without controls, because ALM and solvency outputs are material in a way that demands a different bar.
Why ALM is the natural agent use case
ALM requires repeated integration of data, models and judgement. The questions it answers are stable across cycles, even if the answers move:
- How do assets and liabilities behave under interest-rate scenarios?
- How do management actions affect solvency?
- Which assumptions changed since the previous quarter?
- What explains the movement in capital?
- What are the liquidity implications of stress events?
- What sensitivities should the board see?
- Which model limitations are relevant?
Each of those needs quantitative analysis and narrative explanation. Agents can connect the two.
The research, briefly
Vu’s 2025 AFIR-ERM paper explores AI agents in ALM — LLMs, retrieval-augmented generation, tool-calling and multi-agent collaboration. The core idea is that agents can automate labour-intensive tasks and provide reasoning support inside ALM workflows. The actuarial model office is increasingly being framed as a setting where agentic AI manages data pipelines, model execution, regulatory updates and reporting workflows. The opportunity is significant. The governance requirement is at least as significant — an agent that can run an ALM model or alter assumptions is a very different control problem from an assistant that summarises a report. We apply the controls described on our How we use AI page to all of this work.
Five concrete use cases
Scenario management. An agent retrieves approved economic scenarios, compares them to prior runs, generates scenario documentation, and checks that scenario files match the approved assumptions.
Assumption comparison. An agent identifies changes in lapse, mortality, expense, credit spread, inflation, yield curve or management-action assumptions and drafts a movement commentary that a human reviewer can verify line by line.
Model run orchestration. With proper controls, an agent triggers model runs, checks logs, compares outputs and flags failed runs. It does not override controls or approve results — those gates remain human.
Solvency analysis. An agent assembles solvency movement analysis, reconciles capital components, summarises sensitivities and drafts risk-committee commentary.
Board reporting. ALM outputs are often complex. An agent turns technical results into clear board-level explanation — provided every statement is grounded in approved outputs and the agent never invents the connecting narrative.
The ten ways agentic ALM goes wrong
The risks are specific and worth naming: wrong tool calls; unauthorised assumption changes; use of outdated source documents; hallucinated explanations; hidden model-run failures; inadequate access control; insufficient audit trail; over-reliance by users; leakage of sensitive data; difficulty validating multi-step agent decisions.
Each has a specific countermeasure. The general pattern: the agent proposes, retrieves, compares and drafts. Humans approve assumptions, selections and decisions. Multi-step agent chains require explicit human review gates between steps, not after the final output.
What good governance looks like
A controlled agentic ALM environment has role-based access; approved document repositories; model and assumption version control; tool permissions by user role; logged prompts, retrieved sources and tool calls; human approval gates; reproducible model runs; separation between development and production; validation of agent outputs; monitoring of errors and exceptions; and a written policy on when AI may be used and when it must not be. The IAA’s AI Governance Framework and NIST’s AI RMF are the right scaffolds. Skipping either is a sign the firm has not yet taken agentic ALM seriously.
The South African overlay
South African insurers operating under SAM need robust capital, risk and ORSA processes. ALM also interacts with liquidity, credit risk, market risk, policyholder behaviour and product guarantees. Agentic AI helps with recurring production work, but the regulatory accountability and actuarial sign-off do not move. The named actuary is still the named actuary.
Where this is heading
AI agents are likely to become a major part of the actuarial model office, especially in ALM and solvency where the work pattern fits the technology best. The future is not an autonomous black-box solvency function. The future is a governed agentic workflow where actuaries remain accountable, supported by AI systems that are integrated, traceable and controlled — and where the time savings are spent on better second-line review, not headcount reduction.
If you are scoping what an agentic model office looks like for your firm, our Finance Modernisation practice and our Symphony Automate product cover the orchestration, evaluation and governance.