The Agentic Actuary

John Culkin wrote that in 1967 about television and radio¹. It now describes actuaries and AI. The spreadsheet changed how fast we calculated, but actuaries still designed every model. An agent is different: it can be given an objective, choose its steps, write and run code, test what it finds and come back when it needs a decision².

This essay sets out what that means in practice, where the risks lie, and the habits that keep the work sound and the actuary capable.

Less time moving data and formulas between workbooks. More work in calculations and workflows that are written in code, tested and rerun.

The work follows a simple loop:

Take a monthly claims review. You brief the agent: reconcile claims and exposure, compare frequency and average cost with last month, and investigate material movements. The agent writes and runs the code. You make it reconcile to source and test a case you can check by hand. Then you change the question, perhaps a split by product, and it reruns everything, checks included.

Excel still has a place for reviewing inputs and sharing results. What moves into code is the repeated transformation and calculation that should be tested and rerun.

Nobody knows exactly how actuarial roles will change. In an IFoA survey, 44% of respondents expected significant change, 41% some change and 12% complete transformation³. What we can see is that AI automates tasks, not whole jobs, and it touches the layers of actuarial work unevenly².

Juniors spend most of their time in the first two layers, so the pressure lands on them first. Outside the profession, employment among 22 to 25 year-olds in highly exposed jobs has fallen about 19% behind less exposed peers, mostly through hiring. That finding is descriptive, not causal, and says nothing certain about actuaries⁴.

Kavi Pather describes four things that used to come together and that AI can now separate². The work still gets done, but the professional is no longer formed by doing it. Each needs a deliberate habit to put it back together.

1. Doing and learning

Routine work built judgement through repetition. Use agents freely for delivery, but keep some formation work for yourself: build a simple model from a blank page, estimate before you ask, and use AI as a tutor that questions you rather than a copilot that answers.

2. Writing and thinking

For hard problems, the draft is where the thinking happens. Delegate formatting and rephrasing freely. Write difficult first drafts yourself, and ask the machine to challenge them.

3. Output and understanding

Reading a fluent answer feels like understanding it. Form a view first. Then put the answer away and explain it in your own words, trace each material claim to its source, and name what would change your mind.

4. Capability and accountability

An agent can do the analysis, but it cannot accept a professional duty. Delegating to it adds execution without adding anyone accountable. Know what you are claiming, what you relied on and what is uncertain, and escalate what is beyond your competence.

The core professional skill is deciding what you should think through, what the agent should do, where you should work together, and who owns the result².

• Delegate work that is routine, reversible and easy to check.
• Collaborate on exploration, while keeping your own view and a traceable line of reasoning.
• Lead on material judgement that is hard to verify, and use the agent to attack your conclusion.

The right answer also depends on you. The same delegation is leverage for an expert and lost practice for a trainee. And every choice has an expiry date: what an agent could not do last year may be routine next year, so keep testing where the boundary sits.

We propose seven conventions for directing agents. Each adapts principles that actuarial standards already set out.

1. Frame the decision and set the boundaries. Agree the decision, who will use the work and what the agent may do without approval⁵.
2. Make the basis explicit. Name the data, assumptions and methods, and why they fit^6,7.
3. Keep the workings, not just the answer. Retain the executed code and material choices, not the agent's account of what it did^5,8.
4. Challenge the approach and check the results. Test that the work is right and that it supports the decision. Record the checks⁸.
5. Follow changes through the work. Update affected results and checks, and keep earlier versions^8,9.
6. Deliver something people can use. Connect the advice to its basis in a form people can understand and inspect⁷.
7. Own the professional position. Understand the work, explain its limits and defend the conclusions you accept^7,10.

Apply them in proportion to the decision. They are working conventions, not a professional standard, and they do not replace standards, firm controls or professional review¹¹.

Choosing not to use AI is still possible. It may not stay that way: if a validated AI-assisted review reliably finds errors that manual review misses, a human-only process is not more professional just because it is human².

Responsibility, though, is likely to stay with people. So use agents often and keep relearning them. Delegate what is routine and checkable. Protect the effort that still builds your skill. Make your reliance and uncertainty visible. And use the machine to challenge you, not only to answer.