CEO Corner

Why this is an operating decision for today

Actuarial work has used statistical learning and computational approximation for years. The new opportunity is to connect those methods with systems that can search technical material, assist with code and coordinate controlled analytical work. These capabilities make it possible to redesign the work surrounding a valuation while selectively improving the models within it. Adoption need not wait for an autonomous system capable of exercising actuarial judgment.

There is also a concrete governance timetable. OSFI’s final Guideline E-23, published in September 2025, takes effect on May 1, 2027. It extends model risk management expectations across federally regulated financial institutions, including insurers, and explicitly addresses AI and machine learning. As of this edition, that effective date is still ahead. I would use the preparation period to establish model inventories, ownership and validation evidence that cover both established actuarial models and new AI components.

The International Actuarial Association has also published educational papers on AI governance, testing and documentation. These are resources for professional practice, not universal regulatory approval. Together, the technological opportunity and the developing governance framework support a deliberate programme of adoption. Delay carries an opportunity cost: teams postpone learning where AI works, where it fails and how much review it requires. That learning is specific to an insurer’s products, data and decisions; it cannot simply be bought later with a software licence.

Give each AI component a precise actuarial role

I would distinguish three roles before approving an investment. Statistical learning estimates relationships: mortality differentials, surrender behaviour, claim severity or expense patterns. Numerical approximation accelerates an expensive calculation, for example by estimating a liability surface from a controlled set of full model runs. Generative AI supports work with language and code: retrieving a product clause, proposing a test or drafting an explanation from approved outputs. These roles require different evidence of reliability.

A language model’s fluent explanation does not establish that a mortality estimate is calibrated. A proxy’s close agreement with a reference engine does not establish that the reference engine represents the contract correctly. A predictive model’s strong historical performance does not establish that a proposed management action will cause the predicted outcome. The chief actuary’s first responsibility in an AI programme is to make these distinctions explicit.

My preferred architecture gives the actuarial engine responsibility for specified calculations, gives AI narrowly defined analytical or coordination tasks, and gives authorised professionals responsibility for assumptions and decisions. Where AI changes the numerical method itself, that component enters the model inventory and validation process. It does not disappear inside an application described as an assistant.

Improve the assumption process before increasing its complexity

For a life portfolio, mortality analysis begins with exposure construction, attained age, duration, underwriting class, benefit amounts and the timing of reported deaths. Poisson models with exposure offsets, generalised additive models and survival models provide useful, interpretable starting points. Machine learning can investigate interactions that a broad actual-to-expected ratio conceals, but sparse experience still requires credibility and disciplined pooling. A highly granular estimate is not necessarily a more reliable assumption.

I would require validation on later experience periods and relevant product cohorts, with protection against information leakage across the same policyholder. I would examine calibration by segment, the effect of changing business mix and sensitivity to missing data. Fitting past mortality levels and selecting future mortality improvement are separate tasks. A model can fit recent experience closely while providing little evidence about a long-term improvement assumption.

Lapse modelling deserves the same discipline. Survival or competing-risk models can represent death, surrender and other exits; richer models can investigate the interaction of credited rates, market alternatives, surrender charges and distribution. Yet historical policyholder behaviour may not cover the rate conditions or guarantee incentives that matter under stress. I would retain explicit scenario judgments for those gaps. A predictive association between an intervention and retention also needs causal analysis before it supports an intervention strategy.

Life and annuity valuation must preserve the contract

For a simple term contract, a prospective cash-flow calculation projects premiums, death benefits and expenses using consistent decrement and discount assumptions. Whole life and universal life introduce additional dependencies: account values, premium flexibility, credited rates, surrender values and guarantees. Annuities introduce survival-contingent payments, possible guarantees and policyholder options. AI can help assemble specifications and identify inconsistencies, but each material contractual feature must survive translation into executable logic.

The same demographic change can have different financial consequences across products. Lower mortality generally extends payments under a life-contingent annuity, while reducing expected death claims on a term assurance book. Even then, premiums, expenses, lapses and guarantees affect the total result. An explanation that says only ‘mortality improved’ is inadequate. I would expect it to identify the affected cohorts, the changed cash flows and the consequence for the particular valuation basis.

IFRS 17 adds another important distinction. Fulfilment cash flows and the contractual service margin have different roles; a change in a simple gross-premium reserve is not automatically a change in reported profit. Treatment depends on the measurement model, the source of the movement and whether it relates to future or current and past service. I would use AI to help trace the accounting treatment to approved policy and run evidence, with the classification and conclusion reviewed by the responsible actuarial and finance teams.

Faster stochastic valuation should buy more challenge

For products with material options and guarantees, the computation can become much more demanding. A risk exercise may require an outer set of economic scenarios and an inner valuation at each future state. Regression, least-squares Monte Carlo and machine-learning proxies can reduce the cost of selected calculations. The Society of Actuaries has published practical life and annuity case studies on this use of predictive analytics and machine learning. That establishes a credible area to investigate, not a performance guarantee for a particular book.

I would define the proxy’s permitted input domain before training it. Validation should include stressed states, combinations of risk factors and contracts close to guarantee thresholds. It should examine systematic bias, tail error and sensitivity estimates, not just average fit. In a purely illustrative portfolio with a liability of CU 10 billion, a systematic valuation error of 0.1% represents CU 10 million. An apparently small percentage can matter before considering the error’s effect on hedges or capital.

The reference engine remains essential. An out-of-domain case should trigger a full calculation or an explicit exception, and model changes should trigger reassessment of the proxy. Real-world scenarios used to study future outcomes and risk-neutral scenarios used for market-consistent valuation serve different purposes; the design must preserve that distinction. My investment case would ask how much additional stress testing the saved computation makes possible. Repeating a narrow set of scenarios faster is a smaller achievement than understanding a previously unexamined exposure.

Connect the liability result to assets, liquidity and capital

A liability estimate becomes useful when it changes a decision. In asset-liability management, I would ask whether improved analysis changes the interpretation of duration gaps, key-rate exposures, convexity or cash-flow shortfalls. A duration match can coexist with basis risk, spread risk, illiquid assets and policyholder options. A model trained on ordinary conditions may be least dependable when surrender incentives and asset realisation costs interact.

For capital, the concern moves further into the joint tail. A proxy that reproduces average liability values may still distort the ranking of adverse scenarios, a tail percentile or the allocation of capital across risks. Correlation and dependence assumptions deserve their own challenge. I would report approximation error separately from scenario sampling error and from uncertainty in the underlying assumptions. Combining all three into one unexplained confidence score makes the result harder to govern.

The same principle applies beyond life insurance. In P&C reserving, an AI method should earn its place against relevant chain-ladder, Bornhuetter–Ferguson or other benchmarks, with attention to claims inflation, settlement changes and portfolio mix. Greater predictive complexity does not remove the need to explain why the selected reserve is appropriate for the business.

A valuation close that can answer the next question

Consider an illustrative annuity close in which liabilities increase unexpectedly. I would want the workflow to separate new business and portfolio movements, expected run-off, economic changes, experience deviations, assumption updates and model corrections. Controlled reruns should establish those contributions. Sequential attribution can depend on the order in which changes are applied, so the report must state its convention and identify material interaction effects.

An AI assistant could retrieve the previous approved basis, locate changed assumption records, propose the required reruns and assemble the resulting evidence. It could draft commentary only after the relevant numerical outputs are available. If the model cannot reconcile the movement, the assistant should expose the unresolved residual. It should not supply a persuasive economic story to make the table appear complete.

The review pack would contain a cash-flow run-off chart, a movement bridge, a cohort experience analysis and sensitivity comparisons. Each material figure would identify its source run, model version and assumption basis. The narrative would distinguish a calculated contribution from an interpretation and a proposed action. This is where I see a substantial near-term opportunity: giving the actuary a better starting point for investigation and enough time to pursue the second question.

Make uncertainty and responsibility visible

The professional standard should rise with the speed of production. ASOP No. 56 provides a useful US reference for intended purpose, model understanding, testing, validation and documentation. It is not a global substitute for local standards. The applicable professional and regulatory framework must be identified for the entity and work concerned. Across those settings, the central question remains whether another qualified reviewer can understand the material judgments and limitations.

For generative AI, I would test unsupported numerical claims, incorrect source retrieval, omitted qualifications and responses to conflicting evidence. I would also test whether untrusted material can cause an agent to change its instructions or attempt an unauthorised action. Access to a model repository should not imply authority to release a valuation. Changes to the language model, retrieval sources, prompts or connected tools can change behaviour and require proportionate re-evaluation.

For predictive AI, monitoring should distinguish a changing data population from deterioration in calibration or decision quality. A portfolio mix shift can change observed results without making the model wrong; stable aggregate error can hide deteriorating performance in a small but material cohort. Where outcomes affect customers, lawful use, fairness and the consequences of errors need specific examination. A named reviewer is meaningful only if that person has the evidence, time and authority to challenge the result.

What I would put into the first 90 days

I would begin with one recurring decision and its evidence trail. A life valuation movement review or an annuity assumption investigation is a stronger starting point than a general instruction to ‘use AI’. Establish the current elapsed time, actuarial preparation effort, reviewer effort, unresolved exceptions and rework. Agree acceptance criteria before choosing the technology, including how material errors will be identified.

During the first month, reconcile the selected data, document the model and basis, establish a historical benchmark and assign decision rights. During the second, run the proposed workflow in parallel on historical cases, deliberately including difficult cohorts, adverse scenarios and conflicting documents. During the third, evaluate it in a controlled reporting cycle and decide whether to extend, revise or stop. Ninety days is a proposed learning period, not a promise that a material valuation model can be approved on that timetable.

The economic assessment must include integration, computation, validation, ongoing monitoring and human review. Minutes saved in drafting can be offset by hours spent checking plausible errors. I would measure additional questions answered before a decision deadline, the quality of movement explanations and the stability of results after changes. I would also protect actuarial development: junior colleagues should reproduce calculations and explain failures, rather than learn only to operate an interface.

The decision I would make now

I would authorise a bounded actuarial AI programme now, with a named actuarial owner, independent challenge and measurable acceptance criteria. I would fund the data and validation work alongside the application. I would expect the first result to demonstrate a better investigation or a more reliable decision process, with the limitations visible.

The longer-term prize is an actuarial function able to spend more of its capacity examining assumptions, testing difficult scenarios and advising the business before options close. That capability develops through repeated use, correction and professional challenge. The time to begin is now because building it takes longer than installing the technology.