Control, Collaboration, Augmentation

Every actuary likely remembers their first job. The supervisor hands the trainee a list of steps and often says, “Do step 1 and come back to me.” The trainee returns, the supervisor reviews, and usually sends the work back with corrections. Maybe the trainee misunderstood the
direction. Maybe they did it wrong. Maybe the supervisor failed to explain clearly. This back-and-forth repeats until the job is done.

We have all lived this on one side or both. It is the process that makes sense for working with someone completely new: spell out the steps, check after each one.

Today, we have a different kind of “trainee” called artificial intelligence (AI). Some assume that because AI has the dictionary memorized, we can skip the supervision step. Others let AI design its own process and check only at the end. Both can work, but on close inspection, AI
probably needs more supervision than a new trainee. The trainee at least has a couple of decades of interacting with humans. AI begins with broad written knowledge but adds to it only once or twice a year, and tomorrow it usually will not remember the corrections you made today.

Depending on the system and how it is configured, AI may not retain project-specific instructions or corrections across sessions. Persistent memory, retrieval tools and enterprise knowledge systems can change this behavior, but they also require deliberate governance and validation.

So, the “total oversight” approach we use with a new employee may be needed for AI for a long time, at least until someone gives AI persistent medium-term memory.

We will call this process simply Control. And we see it as a part of a spectrum of control, collaboration and augmentation that an actuary might experience working with AI.

COLLABORATION

This article was written as a collaboration. We planned it together, divided the work by comparative advantage and reviewed each other’s drafts. That is the same approach we should use when working with AI if several project steps suit AI well, but not every step. The steps the actuary may own are those that require judgment, empathy or consideration of long-term consequences, especially when the long term is unlikely to be a simple extrapolation of the present.

This is also how a professional works with a powerful tool: Use it where it does the job well and do the work yourself where you can do it better.

AUGMENTATION

If you have ever had a good mentor, you know what Augmentation feels like. The mentor helps you prepare for each step and, when asked, looks over your shoulder. They are there for planning, data review, assumption setting, model choice, interpretation, documentation and
presentation. Not to do the work, but to offer insights into how each step might be done differently.

This is also a good way to work with AI in the right circumstances. We often ask AI to suggest an outline for a project. We rarely use exactly what it gives us, but it is great at overcoming the blank-page problem. As we go, we might ask AI to review what we have done, tell us what we are missing, or argue against our conclusions. No matter the stage, we can get some reaction, even if we choose to ignore it.

Here are three thoughts we have when considering control, collaboration and augmentation modes:

1. We are always working with AI, not handing work off to it. That is, in our estimation, a difference from some contemporary narratives.
2. AI cannot take responsibility, so it cannot be a decider.
3. AI can generate analysis and recommendations that resemble judgment but lacks the experience and emotional grounding to exercise real judgment.

With these thoughts on the three modes in hand, an expanded spectrum came to mind. The first two modes in that spectrum are the two choices that are most often discussed in AI literature—Automation and Oversight. With the third mode, Direction, we add a new idea, new at least in discussions of supervising AI. That is the idea of defining the process that AI needs to follow in advance, rather than waiting until it finishes and asking what process it followed. This again follows how we tend to supervise inexperienced trainees.

Table 1. ‘Mode’mentum: Expanding on Control, Collaboration and Augmentation​

This full-spectrum mode consideration adds three co-working modes with less oversight than the Control mode. The full Control mode might be too much for some situations, and the intense interaction of Collaboration or Augmentation modes may be too much professional involvement.

Here are seven considerations for choosing an appropriate co-working mode:

1. Consequences of error
2. Complexity of the assignment
3. Degree of professional judgment required
4. Explainability requirements
5. Governance and review expectations
6. Reversibility of mistakes
7. Data sensitivity, confidentiality and permitted system access

Before considering a co-working mode, the actuary may also want to determine whether the data and documents involved may be provided to the AI system. Relevant considerations may include employer policy, confidentiality obligations, cybersecurity controls, vendor terms, data retention, data residency and restrictions on the use of personally identifiable or proprietary information.

Choose Automation when the task is routine, mechanical, when the consequences for error are low and reversible, and when errors are likely to be caught by downstream validation. The AI performs hygienic processing while governance relies on testing, exception reporting and periodic validation rather than continuous human review.

Choose Oversight when AI can produce a useful draft or analysis, but the final communication carries professional, reputational or disclosure significance and explainability requirements. The actuary reviews the entire work product before it is used, applying human judgment to validate the final output.

Choose Direction when the procedure is well defined and repeatable, allowing AI to follow an established process without step-by-step supervision. The actuary governs the process through procedure design, validation, version control, and review of the standardized results to satisfy governance and review expectations.

Choose Control when every stage of the work involves decisions or affects important actuarial results with high consequences for errors. AI contributes throughout the process, but each stage is reviewed and approved, applying professional judgment, before the next stage begins because the output directly influences the final actuarial result.

Choose Collaboration for complex assignments that require substantial professional judgment and benefit from combining human insight with AI's analytical capabilities. Responsibility is shared across the workflow, but the actuary remains accountable for the final professional conclusions. Each contributes where it is strongest, with the actuary retaining responsibility for professional judgment priorities, and final decisions while AI performs complementary analytical work.

Choose Augmentation when the actuary performs the work but benefits from AI acting as a mentor, reviewer, coach or challenger. AI improves quality by identifying omissions, inconsistencies and alternative viewpoints, while the actuary remains the author of the work
product. This mode is particularly valuable when explainability, completeness or professional quality are more important than automation.

Choose Human-only when the assignment requires personal professional responsibility, ethical judgment, uses data that cannot be shared with AI or involves certification that cannot be delegated. AI is not involved.

IN THE MOOD FOR MORE ON MODES?

Here are examples of actuarial work from life, health, P&C and pensions for each of the seven
modes.

Automation—Unsupervised. Bulk metadata extraction from a folder of historical rate filings, where the AI parses each PDF for filing date, line of business, state and filed rate change, dumping results into a tracker. Worst-case error is a typo in a tracking sheet. Another is auto-
converting legacy GGY AXIS or PolySystems batch output filenames to a standardized convention before archiving. A third is having AI generate first-cut commit messages for a nonproduction scratch repo. These are examples where actuaries probably already operate
without thinking of it as AI use.

Oversight—Human-in-the-Loop. The everyday actuary-AI Oversight work is research synthesis. Examples include feeding the new IFRS 17 amendment package into an LLM and getting a one-page brief, then reading the actual amendment yourself before circulating.
Another is asking AI to draft test cases for a new pricing model from the model specification document, then the actuary picks which tests actually run. A third is generating SQL queries against the policy admin warehouse to pull experience data, where the actuary visually inspects the WHERE clauses before executing. For internal audit, drafting a management response paragraph from a finding write-up and then reviewing before sending is a clean fit.

Direction—Human-in-the-Process. The classic actuarial Level 3 is anything cyclical with a fixed SOP. Three examples: the monthly cash flow testing reconciliation routine where the AI follows defined steps to pull GL balances, compare to actuarial system reserves, populate the variance commentary template and flag breaks above threshold. The quarterly NAIC blank exhibit preparation where the AI fills schedules from defined source systems following a documented mapping. The annual asset adequacy memo refresh where the AI updates
economic scenario references, restates last year’s narrative with current numbers, and produces a redlined draft for the appointed actuary. In health, it could be the recurring trend study refresh; in pensions, the annual actuarial valuation memo for a frozen plan where most
language is templated.

Control—Human-in-the-Model. This level shines on staged model builds where the cost of an undetected error compounds downstream. Three concrete options: a VM-22 PBR build for a new non-variable annuity, sequencing through data validation, deterministic reserve calculation, stochastic scenario generation, exclusion testing and reserve aggregation, with the actuary clearing each stage before the next begins. An IFRS 17 GMM contract grouping exercise for a new product where the actuary signs off on the cohort logic, the CSM mechanics, the coverage unit definition and the unlocking treatment in sequence. A long-term care rate increase filing where experience study selection, assumption update justification, projection methodology and required state-specific filing language each get cleared before the next runs. For P&C, the equivalent is a new reserving segmentation where development factor selection, tail extrapolation and reasonableness testing are staged.

Collaboration—Human-in-the-Flow. These are the cases where the work cannot be done well either by AI alone or human alone. Three examples: pandemic catastrophe scenario design where the actuary calls the regime (which variant, which mortality severity assumptions, which behavioral response patterns) and AI runs the sensitivity sweeps and produces tail metrics. M&A actuarial due diligence on a target block, where the actuary directs which contracts and risks matter, navigates the data room negotiation, and reads the qualitative signals, while AI parallel-processes assumption comparisons, restates reserves under acquirer methodology, and flags inconsistencies across files. Climate scenario analysis on a life portfolio where the actuary calls which warming pathways and regulatory responses are credible and AI runs the longevity and morbidity shifts through the in-force.

A health-specific Collaboration is Medicare Advantage bid strategy, where the actuary owns the competitive intelligence and political read on CMS, while AI runs the bid simulations.

Augmentation—AI-Augmented Human. The defining idea here is a peer review or red team workflow. Three concrete examples: an AI that reviews a draft actuarial opinion or memorandum and checks it against ASOP 41 disclosure requirements, flagging missing sections like reliance on others, qualifications and intended users. An AI that red-teams a pricing memo by generating the strongest plausible counterarguments a product committee might raise, so the actuary can sharpen the document before submission. An AI that scores an experience study justification for an assumption change against precedent and flags whether the rationale is unusually thin compared to past cycles.

A particularly resonant Augmentation example for an internal audit is using AI to evaluate whether a model risk management framework narrative covers all the FS AI RMF controls or NIST AI RMF function categories, and to score the maturity of the description rather than just
whether the section exists.

Human—Human-Only. These examples are situations that should stay human-only, rather than where AI is inconvenient. Three of those: a Section 8 ABCD-style disciplinary determination, where weighing professional conduct evidence is a quintessentially human judgment. The decision about whether to use a novel data source in underwriting (genomic markers, wearable continuous biometrics, social media signals) where the question is not whether the data improves accuracy but whether using it is consistent with policyholder dignity and anti-discrimination obligations. A crisis communication moment, like an appointed actuary deciding how to disclose a discovered material reserve deficiency to the board, where every word carries fiduciary weight.

IN CONCLUSION

We believe that how we structure work with AI is a design choice. AI will not mind which we pick, but our results will differ, and the right choice may be far superior. The trick is to fit the co-working mode to the assignment. That assessment is the actuary’s ongoing responsibility, and we believe it will evolve as AI capabilities advance. Persistent memory, deeper reasoning and continuous training would each redraw the lines we have drawn here. As AI capabilities expand, the requirements for actuaries may expand, too. This framework may help keep the assessment ongoing.

Copyright © 2026 by the Society of Actuaries, Chicago, Illinois.