If AI runs the models and drafts the memos, is the actuary the math or the signed accountability for a reserve estimate?

The shift

AI makes model construction and report drafting abundant — it can build the projection, run the sensitivities, pull the assumptions from experience studies, and write the reserve memo at near-zero marginal cost. What it can't do is own the number: sign the statement of actuarial opinion, carry the professional liability, or answer to a regulator when the reserve proves inadequate.

The axioms

• Building the model and running the numbers is the scarce skill the actuary sells. Rests on: the labor of assembling data, coding the projection, and reconciling exhibits being slow and expensive (scarce analyst-hours).
• The credential (FSA/FCAS) certifies someone who can do the technical work. Rests on: technical competence being hard to acquire and otherwise unverifiable — the exam sequence was the proxy.
• A credentialed actuary must sign the opinion and is personally accountable for the estimate. Rests on: accountability — a named, disciplinable person the regulator, board, and courts can hold to the number.
• Setting the assumptions under novel conditions requires human judgment. Rests on: judgment where there's no clean precedent to match — a new product, a regime shift in mortality or claims, a stress no experience study covers.
• Regulators and boards require a human sign-off by statute and standard. Rests on: legal/institutional recognition of the credential, not a capability the model lacks.
• Juniors become credentialed signers by grinding exhibits, tie-outs, and reconciliations for years. Rests on: repetition being the only path to internalized judgment — practice was scarce time, not scarce information.

Invalid axioms

1. Building the model and producing the exhibits is the scarce work the actuary is paid for. AI assembles the data, codes the projection, runs the sensitivities, and drafts the memo faster than a team of analysts and without fatigue. The habit-trap: firms still scope and price reserve and pricing work by analyst-hours spent constructing models, and staff pyramids sized to produce those hours, when the construction is now close to free.
2. The exam sequence certifies value because it proves you can do the technical computation. A model can now do — and increasingly pass — large parts of the technical, formula-driven content. The habit-trap: treating exam progression through the mechanical syllabus as the main signal of a valuable actuary, rather than the ability to judge assumptions and defend them, which the exams test far more weakly.

Unchanged axioms

1. Someone credentialed must sign the opinion and carry the liability for the estimate. A model cannot sign a statement of actuarial opinion, cannot be disciplined by the profession, cannot be sued, cannot lose the credential. The reserve is a number a regulator and board rely on precisely because a person answerable for it stands behind it. The actuary's economic function moves from producing the estimate to being the accountable party who reviewed and owns it — and that accountability didn't get cheaper.
2. Judgment on assumptions under novel conditions still needs a human. The hard calls are the assumptions, not the arithmetic: what mortality or lapse or loss-development pattern holds for a product with no credible history, whether the past regime still applies after a structural break, how much margin the uncertainty warrants. These are exactly the cases where there's no clean pattern to match, and where a plausible, well-drafted answer can be confidently wrong in a way that only surfaces years later when reserves run off. The model proposes a defensible-sounding assumption set; it doesn't know which set survives a regulator's challenge or an adverse-development review.
3. Regulatory and professional sign-off is a legal gate that doesn't dissolve on its own. Statutory reserve opinions, rate filings, and appointed-actuary requirements name a credentialed human by law and standard of practice. This is institutional scarcity, not a capability gap — it won't move at the pace of the models even as the technical skill gap closes, and it's the load-bearing reason the credential retains value.

New axioms

1. Where do future signers come from once the exhibit-grind that grew them is automated? Actuaries learned to smell a bad assumption by building the model by hand, watching reserves develop against their estimates, and being wrong enough times to calibrate. If AI does the grinding, the ladder that produced judgment collapses — open problem: how do you develop the assumption-setting instinct in someone who never had to reconcile the exhibits or own a projection that missed. The signers retiring in fifteen years are the last cohort trained the old way.
2. Who catches a confidently wrong AI-built reserve model before it reaches a filing? AI-generated projections, assumption sets, and memos arrive at a volume no appointed actuary can re-derive from scratch, but review and challenge capacity is fixed. The bottleneck moves from building the model to verifying it and interrogating its assumptions — open problem: the sign-off assumes a human understood the model well enough to defend it, and that understanding is exactly what's eroded when the human didn't build it.
3. What should the credential test for once the computation is commoditized? The exam sequence was built when technical competence was the scarce, testable thing. If that's abundant, the profession has to test and certify judgment-under-uncertainty and the ability to defend assumptions directly — harder to standardize and grade — open problem: nobody has rebuilt the syllabus around that yet.

Where it breaks

Firms automate the exhibit-and-model grind for margin (INVALID #1) while still needing that same grind to grow the actuaries who can eventually sign (NEW #1). Cut the training ground and bank the efficiency in the same motion, and you get a judgment shortage that only becomes visible when the current signers retire and the bench can build models it can't independently judge.

The appointed actuary's signature (STILL HOLDS #1) is meant to certify that a competent human reviewed and owns the estimate — but when the model is AI-built at a volume no one can re-derive (NEW #2), the sign-off risks degrading into a stamp on machine output the signer didn't construct and can't fully interrogate. The liability stays real; the understanding that was supposed to justify it thins.

ai-axioms.vercel.app/axioms/finance/actuary-ai-axioms