In 2012, someone showed me a life insurance policy still in force. Issued in 1928. I'd come to the sector after a long run in credit cards, where you know what you're carrying within six to 18 months. I did the math without meaning to: that policy predated Social Security.
Insurance carries risk differently, and the difference isn't just duration — it's what either side can do about it once the ink is dry. A lender can act on an individual account: reduce a line, renegotiate, price it differently, charge it off. A life insurer cannot unilaterally touch a policy already in force; that right belongs to the policyholder, not the carrier. The industry doesn't just carry decades-long risk. It carries it locked, one contract at a time, with no lever to revisit terms even after the assumptions underneath them have quietly stopped matching reality.
That's an issue the industry's own machinery was never built to answer. Reserving, asset adequacy testing, hedging, reinsurance, the sale of legacy blocks to run-off specialists — all of it real, all of it sophisticated, all of it built to answer whether the aggregate is still solvent. None of it answers a narrower, harder question: does this one assumption, buried in a block that's still passing in aggregate, still hold? The standards touching assumption review ask for judgment, triggered by whatever surfaces the need — not a date on a calendar. The tests that do run on a schedule check the aggregate, not the assumption underneath it.
Here's the tell that this gap stopped being theoretical
In March 2026, the NAIC began piloting a structured AI Systems Evaluation Tool across 12 states. It isn't a general AI-governance gesture. Examiners using it can ask insurers to detail the oversight of AI systems used by actuarial service providers specifically, and to walk through the testing, verification, frequency, and methodology behind them. For actuaries who build or validate predictive models — which increasingly means experience studies and assumption-setting, not just pricing or claims — that means model documentation and assumption justifications becoming subject to regulatory examination, not just internal actuarial standards of practice. That's not an analogy to a proactive posture. It's regulators reaching directly into how AI-touched assumptions get tested and justified — a first, concrete step toward exactly the granular scrutiny this piece is arguing the industry doesn't otherwise have.
(A brief aside: the industry has pushed back hard on this pilot, and not merely over workload. Trade groups' formal comments call it "voluntary for regulators while compulsory for companies," note it has no defined end date, and flag that findings from what's explicitly an exploratory phase could be held against a carrier before the tool itself is finalized. Those are fair governance objections. They're also, in a way, evidence for this piece's own argument: regulators are building this proactive posture iteratively, through judgment and a pilot, before hardening it into a fixed rule — which is closer to the muscle insurers themselves need than a reason to dismiss it.)
That same year — 2012 — I watched a version of this exposure play out in real time, from inside a different seat, in an industry-wide effort that was already well underway. The 2008 crisis had left insurers holding large in-force books of annuities carrying guaranteed crediting or withdrawal rates set in a much higher-rate era. When rates fell and stayed down for years, those guarantees didn't just look generous — they became a standing liability the original pricing had never been stress-tested against. And because nobody can unilaterally rewrite a policy already in force, the fix was never going to be a model update. It was years of patient, unglamorous, one-by-one work: building offers attractive enough that policyholders would voluntarily give up guarantees the industry could no longer comfortably afford to keep. Carrier by carrier, contract by contract, across the sector — not because anyone lacked talent, but because nobody had a standing practice of asking, before the rate environment turned, whether that guarantee still made sense.
That's the exposure insurance carries now, in its own way, with its own kind of contract. It isn't new, and it was never really about AI — the gap between "the aggregate is fine" and "the individual assumption still holds" predates any model running against an underwriting file. What's changing is the cost of not knowing. Checking assumptions individually, continuously, used to be a resourcing decision nobody could justify for a block that was passing anyway. That constraint is loosening — not because the risk changed, but because the tools that make granular, continuing review economically realistic now exist. "We check the aggregate and trust judgment to catch the rest" isn't gone as an answer. It's weaker than it was five years ago. And that industry-wide effort is a preview of something worth sitting with: the fix that took years of manual, one-by-one work only got built because the rate environment forced it. It's a much better position to build that habit before something forces it than during.
What this actually asks of insurance executives
I'm not offering a view on hedging strategy or reserve methodology — that's not my lane. What is squarely in my lane: whether the organizational conditions exist for the people who already own assumption review to work well with AI — not defer to it, not ignore it, but genuinely partner with it.
- The regulator's own posture is worth studying less for compliance value and more as a design pattern — and it's only become buildable as standing practice, rather than crisis response, recently. Assumption review has always had owners. What it hasn't had is a way to run continuously rather than annually or on trigger. That's what AI now makes affordable — not a replacement for an actuary's judgment, but a way to test a guarantee, a lapse assumption, a legacy block, far more often than a person could manage alone.
- The harder question is whether the talent is being built for that partnership. Treating an AI system's output as the answer is one failure mode. Dismissing what it surfaces because it doesn't match what the model has always said is the other. The judgment in between — knowing when to challenge an AI finding, when to trust it, when to go dig further — isn't automatic. It has to be cultivated, the same way any other analytical judgment is.
Picture your next executive meeting on NAIC's AI Systems Evaluation Tool — whenever the Fall National Meeting takes it up, this becomes an agenda item somewhere. The easy question in that room will be whether you're ready for the exam. The harder, better question is whether the people who own your assumption reviews are equipped to work alongside AI — able to use what it surfaces, and just as able to push back on it when something doesn't add up.
The guaranteed-rate annuity problem wasn't solved by finding more talent — the industry had no shortage of it. It was solved reactively, once a crisis forced the question, and it took years longer than it should have because nobody was set up to ask it before the crisis made it unavoidable. That's the real choice in front of insurance leadership now: build the organizational conditions and the talent to ask that question all the time, with AI doing part of the looking and people doing the judging — or wait for the next rate environment, or the next examiner, to force the question again.
