AI-Informed Policyholders Are Drifting from RILA Behavior Assumptions Faster Than VM-21 Can Track

RILA sales hit $21.2 billion in Q1 2026, up 21% year over year, as total annuity sales cleared $100 billion for the 10th consecutive quarter. AI retirement planning tools are now teaching policyholders to optimize buffer utilization and surrender timing against their specific contract terms. Every behavioral dataset behind VM-21 reserves and RILA pricing was gathered before those tools reached the retirement-age market.

Key Takeaways

The SOA behavior study covers 10.5 million contracts and $1.4 trillion of contract value across 2019 to 2021, which is substantial as data and entirely pre-AI as a baseline.
Milliman's first RILA experience study found MYGA-like durational behavior: surrender rates stay very low through the charge period, then spike sharply at expiry.
Historical mid-term reallocation participation runs at 15% to 20% of eligible policyholders. That inertia is the parameter an optimizer removes.
FIA contracts credited well below market rates surrender at more than three times the rate of contracts near market, so the amplifier already exists in fixed products.
Delta-gamma hedge programs are calibrated on mean reversion toward inattention, and the alerts that break the inattention fire in the same market conditions that stress the hedge book.

The Calibration Predates the Tool

The SOA 2019-2021 Variable Annuity Contract Owner Behavior Experience Study is where most VM-21 behavior assumptions come from. It covers roughly 10.5 million contracts exposed to surrender, $1.4 trillion of aggregate contract value, more than 500,000 surrenders, and 3.7 million contracts with withdrawal activity totalling $41 billion withdrawn.

Milliman published the first RILA-specific experience study in June 2025, giving the product class a calibrated baseline for the first time. Surrender patterns resemble multi-year guaranteed annuities more than traditional variable annuities: rates stay extremely low through the surrender charge period and spike at expiry. Channel differences were pronounced, with bank-channel contracts showing elevated surrender in the window immediately after charges end, and large national broker-dealer contracts running below the independent broker-dealer channel. The study covered contracts without guaranteed lifetime withdrawal benefit riders.

Neither study can address the counterfactual, because the behavior it describes did not exist during the observation windows. AI-powered personal finance tools, including retirement-specific platforms running Monte Carlo projections against policy-specific parameters and Social Security claiming optimization, reached mass-market adoption among retirement-age consumers between 2023 and 2025. Every observation in both studies came from policyholders deciding on advisor guidance, periodic statements, and their own initiative.

Where an Optimizer Moves the Number

A RILA buffer is a legible optimization target. The carrier absorbs the first several points of index loss in a term and the policyholder bears the rest, and most products let the policyholder re-elect among index options and buffer levels at each term anniversary. Pricing sets the expected frequency and size of buffer credit utilization from how policyholders have historically managed those elections, and the historical answer is passive: pick at issue, leave it.

Take a product with four index options and two buffer levels at 10% and 20%. An application with the contract data monitors each index against the entry level and the buffer threshold and produces a recommendation at the next election window, toward the deeper buffer on a segment that has already absorbed partial losses or toward more upside on one running strongly positive. It does not have to be well calibrated to change the aggregate. It only has to reduce the inertia that keeps reallocation participation at 15% to 20% of eligible policyholders.

The pricing consequence follows from a small move in a large parameter. If a block's expected buffer absorption at the 10% level runs near 4% of policy-terms a year under passive behavior, and optimized allocation among 25% to 30% of the block lifts that to 5.5% to 6%, the hedge budget embedded in pricing is understated on a line growing 21% year over year.

RILA Behavioral Assumption Dimensions: Historical Baseline vs. AI-Engaged Pattern
Behavior Dimension	Historical Calibration	AI-Engaged Pattern	Primary Source
Mid-term reallocation frequency	15% to 20% per election window	30% to 45% for AI-integrated cohort	SOA 2019-2021 VA Study
Buffer utilization rate (10% buffer)	~4% of terms per year	5.5% to 6%+ for AI-optimized allocation	actuary.info analysis
Post-SC lapse concentration window	Spread across 2 years post-SC expiry	Concentrated 60-90 days around SC expiry	Milliman RILA Study (June 2025)
Rate-differential lapse sensitivity	3x rate when credited rates far below market	Additional multiplier from AI prompt at optimal timing	Milliman FIA Study (January 2025)

Surrender timing sharpens the same effect. Charges typically run six to seven years, and models calibrated on Milliman's RILA study anticipate voluntary exits concentrating in the first two years after they end. An application with contract data can identify, 60 to 90 days ahead of the expiry date, how the accumulated index credit, the remaining charge, and available external rates combine, and prompt at the optimal moment.

That is not a sudden behavioral break. It is a systematic tightening of the surrender distribution toward the moments most favorable to the policyholder and least favorable to the carrier. The amplifier is already visible in fixed products: Milliman's FIA studies through Q1 2024 documented surrender rates more than three times as high on contracts credited well below current market rates. Channel is the usable proxy for exposure, since RILA sold through advisor platforms with integrated planning tools reaches a cohort far more likely to have contract data feeding an optimizer than the career agent channel does.

The Hedge Assumes Inattention

Delta-gamma programs hedging RILA guarantees calibrate partly on behavior models that assume reversion toward a passive baseline, and the assumption has been well grounded. Policyholders are inattentive by default, they do not act on every signal, and because they do not all respond at once, aggregate demand on the hedge book in any short window is lower than a fully rational model would produce.

The alerts remove that dampening exactly where it was doing the most work. An application generating recommendations tied to index performance against buffer levels lifts participation in election windows precisely when index moves are large enough to trigger the alerts, and those windows cluster around the market conditions that already stress the hedge book. Participation running at 30% to 45% in triggered windows against a 15% to 20% historical baseline leaves the book structurally underexposed at the worst moment.

A one-time hedge ratio adjustment does not fix it, because the distribution itself is moving. Tool penetration among retirement-age policyholders is still deepening and carrier data integrations are getting more precise, so a program recalibrated to today's participation rate faces a different distribution in two years. The behavior input belongs alongside the volatility surface and yield curve as a parameter reviewed on a cycle, not a constant set at product launch.

The same bias reaches the reserve. VM-21 takes the conditional tail expectation at the 70th percentile across thousands of scenarios, with lapse, partial withdrawal, and reallocation feeding each one, and it inherits its behavior parameters from a pre-AI dataset. The SOA Research Institute's own 2025 work on AI in investment and retirement anticipated the mechanism, noting the tools may change how advice, allocation, and planning are delivered and create risks retirees could not have been prepared for.

No credible segmentation data yet exists to correct for it. The tools reached scale between 2023 and 2025, and neither the June 2025 RILA study nor any other source can separate behavior by tool engagement. The optimism sits in the tail, which is where VM-21 capital is most sensitive, on the fastest-growing block in the annuity market.

AI-Informed Policyholders Are Drifting from RILA Behavior Assumptions Faster Than VM-21 Can Track

Key Takeaways

The Calibration Predates the Tool

Where an Optimizer Moves the Number

The Hedge Assumes Inattention

Further Reading

Sources