Health Insurance Is Copying Auto Telematics, and Health Has a Harder Problem Than Driving

On 14 August STAT published an argument by two sociologists that health insurers are importing the telematics model from auto insurance: track the behaviour, score it, reward the good and, eventually, charge the bad [1]. They named John Hancock’s Vitality and UnitedHealthcare’s Rewards on the health side, Allstate’s Drivewise, GEICO’s DriveEasy and Progressive’s Snapshot on the auto side, and concluded that the telematics model “is not a health strategy and it is not sound policy” [1].

The comparison is worth taking seriously, because auto insurance has run this experiment for fifteen years and published the results. It is also worth taking further than STAT did, because the health version faces three constraints that auto never had. One of them is the law. One is Apple and Google. One is who owns the sensor.

What auto telematics actually delivered

Usage-based auto insurance has existed at scale since Progressive’s Snapshot went national in 2011. Regulators have started measuring it, and the Maryland Insurance Administration’s July 2025 survey of every auto insurer in the state is the clearest public record [2].

Read those together. After a decade and a half, one driver in eight had opted in. Of those, fewer than one in three saw the saving the programs are sold on, and nearly one in four paid more. The Consumer Federation of America, which pulled the enrolment figure from the same data, notes that insurers advertise discounts “up to” 30% or 40% while most enrolled drivers see no reduction at all [3]. Nationally, TransUnion counted more than 21 million US policyholders sharing telematics data in 2024, a 28% compound growth rate since 2018 and still a small minority of insured drivers [4].

The discount became a surcharge. Snapshot was discount-only for its first four years. Progressive’s 2015 annual report announced that it would begin, “for the first time, increasing rates for a small number of drivers whose driving behavior justifies such rates,” starting in Missouri [5]. Reviews of the program now put the share of participants who see an increase at renewal at about 20% [6]. The CFA lists Allstate, GEICO, Liberty Mutual, Progressive and Travelers as insurers that raise premiums on telematics data, and American Family, Farmers, Nationwide, State Farm and USAA as insurers that say they do not [3]. The pattern is the important part: once enough data existed to price on, pricing on it was too profitable to forgo.

Behaviour did change, modestly, and mostly among the middle. A 2025 naturalistic study gave 86 drivers gamified challenges with €2 and €5 coupons for good scores. Average-risk drivers cut the share of distance driven above the limit from 4.8% to 3.7%, and high-risk drivers reduced speeding intensity from 6.4 to 5.3 km/h over the limit. The cautious group showed no significant change at all, because it had nothing left to improve [7]. A larger US field experiment that simulated usage-based insurance programs reported reductions in speeding, hard braking and rapid acceleration across its treatment groups [8]. Real effects, small in size, concentrated in the people who were neither the best nor the worst.

The data was not always accurate. Smartphone telematics has to decide whether the phone’s owner was driving. A University of Pennsylvania validation of the algorithm class used by leading insurers found 96.5% overall accuracy, but specificity of 91.2% with a standard deviation of 14.8 points: roughly one passenger trip in eleven scored as if the policyholder had driven it, with wide variation between people [9].

And then the data leaked into pricing through the back door. The New York Times reported in 2024 that General Motors’ OnStar Smart Driver program had been selling driving-behaviour data to LexisNexis and Verisk, which passed it to insurers, without meaningful consent. The FTC brought an action in January 2025 and finalised its order in January 2026: a five-year ban on disclosing geolocation and driving-behaviour data to consumer reporting agencies and two decades of consent requirements. California added a $12.75 million settlement and a deletion order [10]. The thing STAT worries about for health, insurers profiting from data sold behind the consumer’s back, has already happened in auto, and it took a federal regulator to stop it.

Six things the auto record predicts

1. Opt-in plateaus. Fifteen years in, telematics is a minority product even where it is heavily marketed. The people who enrol are the ones who expect to win.
2. The program sorts more than it saves. A discount that goes to people who were already safe is a repricing of risk, not a reduction in it. The Maryland distribution, with more policies unchanged than cut, is what sorting looks like.
3. Discount-only does not last. Progressive held the line for four years. Once the data supports a surcharge, the surcharge arrives, and the marketing is rewritten around it.
4. Behaviour change is real, small and middle-heavy. The best participants have nothing to improve and the worst tend not to enrol. The gains come from the median.
5. Data quality becomes a dispute. When a score has a price attached, misclassified trips become complaints, and the regulator starts asking how the score is built. Maryland’s report records exactly that: policyholders unable to learn how their rate was calculated, insurers citing proprietary models [2].
6. A privacy event resets the rules. GM sold the data; the FTC wrote a five-year ban. Health data has stricter statutes and a larger constituency.

Every one of these has a direct analogue in a wearable rewards program. The health version also has problems auto never had.

The health programs, as they exist

John Hancock sold its first interactive life policy in 2015 and, on 19 September 2018, stopped selling anything else: every policy since comes with Vitality, the behaviour-change platform built by South Africa’s Discovery [11]. Vitality PLUS members can obtain an Apple Watch Series 11 or Ultra 3 for $25 plus tax, with the remaining cost spread over 24 monthly payments that shrink with recorded activity [12].

UnitedHealthcare Rewards, launched February 2023, pays up to $1,000 a year, where the employer buys that option, for goals including 5,000 steps a day, 15 minutes of activity, 14 nights of tracked sleep, a biometric screening and a health survey, using the member’s own tracker, smartwatch or phone [13]. Its predecessor, UnitedHealthcare Motion, set a target near 12,000 steps a day and UnitedHealthcare claimed savings of about $222 per member a year in healthcare costs [14].

Aetna’s Attain, built with Apple in 2019 around the same idea, is the cautionary case. Aetna stated at launch that it would not use the data for underwriting, premium or coverage decisions [14], announced in late 2022 that it would sunset the program, and stopped offering it from 1 February 2023 [15].

Medicare Advantage has become the newest channel. Essence Healthcare began providing an Oura Ring and membership at no cost to eligible members in 2025 [16], and Oura’s S-1 names health plans, employers and care providers as its next growth population [17].

The evidence base is Discovery’s. RAND Europe analysed 422,643 Vitality members across the US, UK and South Africa from 2015 to 2018 and found that taking up the Apple Watch benefit was associated with a 34% increase in tracked activity days, roughly 4.8 more days a month. The design that worked best gave the device up front and required repayment if activity lapsed [18]. Discovery reports that its highly engaged life-insurance clients show 73% lower mortality risk and 45% lower morbidity risk than the insured population, and that lapse rates for Gold and Diamond members run up to 67% below non-engaged clients [19].

Those are large numbers and they deserve the caveat Discovery itself would give a regulator: they compare people who chose to engage with people who did not. The RAND study is an association among self-selected members. The mortality and lapse figures describe who engages as much as what engagement does. Which brings us to the first constraint auto never had.

Where the analogy breaks

1. In US health insurance, pricing on the data is illegal

Auto telematics works because the insurer may charge you for what the device sees. Health insurers may not. Under the Affordable Care Act a premium can vary only by age, location, tobacco use, individual versus family enrolment and plan category, and an insurer “can’t take your current health or medical history into account” [20]. There is no lawful path from a step count to a health premium.

What remains is the wellness incentive. For programs that condition a reward on a health outcome, HIPAA and the ACA cap it at 30% of the total cost of coverage, or 50% where tobacco is involved [21]. The EEOC’s parallel limits under the ADA and GINA, which governed programs that ask for health information, were vacated by a federal court effective 1 January 2019 after AARP argued the agency had never justified the 30% figure. The EEOC’s 2021 replacement, which would have allowed only de minimis incentives, was withdrawn. As of 2026 there is no federal ADA or GINA limit at all, only a “voluntary” standard that courts are now testing case by case [22].

So the health copy of telematics is structurally different from the original. Auto reprices the whole book on the data. Health can move money within a capped incentive pool, cannot touch the premium, and operates in a legal fog about how large a “voluntary” incentive may be. That changes the economics entirely: an auto program can fund its discounts from the surcharges and the improved risk selection. A health program has to fund its rewards from somewhere else.

Life insurance is the exception, and it is the market John Hancock is in. Life policies are medically underwritten, so an insurer may in principle use wearable data to price. Regulators noticed early. New York’s Department of Financial Services issued Circular Letter No. 1 in 2019 after finding life insurers using external data with no intuitive connection to health, and in July 2024 adopted a final circular requiring insurers to show that any external data source or model is not a proxy for a protected class and is actuarially justified [23]. RGA, one of the largest life reinsurers, wrote in 2021 that wearable data was “not yet widely or confidently used in underwriting risk assessment” [24]. Five years on, the programs still reward. They do not rate.

2. The platforms decide who may see the data

In auto, the car maker owned the sensor and, in GM’s case, sold what it saw. In health, the sensor is a phone or a device whose data reaches the insurer through Apple or Google, and both have written rules that shape what a program can be.

Apple’s App Store guideline 5.1.3(i) prohibits disclosing HealthKit data to third parties for “advertising, marketing, or other use-based data mining purposes.” It then adds a sentence that reads as if it were written for this market: apps “may, however, use a user’s health or fitness data to provide a benefit directly to that user (such as a reduced insurance premium), provided that the app is submitted by the entity providing the benefit, and the data is not shared with a third party” [25]. An insurer may therefore run a rewards program on iPhone health data, but only through its own app, and it may not receive that data through a wellness vendor, a benefits platform or an aggregator.

Google’s Health API Developer and User Data Policy, effective 24 March 2026, prohibits “transferring or selling user data to third parties like advertising platforms, data brokers, or any information resellers” even if aggregated or anonymised, and prohibits using it “to determine credit-worthiness or for lending purposes.” It permits sharing only to provide features visible to the user, with consent [26]. It does not name insurance, and silence is not permission, as we found when we read six providers’ terms on a different question.

The consequence is architectural. The telematics vendor model, where a specialist collects the data and hands a score to the insurer, is what the MIA found in Maryland: only five insurers collected data directly and eleven contracted it out [3]. On the health side the platforms have largely closed that path for device data. The verification has to happen inside the benefit provider’s own app, or through data the user has expressly routed there.

3. The sensor is owned by the people who were healthy already

An auto telematics device sits in the risk object and measures the behaviour that causes claims. A wearable rewards program measures steps and sleep on a device that 46% of US adults own, and Rock Health describes those owners as younger, wealthier, more urban, healthier and more likely to be commercially insured than the people who do not [27]. Oura’s S-1 puts 63% of its members above $100,000 in household income and more than half with a chronic condition they are already managing [17].

That is selection built into the hardware. Auto telematics has a selection problem because safe drivers opt in. Health telematics has it twice: the healthy opt in, and the healthy own the verifier. We laid out the demographics in the market split and the ownership gap by country in the adoption figures. The short form is that in every market with a national survey, a majority of adults own no wearable, and the minority who do is not the population an insurer most needs to reach.

The measurement itself is also noisier than a car’s. Two devices on the same person disagree about resting heart rate, about step counts, and about which day a night’s sleep belongs to. A telematics score contested at renewal is a nuisance. A wellness reward withheld because a watch and a phone disagree is a member-relations problem with a regulator behind it.

4. The outcome arrives too late to fund the program

Auto claims respond to driving within a policy year, so an insurer can see whether the program paid. Health claims respond to behaviour over years, and the randomised evidence on employer wellness programs is consistent: at BJ’s Wholesale, self-reported exercise and weight management rose but clinical measures, medical spending and utilisation were unchanged at 18 months and again at three years; at the University of Illinois, participants were already healthier and cheaper before the program began, and there was no effect on spending at 12 or 24 months [28][29]. We go through that evidence, and what to measure instead, in the wellness KPI piece.

Which means a health rewards program cannot be funded from claims savings inside the underwriting cycle, because there are none to book. Discovery’s honest business case is the other number: engaged members lapse up to 67% less [19]. The program is a retention and selection product. That is a legitimate thing to build. It is not what “the telematics model” implies, and it is not a health strategy, which is where STAT’s authors and the auto data agree.

What a defensible program looks like

For an insurer, an employer, or a vendor building for either, the auto record and the health constraints point the same way.

Reward, never rate. In health insurance the law settles it. In life insurance the regulators are circling. A program that cannot become a surcharge is also a program members will trust for longer than four years.

Measure selection explicitly. Report the baseline risk of joiners against non-joiners before you report outcomes. Illinois showed what happens when nobody does: a decade of favourable studies that turned out to be measuring who signed up.

Make the verifier universal. If verification requires a wearable, the program is closed to the majority and skewed to the healthy by design. The phone in every member’s pocket records steps, movement and a usable sleep estimate without a device. Accepting it as evidence is the single largest lever against the selection problem.

Respect the platform path. On iOS the benefit provider must own the app and the data must not leave it. On Android the data may not be transferred beyond user-visible features. Design the architecture to those rules from the start; the CFA’s finding that most auto insurers outsourced collection is a design that health platforms have already ruled out.

Publish the distribution. Maryland forced auto insurers to disclose how many premiums rose, fell and held. A health program that publishes its own reward distribution, its participation by income and age, and its verification failure rate will find that regulators ask fewer questions.

Expect the surcharge temptation and rule it out in writing. Progressive lasted four years. The penalty framing is already visible in wellness: RAND found that programs using penalties reported median participation of 73% against 40% for rewards alone [30]. The ACA cap and the vacated EEOC rules mean the legal ceiling is unclear. Decide the ethical one before the actuaries ask.

Where we sit

Sahha supplies the verification layer under several programs of this kind, so we have an interest here and it is worth stating. Our view is that the auto data is the best forecast available, and it forecasts a minority product that sorts risk unless the verifier reaches everyone. That is the reason we read the phone as a sensor alongside every wearable a member chooses to connect: it is the only way a rewards program reaches the 54% of adults without a device, and the only way the participant pool stops being a portrait of who already owned one. It also means we are inside the platform rules described above, not around them, and any program built on our data has to be too.

The short version

Auto telematics is fifteen years old and Maryland’s regulator has counted the results: 13% enrolled, 31% saved, 24% paid more, 45% unchanged, and the largest program now raises rates on a fifth of its participants. Behaviour changed a little, mostly in the middle, and the data was eventually sold behind drivers’ backs until the FTC banned it. Health insurers are copying the mechanism with Apple Watches and Oura Rings, but under the ACA they cannot price on the data, the platforms restrict who may receive it, the verifying device is owned mainly by the already healthy, and randomised trials find no claims savings inside three years. The programs that survive will be the ones that reward rather than rate, verify from the phone everyone carries, and measure their own selection before claiming an effect.