Issue 3 · Regulatory forcing
Telematics Is a Rating Variable Before It Is a Customer Feature
A telematics app can improve engagement and collect useful driving data, but its insurance consequence is a class-plan change with credibility, proxy, and approval implications. Product teams should not discover that after launch.
One system
auto insurance
The other
risk classification
The telematics product and the auto class plan meet when a driving observation becomes a premium relativity, eligibility rule, or coverage decision.
The telemetry feature becomes insurance technology at the moment its data changes who pays what, not when the customer installs it.
An auto telematics app has an easy first story.
It gives a driver a trip score.
It sends a reminder about hard braking.
It creates a reason to open our app after purchase, which is a rare and useful event.
That is a customer feature, and we should build it.
It becomes insurance technology at a more consequential moment. When a score, a kilometre count, or the hour of a trip changes who pays what.
The telemetry feature becomes a rating variable before it becomes an engagement strategy.
Once it moves a premium, it stops being an interface question. It is a change to the class plan: the structure of variables, levels, and relativities we use to turn exposure into price.
The app team made the collection mechanism pleasant.
We still have to explain why the variable belongs in the rating system, what it may do, and where it may not be used.
None of which argues against telematics.
It argues for starting at the correct end of the problem.
Start with the coverage, not the trip
Auto insurance is several different promises gathered into an object customers call car insurance.
Third-party liability is compulsory everywhere we write.
Ontario’s statutory minimum is $200,000.
Quebec’s is $50,000, because bodily injury sits outside the private policy under its public plan.
Collision and comprehensive are different again.
Neither is statutorily compulsory, though a lender or lessor will insist.
Accident benefits differ a third time.
They are separately compulsory in most provinces, regardless of fault.
Ontario makes the distinction unusually visible to us.
A rear-end collision creates an accident-benefits file and a property-damage file under one claim number. Its Fault Determination Rules assign fault from specified collision diagrams. The no-fault benefits channel does not erase fault, and it does not settle the tort question. Telematics has to meet those distinctions before anybody decides one good-driving score applies everywhere. Measured distance may give an exposure rationale for collision or comprehensive.
A behavioural observation may relate to liability frequency.
Neither proposition answers the rules for accident benefits.
Neither changes the public bodily-injury structure in Quebec.
The policy form we issue tells the same provincial story.
Ontario’s OAP 1 is a regulator-approved owner’s policy.
Quebec’s Q.P.F. No. 1 is filed with the AMF. British Columbia, Saskatchewan, and Manitoba run public monopoly insurers for basic auto. Quebec has public bodily injury through the SAAQ, and privately rated property damage. Alberta rates through the AIRB, with a cap on increases for drivers who qualify as good drivers.
There is no Canadian telematics setting for us to pick.
There are provincial decisions with an application attached.
Not a drafting nuisance.
It is our design boundary.
A team that puts a generic score ahead of the coverage and the jurisdiction has done the easy work first.
A rating variable is a filed argument
In Ontario personal auto we file both rates and the risk classification system with FSRA. A factor absent from the approved filing is not a useful data point awaiting a release date.
It is outside the rating logic we filed.
The existing plan is already more particular than a dashboard suggests. It can combine years licensed, at-fault claims over six years, convictions over three, territory, vehicle rate group, annual kilometres, commute distance, and driver-to-vehicle assignment. CLEAR supplies vehicle rate groups by make, model, and year, across three coverages. Those variables are not a bag to draw from at will.
Their relativities operate together, and we fitted them that way.
A generalised linear model estimates a factor’s contribution while holding the others constant. Poisson or negative binomial for frequency. Gamma for severity. Tweedie for pure premium. So a telematics measure that looks decisive alone can add nothing after mileage, territory, vehicle, and driver history are accounted for.
Which is the unglamorous product requirement.
Define the variable, the level, the relativity, the coverage, the population, and the permitted effect. Then make the application faithfully collect what that argument needs.
Not the other way round.
Ontario gives us the sharpest example.
Auto telematics is permitted as a discount mechanism only.
It may reduce premium.
It can’t surcharge a driver or cause a declination. That asymmetry is not an adoption obstacle for a clever pricing team to route around.
It is the product rule.
A customer can earn an excellent trip score, and we remain prohibited from treating the adverse version as an underwriting weapon.
The consequence is practical.
A reversal control has to exist before the first premium effect. The system has to know which score produced a discount, which class-plan version allowed it, and how to withdraw or correct the effect when the input, the attribution, or the approved rule turns out wrong. A bright interface cannot reconstruct that record once a filing question arrives.
More observation does not remove the statistical limit
Usage-based insurance splits into two propositions.
Pay-as-you-drive replaces a self-reported kilometre band with measured distance.
Pay-how-you-drive makes behaviour the variable.
Hard-braking and rapid-acceleration events normalised per distance. Cornering force. Total distance. The share of driving in late-night hours. Phone handling.
Our collection technology moved.
Smartphone SDKs have largely displaced OBD-II dongles, cutting acquisition cost and creating a driver-identification problem the vehicle-attached device never had.
A phone reports a trip to us with confidence.
The rating system still has to establish whose trip it was before applying a relativity. Distance has an unusually clean rationale, because claim frequency scales roughly with distance driven. A per-kilometre premium is closer to the exposure unit pooling assumes than a per-year premium is.
The other measures ask us for more patience.
A hard-braking count may be useful.
The useful count is not automatically credible as a premium cell.
Pooling makes that restraint unavoidable for us.
For independent losses, the standard deviation of the mean falls with the square root of the count. But a new segment has its own small pool.
Split a cell in half and the variance of each estimated mean roughly doubles.
The classical full-credibility standard for claim frequency needs about a thousand claims.
No number of smartphone events will supply those claims.
In Bühlmann credibility, the weight on a cell’s own experience is n over n plus k.
Where noise within the proposed group outweighs signal between groups, that weight approaches zero.
The actuarial answer is to blend the thin result toward a broader class mean.
Not to give it a persuasive colour scale.
There is a second selection problem we inherit.
An opt-in programme selects on driver type before we observe a single trip.
The enrolled population is not the book we are pricing.
Participation and self-selection can matter more than the predictive power of any variable.
Particularly where the programme can only offer a discount.
Behaviour can still carry somebody else’s risk
We present telematics as an escape from blunt proxies.
It observes distance, time of day, harsh braking, and cornering, rather than inferring driving from an adjacent characteristic.
That improves causality and controllability.
It doesn’t end the proxy question.
Late-night driving can track shift work.
Shift work can track income.
A permitted variable can predict for its own loss-related reason and merely correlate with a prohibited characteristic.
It can also draw its predictive power from that correlation.
The remedies available to us differ.
One can be residualised or reweighted.
The other has to go.
Deleting a prohibited field from a data dictionary solves nothing.
Where a prohibited variable correlates with a permitted one, the permitted coefficient absorbs its content.
With enough variables, a model reconstructs what we never showed it.
Ontario’s prohibition on credit information in auto rating is not defeated because an opaque behavioural score has no field labelled credit.
So proxy testing is an operating function beside the class plan.
Not a fairness statement obtained at procurement.
Colorado’s SB 21-169 requires insurers to test external consumer data sources, and the models built on them, for unfair discrimination.
The standard method imputes protected class from surname and geography, then compares outcome distributions.
The NAIC model bulletin on artificial intelligence requires a written governance programme covering the same ground.
Those are not Ontario filing rules.
They are useful evidence that a rating operation needs recurring testing and a remediation path.
Canadian deployment adds a layer.
Quebec’s privacy requirements need primary-source confirmation, including the consent and information path, before anything goes into production.
Nobody should treat the app’s permission screen as a completed privacy design.
Five questions to carry upstairs
The board hears about installations, engagement, and perhaps the share of trips scored.
Those are service measures we chose.
They don’t establish that we have a defensible rating variable.
Better preparation is a file answering five ordinary questions.
What insurance purpose does each observation serve?
Which coverage and which province does it apply to?
Which class-plan level and relativity turn it into a premium effect?
Which prohibited proxies were tested, and what happens when a test fails?
Can we reverse the effect and explain the record?
That sequence also gives us a sane division of labour.
The product team owns collection quality, driver attribution, consent flow, and the ability to show a result.
Actuarial owns the evidence for the factor and its relativities.
Compliance and legal own permitted use and the filing path.
Operations owns corrections, complaints, and the evidence trail.
Nobody owns the whole consequence by owning the app.
I would not launch telematics as a loyalty feature and discover its insurance character after adoption.
The useful product is the one whose rating consequence is more governed than its interface is polished.
What the issue establishes
- separate liability, accident benefits, collision, and comprehensive coverage before assigning telematics data to a pricing decision
- define pay-as-you-drive and pay-how-you-drive as distinct exposure and behavioural rating variables
- map a telematics factor to an Ontario filed class plan, its levels, relativities, and permitted premium effect
- apply credibility limits and self-selection analysis before creating telematics rating cells
- test behavioural variables for prohibited-factor proxy effects and retain a remediation path
- preserve Ontario's discount-only telematics control and account for provincial auto-regime differences
The unfashionable recommendation
Do not launch telematics as a loyalty feature and backfill governance later; treat data collection, class-plan rationale, consent, and proxy testing as the product work.
Asked in the meeting
- When does a telematics app become an insurance rating tool?
- A telematics app becomes an insurance rating tool when its observations change a premium, an eligibility result, a declination, or a coverage-related decision. Installation, trip display, and driver coaching do not by themselves alter the class plan. Distance, late-night driving, harsh braking, cornering, or phone handling become insurance variables when the insurer assigns levels and relativities or another approved decision effect to them.
- What is the difference between pay-as-you-drive and pay-how-you-drive?
- Distance-based usage rating replaces a self-reported annual-kilometre band with measured exposure, usually distance driven. Behaviour-based usage rating makes observed behaviour the variable, using measures such as hard braking, rapid acceleration, cornering force, late-night driving, and phone handling. The first changes the exposure measure; the second introduces behavioural classifications that require separate credibility, proxy, and approval analysis.
- What can an Ontario auto telematics program do to a premium?
- Ontario permits auto telematics as a discount mechanism only. Driving data may reduce a premium, but it may not surcharge a driver or cause a declination. Ontario insurers file both rates and risk classification systems with FSRA, so the program must preserve the approved class-plan logic rather than treat a behavioural score as unrestricted underwriting authority.
- Why is a telematics score not enough to set an auto premium?
- A score is an input or a calculation; a class plan is the filed structure of rating variables, their levels, and their relativities. In Ontario, the plan typically combines years licensed, at-fault claims in the past six years, convictions in the past three years, territory, vehicle rate group, annual kilometres, commute distance, and driver-to-vehicle assignment. A telematics factor needs an equally legible role within that structure.
- Which auto coverages need to stay separate when telematics is introduced?
- Third-party liability, accident benefits, collision, and comprehensive have different purposes and sources of obligation. Liability insurance is compulsory in every Canadian province, while collision and comprehensive are not statutorily compulsory and are often required by a lender or lessor. Accident benefits are separately compulsory in most Canadian provinces regardless of fault. A behavioural measure cannot be assumed to have the same rationale across those coverages.
- Why cannot an opt-in telematics sample be treated as the whole auto book?
- An opt-in telematics program selects on driver type before the insurer analyses the observations. Participation rates and self-selection therefore affect the experience being measured: the enrolled population is not automatically representative of the book being priced. A change in score distribution may reflect who enrolled, not a change in driving risk. Credibility and deployment testing need to account for that selection.
- How much experience does a new telematics rating cell need?
- A rating cell needs credible experience, not merely many collected signals. Under the classical full-credibility standard for claim frequency, results within plus or minus 5 percent of expected value with 90 percent probability require about 1,082 claims. Splitting a cell raises the variance of its estimated mean, so thin telematics cells are blended toward a broader class mean or prior indication.
- Can a driving behaviour variable still be a prohibited proxy?
- Yes. Removing a prohibited characteristic from model inputs does not prevent a permitted variable from carrying its predictive content. A behavioural measure that predicts losses for its own reasons but merely correlates with a protected trait calls for a different remedy from one whose predictive power derives from that correlation. The first may be residualised or reweighted; the second may need removal.
- What does a smartphone SDK change in a telematics program?
- Smartphone SDKs have largely replaced OBD-II dongles for telematics collection. They lower acquisition cost and gather measures including distance, hard braking, rapid acceleration, cornering, late-night driving, and phone handling. They also create a driver-identification problem that the vehicle-attached dongle did not have. A rating decision therefore needs a control for whose behaviour is being attributed to the policy.
- Why does mileage make a better insurance variable than a loyalty signal?
- Claim frequency scales roughly with distance driven, so a per-kilometre premium is closer to the exposure unit assumed by pooling than an annual premium detached from measured use. That does not make every measured kilometre a permissible rating factor in every province. It establishes the loss rationale that still has to be tested for credibility, proxy effects, legality, filing, and approval.
- How do provincial auto systems change a telematics rollout?
- Provincial auto systems do not share one rating rule. Ontario uses a filed risk-classification system and permits telematics discounts only. British Columbia, Saskatchewan, and Manitoba have public monopoly insurers for basic auto coverage. Quebec places bodily injury with the public SAAQ while property damage is privately rated. Alberta rates through the AIRB, with a cap on increases for qualifying good drivers.
- What should management have before telematics affects a premium?
- Management needs the variable's intended insurance use, the coverage to which it applies, its class-plan level and relativity, the evidence of credibility, the prohibited-proxy test, the consent and data-control path, the applicable filing status, and a reversal process. The board question is whether those records can explain a premium effect before the feature is released, not whether the app has a high installation rate.