Every province's customers deserve Quebec's automated-decision notice

I think Canadian insurers should give customers outside Quebec the notice that Quebec's privacy law already obliges us to give customers inside it. When a renewal price, a referral or a decline is produced by automated processing without a person deciding, the customer should be told, should be able to learn the main factors behind it and should be able to put their view to someone who can change the outcome. Haverlock has done this in all five of our provinces since March 2025. I want to explain why, what it cost and what it did to the way we build models.

What Quebec asks, and what the rest of us skip

Quebec amended its private-sector privacy law in stages between 2022 and 2024. One provision, in force since September 2023, requires a business that uses personal data to make a decision based exclusively on automated processing to inform the person concerned. On request, the business must explain the personal data used, the reasons and principal factors behind the decision, and the person's right to have that data corrected, and it must give them a chance to submit observations to someone able to review the decision. I am describing the obligation in general terms; our privacy counsel reads the text more closely than I do. Its intent is clear enough.

Federal private-sector privacy law, which governs most of our customers outside Quebec, is older and says far less about automated decisions. Proposals to modernise it have come and gone in Parliament. An insurer can therefore run the same rating engine in Ontario and Quebec and owe customers in the two provinces quite different explanations of the same kind of decision. Many insurers accept exactly that, usually by treating the Quebec notice as a compliance exception produced by a separate document process.

Insurers also argue, with some force, that few of their decisions are exclusively automated. A broker presents the price. An underwriter can override it. Auto rates in most provinces where private insurers write the business are filed with a regulator, and in several are approved by one. All of that is true, and none of it changes what the customer experiences: a number produced by a model they cannot see, from data they may not know we hold.

A notice changes the model before it reaches the customer

We set three triggers. A customer receives a notice when an automated decision declines them in our direct channel, refers their application for additional underwriting, or raises their renewal premium by more than 15 per cent for reasons other than a filed change to base rates. The notice says that a model was used, names in plain language the three factors that contributed most, and gives a telephone number and a web form for asking a person to review it. In the first twelve months about 2.1 per cent of renewals triggered a notice and 6,400 customers asked for a review. We changed the outcome in about 900 cases, most often because our data was wrong: a claim attributed to the wrong driver, a vehicle recorded with the wrong trim, a roof replaced years ago and never updated.

What I did not anticipate was the effect upstream. Once every model that could trigger a notice had to produce its main factors in words a customer would accept, my data scientists began asking that question at design time. Two variables in our home model, both derived from aggregated third-party data about the neighbourhood, could not be described in any sentence we were prepared to send to a customer. We removed them. The model lost a little predictive power, which our pricing actuaries measured at about a tenth of a loss ratio point. I think that was a fair price for a model we can explain.

The review requests also became a data quality programme we had not budgeted for. Nine hundred corrected outcomes meant nine hundred records that had been wrong in the systems feeding our models. Every successful review is now routed to the owner of the field involved, and error rates on driver attribution and vehicle trim have both fallen by more than half since the notices began.

The objections, taken seriously

The first objection is cost. The programme costs about C$1.9 million a year, mostly for eleven reviewers and the systems work that generates factor explanations. For an insurer of our size that is not trivial, and competitors who give only the minimum notice do not bear it. Our chief financial officer asked whether we were paying for a principle. We were, partly. We were also paying for a stream of corrected data that is hard to obtain any other way.

The second objection is gaming. If customers know which factors move their price, some will misstate them. I take this seriously, but the factors we disclose are ones we can verify, such as claims history, vehicle and location, and we do not disclose their weights. Our fraud controls continue to apply. In the first year we found no pattern of misstatement linked to the notices, though I would not claim one year settles the question.

The third objection is that brokers, not insurers, own the customer conversation. Several of our broker partners were uneasy about a notice arriving from Haverlock about a price they had presented. We changed the design so that the broker receives the notice first and may deliver it within five working days, and most now do. The broker relationship survived the change, and the customer gained an explanation.

The fourth objection is that the law does not require any of this outside Quebec. That is correct. I would only observe that privacy law in Canada has moved in one direction for a decade, and that an insurer which builds its models to explain themselves now will not have to rebuild them in a hurry later. Machine learning makes the point sharper. A gradient-boosted pricing component is harder to explain than a rating table, and the time to find out whether it can be explained is before it is filed, not after a complaint.

None of this replaces the regulators who review our rates, our complaints process or the ombudsman services available to customers. It sits alongside them and reaches people those routes do not, because most customers who receive an unexplained increase never complain; they simply leave or pay.

I would put one question to any insurer's board. If a customer in Ontario asked why her renewal had risen by a fifth, could the company tell her which data drove it, and could a person change the answer if the data were wrong? If the honest answer is that only a customer in Quebec would receive that explanation, the board should ask why a postal address decides whether someone deserves one.