Decisions that slowed the bank down for good reasons

A renewal register that raised our provisions and refused to set prices

In early 2024 our credit reports showed residential mortgage arrears of 0.17%. Beneath that figure sat 51,200 mortgages worth C$10.8 billion renewing in fiscal 2025 and 2026, most of them five-year fixed terms signed in 2020 and 2021. At contractual amortisation, payments would rise by about 19% on average and by more than 30% for one borrower in eight. Arrears would show strain only after households had exhausted their savings.

I proposed a renewal register: the projected payment at renewal, payment as a share of observable income and, for the 58% of mortgage holders whose main chequing account was with us, three behavioural signals from deposit and card data. I proposed two uses, outreach and provisioning, and ruled out a third. Patrick Melanson, our executive vice-president of personal banking, wanted the register to inform renewal pricing, with sharper rates for customers likely to switch lenders. Our privacy office and I argued that pricing on hardship signals exceeded what customers had been told and would penalise the people the register was meant to help. Graham Oakhurst, our chief financial officer, opposed the provisioning use because the signals had never been through a credit cycle.

Our chief executive ruled out pricing and accepted provisioning on two conditions: review by model validation, and two quarters as a reviewed adjustment before the method entered the expected-loss model. We gave up an estimated two points of renewal retention and accepted a higher allowance.

The register identified about 6,800 households under significant strain. Advisers called 4,900 of them six to nine months before maturity, and about 1,620 changed their term, payment date or amortisation within policy, or entered a hardship arrangement. Stage 2 balances rose by C$1.9 billion and the allowance by C$17 million. In June 2026, 90-day arrears in the renewing cohort were 0.29%, against a forecast of 0.55% made before outreach began; I report that difference without claiming all of it.

Renewal strain is now a standing item in the quarterly credit report. What I would do differently concerns the 42% of borrowers whose deposits sit elsewhere. My first design accepted that we could not see them. We should have asked every renewing borrower about changes in income from the first contact, not from mid-2025.

Counting every model before the revised guideline arrived

When OSFI published its revised guideline on model risk management, our inventory listed 88 models. The guideline defines a model broadly, explicitly includes artificial intelligence and machine learning, and expects risk to be managed across a model's life, including its data. I suspected our list described what owners had chosen to register rather than what the bank used. Treasury relied on a prepayment spreadsheet for hedging decisions, marketing ran propensity scores, and operations had begun using a generative AI tool to draft replies to customer letters.

Yolande Boudreau-Haché, our chief risk officer, and I proposed a census rather than an update. Every business unit would list each spreadsheet, vendor score, rules engine and AI tool used to inform a decision, and a joint panel from model risk and my office would classify each one, recording a reason for every item judged not to be a model. Owen Haldane, our treasurer, objected that his prepayment spreadsheet was a set of assumptions, not a model, and that classifying it would put treasury decisions into a months-long validation queue. Our head of marketing argued that propensity scores carried little risk because they only chose whom to contact. Operations asked for the drafting tool to be exempt because staff reviewed every letter.

The census examined 274 candidate items and classified 163 as models, 41 of them high materiality. The prepayment spreadsheet was among them, since it informed hedging on a C$27 billion mortgage book. The drafting tool was classified as a model, which delayed its wider release by seven months while we tested it on French-language correspondence, where it performed noticeably worse. Nine propensity scores were classified as models and fourteen were not, with reasons recorded.

We gave up speed. The validation queue grew from 30 models to 71, and the bank funded six additional validators in the chief risk officer's team at about C$1.4 million a year. My office built the inventory platform, traced every model's data lineage and named a data owner for each high-materiality model. We retired or rebuilt 37 spreadsheets.

The risk committee now receives the census each year, including the 111 items judged not to be models. What I would do differently is settle the definition before counting. We changed our classification rules after the first ninety items and had to review them again, which cost six weeks.

Delaying a small-business credit model to fix how we matched French names

In 2023 my team completed a gradient-boosted model for small-business lines of credit and term loans up to C$250,000, to replace a scorecard in use since 2014. Launch was planned for September. Before release I required fairness testing: approval rates at similar observed risk, compared by province, sector, owner's gender where known and language of service. Language was never a model input.

The tests showed French-preferring applicants in rural Québec, northern Ontario and New Brunswick approved about 11 percentage points less often than comparable English-preferring applicants. The cause was upstream of the model. Our matching routine failed to link 7.9% of French-preferring applicants to their credit bureau files, against 2.2% of English-preferring ones, because it treated accented letters as different characters, split compound surnames at the hyphen, ignored older dit name forms and parsed French addresses poorly. Unmatched applicants received default values and looked like thin files.

Sandra Kilbride, our executive vice-president of business banking, argued for launching on time and sending unmatched applications to manual review. It was a reasonable proposal. I recommended a five-month delay after we reran three years of scorecard decisions and found the same gap in the old process, and after we found that manual review rates for unmatched files varied widely by region. Our chief executive accepted the delay.

We gave up five months of faster decisions. Some applicants went elsewhere, and business banking estimated about C$31 million of lending that might have been approved sooner. We rebuilt matching with accent-insensitive comparison, compound-name handling, a table of name variants compiled with francophone branch staff, and French address parsing.

The model launched in February 2024. Match failures fell to 1.4% for French-preferring and 1.6% for English-preferring applicants, and the approval gap at similar risk narrowed to about 2 points, which we monitor monthly. Most applications are now decided in under two days rather than six, and early delinquency has tracked the validation estimate. We wrote to 312 applicants declined under the old scorecard whose files had failed to match and offered a review; 87 were approved.

Fairness testing at Aubreuil now begins with record matching by language. What I would do differently is ask the francophone branch staff first. Several had known for years that certain names never matched.

Asking customers their language instead of guessing it

In late 2024 the first renewal letters prompted by our register went out in the language recorded for each customer. For 26% of active customers that language had never been confirmed. It had been inferred, under a rule adopted in 2009, from the customer's surname and home branch. Within a month we had 140 complaints from French-speaking borrowers who had received an English letter about one of their largest financial decisions. Over the previous year the bank had logged 1,930 complaints about communications in the wrong language. When I asked the team to run my own record through the rule, it placed me in English.

I proposed abolishing the inference. Every customer without a confirmed language would be asked at the next contact, and through a campaign, which language they wanted, and would receive bilingual communications until they answered. Our head of marketing opposed the change: the inference was right most of the time, bilingual mailings were longer and dearer, and response rates would fall when offers arrived in two languages. Our chief compliance officer supported the change and asked that we also separate language of service from language of documents.

We gave up a rule that was cheap and usually right. Bilingual printing and longer documents added about C$0.7 million a year, and marketing response rates among unconfirmed customers fell by roughly a fifth.

Between January and November 2025, confirmed language rose from 74% to 96% of active customers. Once customers answered, the inferred language proved wrong for about 61,000 of them, in both directions: French speakers with English surnames, and English speakers attached to francophone branches. By the end of fiscal 2025, complaints about wrong-language communications were running at an annual rate of 410. We now record language of documents separately, because about 4% of customers want to speak with us in French and receive mortgage documents in English for a lawyer or co-borrower. Language of service is excluded from every model as an input and used to test outcomes.

What changed most is that our fairness tests now rest on what customers told us rather than on a guess. My regret is how long I took to look. I had been chief data officer for three years before I examined a rule that touched every customer, and it took a complaint about a renewal letter to make me look.