Why RBC's Call Center AI Cuts 80,000 Hours But Raises an OSFI Flag

RBC has deployed natural language processing to handle routine customer inquiries as part of broader AI efficiency initiatives. BMO followed with similar numbers. The hours themselves aren't the story. What those hours represent is: both banks are using AI to decide who gets contacted about a late payment, when the contact happens, and what repayment plan gets offered. Regulators are watching this decision layer - the targeting logic that determines which borrower gets called, when, and with what offer. Traditional collections waited for a missed payment, then followed a rigid call schedule: 15 days overdue triggers email one, 30 days triggers call one, 60 days triggers call two. AI replaces that timeline with propensity modeling. The system scans transaction patterns, paycheck deposits that stopped arriving on time, credit card spend that spiked before a mortgage payment, recurring charges that overdrew the account twice in three months, and flags accounts likely to default before the borrower realizes they're in trouble. The bank calls at 10 days, not 30, with a term extension offer tailored to that household's cash flow. The loan never technically goes delinquent.

That's the pitch. It works, statistically. Intervention before the 30-day mark keeps the loan out of the "impaired" bucket on the balance sheet, which matters when gross impaired loans across Canada's Big Six banks hit $37.5 billion in 2026 and every basis point of loan loss provision eats into earnings. For the customer, early intervention can mean avoiding the credit score hit that comes with a 60-day late flag. Some banks frame the AI as a financial health assistant, not a collections agent. The customer experience data backs them up on one narrow point: people negotiating with a bot report feeling less judged than people negotiating with a human agent who can hear the stress in their voice.

Where the model breaks

The problem regulators see isn't efficiency. It's legibility. OSFI's model risk management guideline, updated in 2025, requires banks to explain why a specific credit decision was made. A rule-based system could do that: "We called because the loan is 30 days past due." A machine learning model trained on 15 years of transaction data and 400 behavioral variables cannot always produce that explanation in a form a human can verify. The model might flag a postal code, an employment type, or a pattern of weekend ATM withdrawals as predictive of default, but it won't tell you which variable tipped the score or whether that variable is proxying for something the bank isn't allowed to consider, like race or immigration status. That's the black box problem. It doesn't require malice. It just requires complexity.

Digital redlining is the tail risk. If the AI learns that certain neighborhoods or job types default more often, it will weight them more heavily in the score. If those neighborhoods or jobs correlate with protected demographics, the bank is now making credit decisions based on a proxy it didn't design and can't fully audit. The institution might not know until a regulator runs the output data and sees the pattern. By then, thousands of repayment plans have already been offered or withheld.

The talent reallocation story is real but incomplete. Humans didn't disappear; they moved to high-complexity cases where the borrower has irregular income, multiple properties, or a situation the model can't parse. Those agents need judgment. The question isn't whether AI saves time. It does. The question is whether the thing being optimized - speed to contact, probability of repayment within 90 days - aligns with the thing the system is supposed to protect: fair access to credit and transparent decision-making. Speed and fairness sometimes point in opposite directions. The model doesn't notice.