Canada’s major banks have drawn a hard public line against AI agents. Payroll departments need to do the same, says expert
Canada’s major banks have drawn a public line against AI agents accessing sensitive financial data. At the Toronto Global Forum on Oct. 21, Bank of Montreal (BMO) chief executive officer Darryl White warned people against giving financial and account information to AI agents, saying they were “not protected.” A week later, BMO, Royal Bank of Canada, Canadian Imperial Bank of Commerce, and TD Bank all stated they will block AI agents from directly accessing online bank accounts and other financial information, according to CBC News.
With Canada’s big banks strongly indicating that they don’t trust AI agents with sensitive financial information, it raises the question for HR leaders for their own organizations: who is guarding the payroll?
The HR and payroll function sits on some of the most sensitive personal information an organization holds – social insurance numbers, direct deposit banking details, salary records – and the governance frameworks protecting that data from AI agents are, in most Canadian workplaces, far less robust than those inside a financial institution, according to John Wunderlich, a privacy and identity consultant and principal of John Wunderlich & Associates in Toronto.
“In no way, shape, or form should AI be let anywhere near any of the payroll calculations, or any of the things that produce those outputs,” says Wunderlich. “AI is not deterministic. AI is probabilistic.”
Payroll calculations – hours worked, provincial statutory remittances, tax deductions – follow precise, deterministic rules. AI systems produce probabilistic answers. Combining the two in operational payroll processes introduces a category of risk most organizations have not yet confronted, says Wunderlich.
One U.S. survey reported in HRD Canada found that one-third of employees felt uncomfortable with AI calculating their pay, while nearly half in another survey said they would be uncomfortable with AI tools responding to their payroll inquiries.
The bright line AI must not cross
Wunderlich draws a hard governance boundary around the operational core of any payroll function: the systems that calculate pay and disburse funds to employees’ bank accounts. That layer should be entirely off-limits to AI agents, he says.
“Don’t let it near your operational systems that actually calculate and disburse the funds,” says Wunderlich. “Your payroll function has access to your banking system and sends information to your employees’ bank accounts. Don’t let AI anywhere near that.”
There is a role for AI in payroll, says Wunderlich – as a research tool helping professionals scan legislative updates across multiple jurisdictions or flag code vulnerabilities before they reach production. “I think of it as an enthusiastic research assistant that needs close scrutiny,” he says.
Who carries the liability when something goes wrong
Deploying an AI agent does not insulate an organization from accountability for what that agent does, according to Wunderlich. “What you can’t do as an organization is outsource accountability,” he says. “So whatever the agent does, or whatever the agent working for your agent’s agent’s agent does, it all comes back to you.”
The legal principle is vicarious liability. If an AI agent causes harm – a data breach, a wrongful payment, a privacy violation – the employer bears responsibility. The Air Canada chatbot case, in which a court found the airline liable for incorrect information provided by its AI agent, is now a widely cited cautionary example.
As HRD Canada has reported, HR teams are not fully reckoning with the legal risks already in play from AI deployment. Approximately 40 per cent of Canadian organizations have no official AI policy for their employees, according to Rise People’s State of Canadian HR in 2026 report, meaning staff may be feeding payroll-related queries or sensitive data into AI tools without any organizational oversight in place. The organization remains liable for the outcomes regardless.
When humans in the loop stop paying attention
A subtler risk involves organizations that believe they have addressed AI governance by inserting a human in the loop. In practice, Wunderlich says, that oversight can quietly become an “accountability sink” – a term from the business book The Unaccountability Machine by Dan Davies that refers to structures that absorb or obscure the consequences of a decision so there is no direct individual accountability.
“If you’re the human in the loop and you’re asked to approve these things, it becomes like clicking ‘I approve’ on privacy policies,” he says. “People just do it. And what that does, it gives you the appearance of human-in-the-loop accountability.”
As HRD Canada has examined in its coverage of the top pitfalls HR leaders must avoid when deploying AI in HR, automation fatigue poses a genuine governance risk. Ontario’s AI disclosure rules under the Working for Workers Four Act, 2024, which came into force Jan. 1, 2026, are a step forward – but transparency is not enough, says Wunderlich. “Transparency can end up being the same kind of fig leaf that consent is on the internet because it’s meaningless,” he says. “If you want to raise the regulatory floor, what you want to do is put in some kind of safety rules.”
HR and payroll functions are among the more compliance-ready in any Canadian organization. That same instinct, as shown by Canada’s big banks, needs to be applied to AI governance as Canadian organizations continue to adopt AI – ahead of regulation, and before something goes wrong.
