Canadian CFOs Hit Data and Control Barriers Scaling AI

Canadian midmarket finance chiefs are testing artificial intelligence, but unreliable data links and spending constraints are making it hard to use AI safely in core finance work. Those problems surfaced at a recent CFO Alliance roundtable in Toronto, founder and CEO Nick Araco Jr. told CFO.com.

Some finance leaders said employees were using personal AI accounts, outside company oversight. One company responded by blocking browser-based AI and limiting a desktop tool to selected drives, aiming to restrict which files it could access.

The risk became tangible at a century-old manufacturer. After connecting Claude to Oracle NetSuite for reporting, the company used AI to prepare a cash forecast. The system lost its connection to the finance data and rebuilt the forecast with incorrect figures, which went unnoticed until someone checked.

The company changed its process afterward: AI can help design a report, but a developer builds it inside the company’s system. That keeps the financial output tied to the source data and puts a person between an AI suggestion and a production report.

Vanessa Galarneau, co-founder, CFO and COO of Pluvo, said Canadian finance teams have similar interest in AI to U.S. counterparts but often have fewer resources. She told CFO.com that Canadian CFOs face a higher bar for approving spending without a clear return.

The challenges fit a broader gap between AI adoption and results. KPMG’s 2026 survey found 83% of Canadian finance respondents had moved beyond planning for broad AI use, yet only 36% reported better forecast accuracy. Just 66% said they could efficiently produce evidence for an AI audit, below the global 82% figure.

A Bank of Canada study published in June found personal AI use among business leaders was widespread, while use in production remained limited. It said Canadian firms were still at an early stage of adoption.

For finance chiefs, the immediate test is not simply whether an AI tool can produce an answer. They must know which data is authoritative, where human approval belongs and whether the result can be traced back to its source.