Canadian CFOs Confront 'Shadow AI' and Broken ERP Forecasts

A century-old Canadian manufacturer rebuilt its cash forecast on wrong numbers after an AI tool's connection to Oracle NetSuite dropped without anyone noticing. The company had connected Claude to its ERP to produce KPIs, and its controller had created an AI skill to build the forecast. "Nobody caught it until they went looking," said Nick Araco Jr., founder and CEO of the CFO Alliance.

The failure, described at a recent CFO Alliance roundtable in Toronto, captures the core problem Canadian midmarket finance leaders face: turning AI experiments into dependable financial processes. Finance chiefs at the event flagged unreliable data-sharing systems and technology spending they could not allocate properly.

A different point in the adoption cycle

Araco said the Toronto leaders raised issues that had not surfaced at the organization's earlier U.S. roundtables, including the contract risk around what Deloitte has quantified overseas as "shadow AI" — AI use running through employees' personal accounts. Some Canadian CFOs said they are watching companies south of the border move first and applying those lessons.

"Toronto is having the U.S. conversation with the benefit of watching first," Araco said. He found the Toronto finance leaders more deliberate about governance as they worked through the same questions.

Vanessa Galarneau, a Canadian finance executive now in San Francisco, co-founded Pluvo, a company she describes as an "AI-native layer built for strategic finance," where she serves as CFO and COO. She said Canadian finance teams have the same appetite for AI as their U.S. counterparts but fewer resources to move experiments into production. Silicon Valley companies often have growth capital to support R&D across the business, she told CFO.com, while Canadian CFOs face a higher bar for spending without a clear return.

"There's less margin for error [in Canadian markets]," Galarneau said. That constrains hiring of technical talent and raises the stakes for every funded experiment.

The invisible leak

The first concern inside the roundtable involved AI use beyond approved tools. "The leak isn't an enterprise tool. It's the personal account," Araco said. Employees can now move company information outside a controlled environment within seconds. "At that moment, the company's zero-data-retention contract with its vendor means nothing," he said.

Customer nondisclosure agreements add risk, Araco said, because employees may not connect the NDA their company signed to the spreadsheet they paste into a consumer AI tool. He said this contract angle was more prominent in Toronto than at any recent U.S. roundtable.

One company at the event now blocks browser-based AI and requires employees to use a desktop application restricted to certain drives, keeping output local and limiting the tool's reach. Another CFO said they had identified themselves as the CFO in the settings and instructed the tool to treat everything as confidential. One participant responded: "I wouldn't trust that as far as I can throw it."

Galarneau assumes employees at most companies already use personal AI accounts, and the absence of reliable data makes the exposure harder to manage. "I don't think anyone has a clean number, and that's part of the problem," she said.

Design versus production

The manufacturer's forecast failure produced a new operating rule: finance can use AI to determine what a report should include and how it should look, but a developer then builds the report directly inside the system.

"AI is for design, and systems are for production," Araco said. He called it "the most practical thing" he had heard across the roundtable series, because it gives finance teams a repeatable process for human involvement.

Galarneau said finance must establish which data is authoritative, decide where approvals belong, and preserve the ability to trace an answer back to its source. "A lot of pilots prove the model can do something useful," she said. "Far fewer prove that it is worth operationalizing and safe enough for finance to rely on."

The required talent is scarce on both sides of the border, she said: people who understand finance and can translate its processes into technical infrastructure. This gap is creating international demand for great finance engineers. "You cannot just bolt [model context protocol] onto old finance infrastructure and expect it to become AI-native," Galarneau said.

Counting the cost

"Token is the cost that nobody can see," Araco said. One Toronto finance leader could see an engineer's $5,000 spend on AI across 30 days, with no idea which project produced the expense.

Customers were also driving up costs by submitting broad requests to AI agents in operation, some CFOs said. Those teams began assigning models to specific work and found a newer low-cost model matched the flagship model on one job — cutting that group's AI spending by around 30%.

"If I'm driving to the corner store, I don't need a Ferrari," one CFO said. Another participant summed up the ownership question: "We don't need a chief token officer." Araco captured the broader adoption dilemma in one line: "Do I want to be customer 46, or customer 46,001?"

One finance leader said AI had expanded what the company could do without delivering an obvious efficiency gain. "We're not more productive. We're doing things we never did before. How do you price that?" the executive asked.

Accountability remains critical as AI takes on routine work. One participant said AI could increase annual report production from 25 to 40, though every figure would need line-by-line checks. The finance leader told the room: "I can't tell a judge the AI said so."