76 per cent of accountants and bookkeepers said they saw an increase in clients using public AI tools for advice, a recent survey showed.Scott Olson/Getty Images
The deadline for filing taxes in Canada for 2026 is April 30. As the big day approaches, Globe Advisor and Globe Investor have teamed up to offer advice on how to maximize returns, find credits and avoid an audit. The full series can be found here.
Ross Pasceri had no idea where the numbers were coming from.
His firm had just started experimenting with the use of artificial intelligence tools to summarize new tax rules for their online publication when one sentence caught the accountant’s eye.
Mr. Pasceri said the AI had incorrectly “spit out” that the 2026 capital gains limit had increased from about $1.25-million to $1.3-million. “The CRA didn’t announce it, it’s not in the legislation, and the Department of Finance didn’t announce it – it was totally hallucinating," he said.
Tax specialists such as Mr. Pasceri are spending more hours fixing mistakes made by publicly available AI tools such as Google Gemini and ChatGPT, while also learning how to optimize their use for business. But rather than ushering clients away from the technology completely, many accountants are eager to encourage them to learn how to use it properly to maximize returns.
In a tough job market, what AI skills do you really need?
According to a survey released earlier this year by Dext, an AI bookkeeping software, 76 per cent of accountants and bookkeepers surveyed said they saw an increase in clients using public AI tools for financial, tax or bookkeeping advice in 2025.
But rather than the technology simplifying their workload, nearly half of those catching AI-related errors said they spent up to three hours a month fixing mistakes caused by generated advice. Another 27 per cent said they spent between four and six hours doing so.
Incorrect interpretations of business expenses were the most common issue, reported by 44 per cent of respondents. Incorrect tax claims or charges followed closely at 43 per cent.
Some of those slips added up to downright fraud. More than 40 per cent of those who responded to the survey expect increased misuse of AI outputs to justify inappropriate or fraudulent claims.
“If a tax return contains errors, you’re responsible – even if AI helped you,” said Ryan Minor, director of tax at Chartered Professional Accountants of Canada. “Obviously, you can’t 100 per cent rely on it.”
But, Mr. Minor said, the potential for tax-filing efficiencies with AI applications is enormous. He highlighted tools that would be able to scan and look for tax planning opportunities, dictate phone calls and prepare memos based on your discussions.
For business owners, “AI can look at a return, figure out what industry the business operates in and suggest things you may have missed,” he said.
Few other fields have been under as much of a threat from AI as accounting. It regularly ranks among the most automatable jobs, given the large amount of routine tasks and data entry.
According to a 2025 report from KPMG, more than 82 per cent of the Canadian companies surveyed said they were using or piloting AI in finance, compared with 71 per cent globally. Accounting functions were the leading areas of adoption.
Despite the rapid pace of innovation in this field, experts say AI is far from ready to replace humans – or even replace traditional accounting software, for that matter.
“These models are trained with a big collection of stuff,” said Mu Zhu, associate dean of AI strategy at the University of Waterloo. Such a model “might read an article and say, ‘Okay, this is about tax,’ but it may not remember that the discussion was specific to the United States.”
General-purpose AI tools are also not trained with tax-return data, he said.
A common scenario involves ChatGPT incorrectly telling clients transactions that can be deferred under U.S. tax law can also be deferred in Canada, Mr. Pasceri said. “AI isn’t geared to answer questions solely in the Canadian context.”
Bots may also struggle to keep pace with legislative changes and fast-changing tax rules. “Companies like OpenAI don’t release new versions every day,” Mr. Zhu said, adding that training and updating large models requires significant time and resources.
While AI models can identify patterns and relay general rules they learn from government websites and other source materials, programs such as TurboTax rely on rules specifically programmed into the system.
Put simply, these traditional tools use “a bunch of ‘if-then’ statements,” Mr. Zhu said. When tax laws change, engineers update the program so the latest version reflects the current rules.
Old-school tax software also isn’t prone to the same kind of hallucination issues.
Mr. Pasceri recalled one case in which a client cited a technical interpretation from the Canada Revenue Agency, not realizing that the AI tool they were using had hallucinated it.
“It was a completely made-up document,” he said. “The scariest part – it looks credible on the surface.”
Where AI systems excel is in identifying credits or deductions that might apply to a specific taxpayer’s situation, according to Mr. Minor.
For example, someone caring for a disabled relative or renovating a home for accessibility could describe their situation to an AI bot and get a list of tax credits they had not previously considered.
“It’s useful for scoping out potential opportunities,” he said.
Public AI tools are also helpful for unpacking complex CRA form instructions (such as “enter the amount of UCCB repayment from line 213″) and breaking down exactly where to find necessary details. For instance, such tools could categorize transactions or extract text from old receipts or screenshots of Uber charges to input into forms.
Using dedicated, Canadian AI tools specifically designed for taxes, such as TaxGPT and CloudTax, can also potentially help tax-filers get better outcomes because these tools refine the scope of data they’re trained on, keeping it more relevant to national tax rules. Individual tax firms are also building out their own software to attract users.
Where AI tools will inevitably still lag, Mr. Zhu said, is understanding how unique overlapping traits – a person’s marital status, dependants, health conditions and countless other factors – interact to influence tax outcomes. “It can’t adapt to the nuance of the individual’s situation,” he said.
