Can You Trust AI Tax Advice in Canada? A Former Auditor’s Take

More business owners are doing something that feels harmless. They have a tax question, they type it into ChatGPT or a website chatbot, they get a clean, confident answer in seconds, and they act on it. No wait time, no invoice, no jargon. It feels like progress.

Here is the problem. In 2025, Canada’s Auditor General tested the CRA’s own tax chatbot and found it gave correct answers only about a third of the time. If the tax agency’s own AI is wrong more often than it is right, it is worth asking what you are really getting when you trust AI tax advice for your own return. We spent years inside the CRA before starting Better Books Canada, so here is the straight version.

The short answer

Can you trust AI for your Canadian taxes? Not on its own, and not for anything with money on the line. AI tools are genuinely useful for understanding a concept or drafting a question. They are not reliable for a filing position, a dollar figure, or a rule that decides how much tax you owe.

The reason is simple. These tools are built to sound right, not to be right. Canadian tax is exactly the kind of subject where sounding right and being right come apart.

The CRA’s own AI is wrong more than it is right

Start with the clearest example, because it is the CRA’s own. In its 2025 report on the CRA’s contact centres, the Office of the Auditor General of Canada tested the agency’s online chatbot, Charlie, and found it answered correctly only about a third of the time. That sits well below the CRA’s own internal target of 70 percent.

The human side was not much better. When the Auditor General placed test calls to live CRA agents, general questions about individual taxes were answered accurately only 17 percent of the time. Questions about business taxes and benefits came in around 54 percent. You can read the findings in the Auditor General’s 2025 report on the CRA’s contact centres.

The CRA has since upgraded Charlie to a newer generative version and says early testing showed around 90 percent accuracy, though that figure has not been independently verified in real world use. The point still stands. The organization that writes the rules could not keep its own tools consistently accurate, and it spent roughly 18 million dollars trying.

Why AI gets tax so confidently wrong

AI language tools work by predicting the most likely next words based on patterns in their training data. They are not opening the Income Tax Act and reasoning through your situation. They are producing text that reads like a correct answer. Most of the time that is close enough. With tax, close enough is often just wrong. A good example is the quick method remittance rates for GST/HST, which we have seen quoted from the wrong column and applied to the wrong base by tools that sounded certain.

Canadian tax is a hard case for these tools for a few reasons. The rules change every year, so a model trained on older text repeats last year’s number with total confidence. Much of the tax writing online is American, so answers quietly blend IRS rules into Canadian ones. And the real answer usually depends on details the tool never asked about, like your province, your incorporation status, or whether an expense was truly for business. Anything that crosses a provincial line is where that gap shows up fastest, which is why we wrote out the Quebec tax rules for an Ontario resident in full.

We saw the same failure pattern from the other side of the desk. During our years reviewing files at the CRA, the returns that went wrong were rarely the work of someone acting in bad faith. They were usually built on one confident, wrong assumption that cascaded through the rest of the return. An AI answer that is 90 percent right and 10 percent wrong is the perfect machine for producing exactly that kind of mistake, because the wrong part sounds identical to the right part.

When AI is wrong, you pay, not the AI

Here is the part that matters most, and it is the part people miss. If you act on a wrong AI answer and underpay your tax, the cost lands on you. Not on the tool, and not on the CRA.

It is worth contrasting two situations. In 2024, a British Columbia tribunal ordered Air Canada to pay a customer who had relied on wrong information from the airline’s website chatbot. The tribunal decided a company is responsible for what its chatbot says, and would not let Air Canada treat the bot as a separate party. You can read that decision in Moffatt v. Air Canada.

Your tax return does not work the same way. If a CRA agent, the CRA chatbot, or ChatGPT gives you a wrong answer and you file on it, you are still responsible for the correct amount of tax. Bad advice does not erase what you owe. The most you can usually ask for is relief from the penalties and interest, not the tax itself, and even that is not automatic. It is also worth knowing what triggers a CRA audit in the first place, because a chatbot answer that inflates a deduction lands on exactly the criteria the agency screens for.

The CRA can cancel or waive penalties and interest in certain situations, including where its own error contributed, but you have to apply and make the case. The details are on the CRA’s page on taxpayer relief provisions. So the asymmetry is real. A private company can be made to answer for its chatbot. When it comes to your return, the responsibility sits with you, which is exactly why a confident wrong answer is so expensive.

Where AI actually helps

None of this means AI is useless for tax. Used well, it is a solid starting point. It is good at explaining a term in plain language, giving you the general shape of how something works, helping you organize your receipts and questions before a meeting, and turning a vague worry into a specific question you can actually get answered.

The line is simple. Use AI to understand, not to decide. The moment a real dollar figure, a deduction, or a filing position is involved, the answer has to be verified against the actual rule, or checked by a person who is accountable for it.

What to do instead

You do not have to choose between a chatbot and nothing. Here is the approach we would give any business owner.

• Verify every hard number against canada.ca. If an AI tells you a rate, a limit, or a deadline, confirm it at the source before you rely on it. Tax figures change every year.
• Keep your own records. AI and the CRA both only know what has been entered or reported. Solid records are what let you prove your position, and they are the backbone of the record keeping requirements for your business.
• Know which questions need a human. Anything involving what you can write off as a small business owner, your incorporation, payroll, or HST and GST is not a chatbot question.
• Understand what the CRA does automatically. For simple personal situations there is real free help, which we cover in our guide to automatic tax filing in Canada. For a business return, that easy button does not exist.

And if you ever want to see how a file looks to the person reviewing it, our guide on what auditors look for during a CRA audit walks through it. The habits that hold up in an audit are the same ones that protect you from a bad AI answer.

Frequently asked questions

Can I use ChatGPT to do my Canadian taxes?

You can use it to understand concepts or draft questions, but not to prepare or file a return you rely on. It regularly blends outdated or American rules into Canadian answers, and you remain responsible for any errors. Verify anything with a dollar figure attached against canada.ca or with a professional.

Is the CRA’s chatbot accurate?

Not reliably. Canada’s Auditor General tested the CRA chatbot in 2025 and found it correct only about a third of the time, below the CRA’s own 70 percent target. The CRA has since upgraded it and claims higher accuracy, but that has not been independently verified.

If AI or the CRA gives me wrong tax advice, am I off the hook?

No. You are responsible for the correct amount of tax on your return regardless of who gave you the wrong answer. You can sometimes get penalties and interest cancelled or waived through the CRA’s taxpayer relief provisions, especially where the CRA’s own error contributed, but the underlying tax still stands and relief is not automatic.

What is AI actually good for when it comes to tax?

Explaining terminology, outlining how a rule generally works, and helping you organize records and questions before you talk to a professional. Treat its output as a first draft to verify, never as a final filing decision.

When should I stop using AI and call a bookkeeper or accountant?

The moment real money is on the line. Deductions, incorporation, payroll, HST and GST, and anything audit related need a person who is accountable for the answer. A confident wrong answer on those costs far more than getting it done right the first time, and usually far more than what a bookkeeper costs in Canada. Which of the two you actually need depends on the question, and we set that out in bookkeeper vs accountant.

Get an answer you can actually rely on

AI is a useful assistant and a poor final authority, especially when the bill for a wrong answer is yours to pay. If you have a tax or bookkeeping question where real money is on the line, reach out to our team and get an answer from people who spent years on the other side of the CRA’s desk. We would rather you ask us than trust a confident guess.

Last updated: June 2026. Tax rules and CRA services change, so verify current figures and programs against canada.ca.