Why Financial Advisors Still Matter: AI Gets It Wrong 57% of the Time
A software engineer in Waterloo asks ChatGPT whether she can deduct her mortgage interest on her tax return. The answer comes back confident and detailed: yes, mortgage interest is deductible, here's how to claim it. She files accordingly. Six months later, CRA sends a reassessment. The deduction doesn't exist in Canada. It's a U.S. provision. The chatbot conflated jurisdictions and cost her $4,200.
This isn't an edge case. A 2026 study commissioned by fintech firm Saturn tested 18 AI models on financial questions and found them wrong or incomplete 57% of the time. More than half. The models don't hedge. They don't say "I'm not sure." They deliver wrong answers in the same authoritative tone they use for correct ones.
The Confidence Problem
The danger isn't that AI makes mistakes. Humans make mistakes. The danger is that AI makes mistakes while sounding exactly like it knows what it's talking about. A human advisor who's unsure will usually signal it: "Let me double-check that," "I'll need to look at your specific situation," "That depends on a few factors." The chatbot generates fluent, structured, professional-sounding text whether the underlying answer is right or fabricated.
That fifty-seven percent failure rate covers queries most Canadians would consider basic. What's my TFSA limit this year? How does the First Home Savings Account work? Should I pay down my mortgage or invest the difference? These aren't exotic tax shelter questions. They're the weeknight Google searches that now increasingly get routed to an LLM instead of a government website or a licenced professional.
The models are especially bad at Canadian content. They're trained predominantly on U.S. financial data, so they'll confidently recommend 401(k) strategies for your RRSP, suggest itemizing deductions that don't apply here, or hallucinate a tax credit that sounds plausible but doesn't exist. When asked about the Canada Learning Bond, one model cited a dollar figure that was off by thirty percent and referenced outdated program details.
What AI Does Well
AI is effective at explaining concepts. Ask it what inflation is, how compound interest works, or why diversification matters, and you'll get a coherent, accurate summary most of the time. It's a decent financial literacy tool at the "explain this term" level. The accuracy collapses when you ask it to do anything with your actual numbers. Multi-step calculations, paying taxes across multiple accounts, risk assessment that accounts for your specific timeline and obligations, these require logic and precision the models don't reliably have. They're built for linguistic fluency, not arithmetic.
Roughly thirty-seven percent of Canadians have used AI for financial research or decision-making help. Most don't know that eighteen percent of AI-generated financial advice contains entirely fabricated regulatory citations. The model invents a rule, attributes it to a real institution, and moves on. You won't catch it unless you already know the rule is fake.
The Liability Gap
If a CFP gives you advice that costs you money, there's a professional liability framework. The advisor can be sued. Their licence can be pulled. There's a complaint process. If a chatbot does the same, there's nothing. No regulator oversees it. No one is on the hook. Canadian securities regulators are monitoring AI use but haven't implemented specific rules. You're in a buyer-beware environment with a product that never admits it might be wrong.
The Office of the Privacy Commissioner has separately raised concerns about entering sensitive financial information into public AI models. Data you input can end up in training sets, be exposed in breaches, or get used in ways you didn't anticipate. Most people don't read the terms before pasting in their income, debts, and account balances.
Where This Ends Up
The future probably isn't AI instead of advisors. It's AI behind advisors, back-office work, research summaries, scenario modelling, with a human accountable for the final output. That model keeps the efficiency gain without the fifty-seven percent error rate landing on the client.
For now, treat the chatbot like a search engine that occasionally lies. Use it to learn concepts. Don't use it to make decisions with actual money on the line. And if it tells you that you can deduct your mortgage interest, double-check that against literally any other source before you file.
A software engineer in Waterloo asks ChatGPT whether she can deduct her mortgage interest on her tax return. The answer comes back confident and detailed: yes, mortgage interest is deductible, here's how to claim it. She files accordingly. Six months later, CRA sends a reassessment. The deduction doesn't exist in Canada. It's a U.S. provision. The chatbot conflated jurisdictions and cost her $4,200.
This isn't an edge case. A 2026 study commissioned by fintech firm Saturn tested 18 AI models on financial questions and found them wrong or incomplete 57% of the time. More than half. The models don't hedge. They don't say "I'm not sure." They deliver wrong answers in the same authoritative tone they use for correct ones.
The Confidence Problem
The danger isn't that AI makes mistakes. Humans make mistakes. The danger is that AI makes mistakes while sounding exactly like it knows what it's talking about. A human advisor who's unsure will usually signal it: "Let me double-check that," "I'll need to look at your specific situation," "That depends on a few factors." The chatbot generates fluent, structured, professional-sounding text whether the underlying answer is right or fabricated.
That fifty-seven percent failure rate covers queries most Canadians would consider basic. What's my TFSA limit this year? How does the First Home Savings Account work? Should I pay down my mortgage or invest the difference? These aren't exotic tax shelter questions. They're the weeknight Google searches that now increasingly get routed to an LLM instead of a government website or a licenced professional.
The models are especially bad at Canadian content. They're trained predominantly on U.S. financial data, so they'll confidently recommend 401(k) strategies for your RRSP, suggest itemizing deductions that don't apply here, or hallucinate a tax credit that sounds plausible but doesn't exist. When asked about the Canada Learning Bond, one model cited a dollar figure that was off by thirty percent and referenced outdated program details.
What AI Does Well
AI is effective at explaining concepts. Ask it what inflation is, how compound interest works, or why diversification matters, and you'll get a coherent, accurate summary most of the time. It's a decent financial literacy tool at the "explain this term" level. The accuracy collapses when you ask it to do anything with your actual numbers. Multi-step calculations, paying taxes across multiple accounts, risk assessment that accounts for your specific timeline and obligations, these require logic and precision the models don't reliably have. They're built for linguistic fluency, not arithmetic.
Roughly thirty-seven percent of Canadians have used AI for financial research or decision-making help. Most don't know that eighteen percent of AI-generated financial advice contains entirely fabricated regulatory citations. The model invents a rule, attributes it to a real institution, and moves on. You won't catch it unless you already know the rule is fake.
The Liability Gap
If a CFP gives you advice that costs you money, there's a professional liability framework. The advisor can be sued. Their licence can be pulled. There's a complaint process. If a chatbot does the same, there's nothing. No regulator oversees it. No one is on the hook. Canadian securities regulators are monitoring AI use but haven't implemented specific rules. You're in a buyer-beware environment with a product that never admits it might be wrong.
The Office of the Privacy Commissioner has separately raised concerns about entering sensitive financial information into public AI models. Data you input can end up in training sets, be exposed in breaches, or get used in ways you didn't anticipate. Most people don't read the terms before pasting in their income, debts, and account balances.
Where This Ends Up
The future probably isn't AI instead of advisors. It's AI behind advisors, back-office work, research summaries, scenario modelling, with a human accountable for the final output. That model keeps the efficiency gain without the fifty-seven percent error rate landing on the client.
For now, treat the chatbot like a search engine that occasionally lies. Use it to learn concepts. Don't use it to make decisions with actual money on the line. And if it tells you that you can deduct your mortgage interest, double-check that against literally any other source before you file.
Sources
Read Next
What 27 Rising Stars Under 40 Reveal About Canada's Next Wealth Model
The Deficit Fell $11.6 Billion. Your Mortgage Rate Didn't Get the Memo.
Banks Call It a Transfer Fee, but It Works Like an Exit Penalty
Leveraged Insurance Ends When Regulators See the Downside, Whatever the Math Says