By Rob Morrison, Chief Marketing Officer, Otivo
Two people can ask an AI the same money question and end up roughly US$50,000 apart by age 60. That's not a hypothetical — it's what researchers at MIT Sloan and Stanford Graduate School of Business found when they simulated the lifetime effects of following AI-generated financial guidance. The gap didn't come from the technology being poor; it came from what the person asking already knew, and how much of it they thought to mention. Here's what the research found, why the finding cuts against the technology's biggest promise, and what separates a good answer from good advice.
Quick answer
AI financial advice quality depends heavily on the prompt. MIT Sloan and Stanford research published in 2026 found large language models gave sound general guidance, but recommended lower equity allocations and lower saving rates to less financially literate users and to those new to AI — compounding into US$50,000 to US$100,000 less wealth by age 60.
What did the research actually find about AI financial advice?
The study set out to test whether following AI financial advice over a lifetime leaves people better or worse off, and the headline finding was that it mostly leaves them better off. Taha Choukhmane, Weidong Lin and Matthew Akuzawa of MIT Sloan, with Tim de Silva of Stanford Graduate School of Business, asked a sample of 1,000 adults to write their own prompts seeking spending and investing advice, ran those prompts through GPT-5.2, GPT-5.6 and Gemini 3 Flash, then simulated what would happen to people aged 22 to 89 who followed the answers. Their paper, "AI Financial Advice: Supply, Demand, and Life Cycle Implications", won the Swiss Finance Institute Outstanding Paper Award in 2026.
The researchers expected problems. What they found was that the models consistently steered people towards saving during their working years, drawing down in retirement, holding diversified stock funds, and reducing stock exposure after age 45. As Choukhmane put it, the advice lining up with what academics consider sound financial principles was not a given.
Where it fell short was nuance. The models leaned on rules of thumb and handled change badly. Someone who lost a job was told to cut spending sharply even when they had savings sitting there to draw on. Portfolios were left to drift rather than rebalanced. These are the moments that matter most in a financial life, and they're precisely the moments a single question-and-answer exchange can't see coming.
One caveat before the dollar figures. This is a US study, built on US tax rules and Social Security assumptions, and the amounts are US dollars. The pattern travels to Australia. The numbers don't.
Why does the person asking change the answer?
Because the model only knows what the prompt tells it, and different people tell it different things. Following the advice generated by prompts written by men, by more financially literate users, or by people who had used AI for financial questions before produced around 5% more wealth close to retirement. That sounds modest until it's expressed in dollars.
Two specific gaps came out of the simulations. The model recommended higher equity allocations in response to prompts written by men and by users with high financial literacy, and over a working life that difference compounded to roughly US$50,000 — about 4% — less wealth at age 60 for women and for less financially literate users. Separately, the model recommended lower saving rates to people who had never used AI for financial advice before, leaving them with almost US$100,000, or 6%, less at 60 than users with prior experience.
The gender result has a second layer worth sitting with. About two-thirds of that gap traced back to how men and women wrote their prompts — different topics, different vocabulary, men reaching for words like strategy and growth while women wrote about family and paying for things. The remaining third came from the model giving different advice when the identical prompt was labelled as coming from a woman.
Consider a 34-year-old on $85,000 who has never asked an AI a money question and doesn't think to mention her employment stability, her existing savings or how long she intends to keep working. She isn't withholding anything. She simply has no reason to know those details change the answer. The tool answers the question she asked, competently, and she has no way of telling that a differently worded version of the same question would have produced better guidance.
This is the awkward part. One of the strongest arguments for AI in financial services is that it can put help within reach of people who could never afford to sit down with a professional. Yet the people with the least financial knowledge appear to get the least out of it. The old advice gap was largely about cost. The new one looks like it's about knowledge.
That matters in Australia already. ASIC's Moneysmart research, released on 16 March 2026, found that 18% of Australians aged 18 to 28 are using AI platforms for financial information and guidance, and that 64% trust those platforms for money advice — including 16% who trust them completely.
What can't a general-purpose AI know about your money?
Whatever you don't tell it, which is usually the part that matters. A general-purpose tool starts every conversation from nothing and works with what it's handed.
Take a question as ordinary as whether to put an extra $500 a month into the mortgage. Answering it properly means knowing the loan's interest rate, the super balance, the marginal tax rate, what other debts exist, the person's age and when they intend to stop working. Someone asking that question in a chat window is unlikely to volunteer all of it — not out of evasion, but because there's no obvious reason to think a mortgage question has anything to do with super. That's the sort of question Otivo's debt advice module is built to work through, weighing repayment against the rest of a household's position rather than in isolation.
A good adviser doesn't hand someone a blank page and ask for the perfect question. They ask the questions themselves. They find out what's relevant, including the things the person didn't know were relevant. The interesting design problem in AI financial advice isn't making models smarter. It's building systems that already know what to ask.
The three gaps between a good answer and good advice
Reading the research alongside how licensed advice actually works in Australia, three gaps separate a competent answer from advice you can rely on. Call them the context gap, the continuity gap and the accountability gap.
The context gap
A general-purpose tool knows what's in the conversation. A licensed advice platform starts from a financial position — income, super, debts, cover, goals — and brings the relevant parts into the answer without being prompted. The researchers demonstrated this themselves: when they replaced everyday prompts with structured ones containing full financial information and explicit assumptions, the advice improved. Which means the fix is structure. And structure is something a system can supply rather than something a user has to know how to provide. how a general-purpose chatbot differs from licensed advice
The continuity gap
Financial advice isn't a series of isolated questions, because a financial life isn't a series of isolated moments. Salaries change. Markets move. Children arrive, relationships end, inheritances land, retirement approaches. The guidance that fit five years ago may not fit now. This is exactly where the models struggled — job loss, drawdown, rebalancing — and it's structural rather than a temporary limitation. A system that forgets everything between conversations can't notice that something changed. Advice that lives with someone over time is a different product from advice that answers a question.
The accountability gap
The MIT and Stanford researchers measured whether AI guidance produced good economic outcomes. That's a different question from whether it meets a regulatory standard. In Australia, personal financial advice is a licensed activity carrying a best interests duty, disclosure obligations and access to external dispute resolution through AFCA. General-purpose AI tools sit outside that framework. The study also found something that sharpens the point: the models regularly named specific account types, products and providers that respondents had never mentioned, in one case surfacing a particular fund manager in 6% of responses when fewer than 0.4% of prompts referred to it. what the law says about AI and financial advice in Australia
What would actually close the advice gap?
Not teaching millions of Australians to write better prompts. That's the answer that puts the burden back on the person with the least capacity to carry it, and the research is fairly clear that the burden is where the damage happens.
The alternative is to build the questions in. If structured prompts containing someone's full financial position produce better guidance — and they do — then the sensible response is to construct systems that hold that position already, ask what's missing, and carry it forward. Choukhmane's own suggestion for consumers is to treat AI first as a way to build financial understanding rather than something to follow, and he notes that for people who can't afford professional help, AI is a genuinely useful and inexpensive option. Both things point the same direction: the value isn't in the answer, it's in what the system knows before the question arrives.
There's a reasonable measure of success buried in all this. It isn't how capable the models become. It's whether someone who knows almost nothing about money can get help that's as useful as what someone who knows exactly what to ask receives.
Frequently asked questions
Is AI financial advice reliable?
Research from MIT Sloan and Stanford published in 2026 found that large language models generally gave sound guidance on saving, diversification and reducing investment risk with age, but handled change poorly — over-correcting after a job loss and allowing portfolios to drift instead of rebalancing. Reliability also varied with the prompt. General-purpose AI tools in Australia sit outside the licensing framework that governs personal financial advice.
Does the way you word a question change the financial advice AI gives?
Yes, and measurably. In the MIT Sloan and Stanford simulations, prompts written by men, by more financially literate users, and by people with prior AI experience produced advice that led to roughly 5% more wealth close to retirement. Adding full financial detail and explicit assumptions to a prompt improved the quality of the advice across the board.
Is an AI financial advisor the same as a licensed financial advice service in Australia?
No. Whether spelled adviser or advisor, the term is restricted under section 923C of the Corporations Act to individuals listed on ASIC's Financial Adviser Register, so it doesn't describe a digital tool. More practically, a general-purpose AI tool answers questions; a licensed digital financial advice service operates under an AFSL, owes a best interests duty when providing personal advice, and gives access to external dispute resolution.
What information does an AI need to give useful financial guidance?
The research points to age, employment status and stability, income, savings and investment balances, debts, and intended retirement timing, along with explicit assumptions about the economic environment. Prompts containing that detail produced better simulated outcomes than typical everyday prompts. The practical difficulty is that knowing which details matter is itself a form of financial literacy.
Otivo was built to solve exactly this problem. Nobody should need to know the perfect question, or which pieces of their financial life are relevant, before they can get good advice.
It builds a picture of your financial position — your income, super, debts, goals and other relevant circumstances — and uses that context when you ask a question. If something important is missing, it can ask. And because that information stays connected to your financial plan, each conversation doesn't have to start from scratch. The retirement planning module is one place to see what that looks like.
Importantly, all of this happens inside a licensed financial advice service. Otivo Pty Ltd holds AFSL and Australian Credit Licence No. 485665, combining the accessibility and convenience of AI with the safeguards, obligations and accountability of licensed advice.
Because the real opportunity for AI isn't simply giving people better answers. It's giving more Australians access to personal, licensed financial advice without requiring them to become financial experts first. Good financial advice shouldn't depend on how good you are at asking for it.
Sources
- Choukhmane, T., de Silva, T., Lin, W. and Akuzawa, M., "AI Financial Advice: Supply, Demand, and Life Cycle Implications", MIT Sloan School of Management and Stanford Graduate School of Business, 2026. Summarised at mitsloan.mit.edu
- ASIC, media release 26-049MR, "ASIC urges Gen Z to sense-check money advice as social media fuels riskier financial decisions", 16 March 2026. asic.gov.au
Disclaimer
The information in this communication is current as at September 2026 and has been prepared by Otivo Pty Ltd ABN 47 602 457 732, AFSL and Australian Credit Licence No. 485665. This content is general information only and has been prepared without taking into account your objectives, financial situation or needs. It is not personal financial or taxation advice and should not be relied on as such. Before acting on any information, you should consider its appropriateness having regard to your personal circumstances. This material must not be reproduced in whole or in part, or posted on any social media platform, without the prior written consent of Otivo Pty Ltd.