MIT study finds AI financial advice works—with caveats

AI financial advice can produce sizable savings buffers for virtually everyone over age 30, according to new research from the MIT Sloan School of Management. But the study also found that the quality of the guidance depends heavily on how users phrase their prompts—and that differences in those prompts can compound into substantial wealth gaps.

Taha Choukhmane, an assistant professor of finance at MIT Sloan, and co-authors tested how large language models advise people on spending, saving, and investing over their lifetimes. The models generally recommended saving during working years, drawing down assets in retirement, investing heavily in diversified stock funds, and reducing stock exposure after age 45.

“We were somewhat surprised by how good the advice was. Especially when you read the kind of questions people asked, it was not a given that the advice would line up with what academics think are good financial principles.”

How the financial advice study worked

The researchers built a life-cycle financial model covering how incomes, employment, investments, and taxes typically change over time. That model provided a benchmark for evaluating what constituted sound financial decisions.

Recommended reading

Startup targets Nvidia with 128TB AI server

They then asked a sample of 1,000 adults to write their own prompts requesting financial advice from GPT-5.2, GPT-5.6, or Gemini 3 Flash. The researchers simulated what would happen if people aged 22 to 89 followed the resulting recommendations repeatedly over time.

The simulations covered decisions about:

• Spending and saving
• Stock-market participation and portfolio allocation
• Investment risk over the life cycle
• Responses to changes such as unemployment

The team repeated the exercise with structured academic prompts containing each person’s age, job status, income, savings balances, and assumptions about the economic environment. Those prompts also specified normal life expectancy, living expenditures, retirement age, employment risk, income risk, current U.S. tax law, and Social Security rules that would not change.

The researchers compared both forms of simulated AI advice with what people were already doing without AI, as well as with the advice generated by the academic prompts.

Better prompts improved saving and investing advice

The models performed better than the researchers expected even when responding to ordinary users' questions. Their recommendations generally encouraged higher savings, greater participation in the stock market, diversified portfolios, and age-appropriate risk-taking.

However, the models were weaker at handling financial shocks and ongoing portfolio management. They relied too much on simple rules of thumb for spending and saving, advised people who had lost jobs to cut spending too sharply even when they had savings, and allowed portfolios to drift instead of actively rebalancing them.

A typical user prompt might be: “Where should I invest starting with $50 and consistently adding $25 a month after?” By contrast, a structured prompt could define the user’s income, expenses, job and income risks, expected retirement age, and tax assumptions.

“Regular people are not writing their prompts the way a finance professor is.”

The authors said this makes AI a potentially affordable and widely accessible source of guidance, particularly because traditional human financial advice can involve significant costs, biases, and conflicts of interest. But the study did not establish that every recommendation was suitable for every individual, and it found that the models often failed to rebalance portfolios actively.

Prompt differences produced wealth gaps

The advice also varied according to a user’s gender, financial literacy, and prior experience with AI. Following recommendations generated from prompts written by men, more financially literate users, or people with previous AI experience produced about 5% more wealth near retirement.

The researchers found that:

• The models recommended higher equity allocations in response to prompts written by men and by users with high financial literacy.
• Women and less financially literate users ended up with roughly $50,000, or 4%, less wealth at age 60.
• People without prior experience using AI for financial advice received lower recommended saving rates and ended up with almost $100,000, or 6%, less wealth at age 60 than people with prior AI experience.

Some of the disparity came from how people asked questions. Women were more likely to use words such as “family,” “grocery,” and “pay,” while men more often used terms including “strategy,” “crypto,” and “growth.” But the model also changed its recommendations when the same prompt was labeled as coming from a woman rather than a man. About two-thirds of the gender gap in wealth outcomes came from differences in prompts, while the remaining third came from that change in the model’s advice.

Choukhmane said the latter pattern could reflect reasonable inferences about differences in preferences, life expectancy, or income risk—or bias learned from training data. The study argues that progress requires clearer benchmarks for how financial advice should vary across demographic groups.

Advice for consumers and financial firms

The researchers recommend that consumers use AI first to build financial understanding rather than simply follow its recommendations. Prompts grounded in life-cycle planning, portfolio theory, and explicit real-world assumptions improved advice on saving and spending and reduced reliance on basic rules of thumb. Users should also ask models to guard against demographic bias.

Choukhmane said AI can complement a human advisor, helping clients implement advice between meetings. For people unable to afford an advisor, he described AI as an inexpensive way to access financial guidance—while acknowledging that the people most likely to benefit may have the least financial literacy or the least ability to write detailed prompts.

The findings also have implications for financial companies. LLMs recommended specific account types, financial products, and providers that users had not mentioned. Vanguard products appeared in 6% of model responses and iShares products in 3.4%, even though fewer than 0.4% of prompts mentioned either company.

That suggests product discovery may increasingly depend on how financial offerings are described by LLMs, rather than only on traditional marketing or search visibility.

The paper, “AI Financial Advice: Supply, Demand, and Life Cycle Implications,” won the Swiss Finance Institute Outstanding Paper Award 2026. Its authors are Taha Choukhmane, Weidong Lin, and Matthew Akuzawa from MIT Sloan, and Tim de Silva from Stanford Graduate School of Business.