AI Financial Advice Is Better When You Provide the Whole Picture

New research covered by MIT Sloan suggests that AI can provide surprisingly reasonable financial guidance, but the quality of that advice depends heavily on how the question is asked.

The researchers collected financial questions from approximately 1,000 adults and simulated what could happen if people followed the resulting AI advice throughout their lives. The models generally encouraged people to save more, hold diversified investments and gradually reduce investment risk as they aged.

The advice was not consistently good. AI relied too heavily on simple rules of thumb, responded poorly to disruptions such as unemployment and often allowed investment allocations to drift instead of recommending active rebalancing.

Better prompts produced better results.

The researchers created what they called an academic prompt. It gave the model a complete description of the person’s income, employment, savings and investments. It also instructed the model to use life-cycle planning and modern portfolio theory, act in the person’s best interest and state its assumptions about taxes, investment returns, retirement and financial risk.

This reduced the model’s reliance on generic rules such as always saving a round percentage of income or never withdrawing more than four percent in retirement. It also produced better advice following income shocks, although even the improved prompt did not solve the problem of passive portfolio drift.

The study also found troubling differences based on who appeared to be asking. Advice generated from prompts written by men, financially experienced people and previous AI users produced better simulated retirement outcomes. Some of the difference came from how people framed their questions, but the model also changed its advice when the same prompt was assigned a different gender.

That means a good financial prompt should provide relevant facts directly and tell the model not to infer risk tolerance, financial ability or goals from demographic stereotypes.

Here is a reusable starting point:

Help me evaluate my finances using life-cycle planning and modern portfolio theory. Act in my best interest, explain your assumptions and do not infer my risk tolerance, knowledge or goals from my gender or other demographic characteristics.

Do not recommend anything until you identify important missing information and ask me for it.

I will provide rounded amounts for:

• My age range and expected retirement age
• Household size and number of dependents
• Employment status and annual after-tax income
• Essential and optional monthly spending
• Debts, balances, interest rates and minimum payments
• Cash and emergency savings
• Retirement accounts and any employer contribution
• Taxable investments
• Insurance coverage
• Major goals and their expected timelines
• Expected changes to income or expenses
• My ability and willingness to accept investment losses

Based on that information:

1. Identify my most important financial priorities in order.
2. Recommend realistic monthly and annual spending and saving ranges.
3. Calculate an appropriate emergency-fund target.
4. Explain how I should balance debt repayment, cash savings and investing.
5. Suggest a diversified asset allocation appropriate for my goals and risk capacity.
6. Define when and how the portfolio should be rebalanced.
7. Stress-test the plan against unemployment, a major unexpected expense and a significant market decline.
8. Explain what assumptions could materially change the recommendation.
9. Identify anything that should be verified with a fiduciary financial planner, accountant, attorney or tax professional.

Do not invent missing facts. Do not recommend individual stocks, cryptocurrencies or branded financial products unless I specifically request them. When mentioning a product, explain the costs, risks and reasonable alternatives.

Readers should not include names, addresses, Social Security numbers, exact birth dates, employer names, bank names, account numbers or login information. Rounded balances and general descriptions should be sufficient for planning.

AI still cannot guarantee that tax rules are current, understand every unusual circumstance or assume the legal obligations of a fiduciary advisor merely because a prompt tells it to. It is better used to understand options, organize questions and test a plan than to blindly authorize transactions.

The notable lesson is not simply that AI can provide financial advice. It is that people with enough financial knowledge to ask complete, structured questions receive substantially better advice than people who may need the most help.

A useful financial AI should help close that knowledge gap rather than quietly make it larger.

Read “AI financial advice is surprisingly good, especially if you ask the right questions.”