The Same AI Gave Two People Financial Advice. One Prompt Was Worth $50,000 More.

If you've asked an AI tool for financial advice and gotten something vague back — "consider diversifying," "consult a professional" — the problem probably wasn't the model. It was what you told it.

What MIT Sloan found

Researchers at MIT Sloan published a study on August 1 comparing AI-generated financial advice against the structure of the prompts that produced it. The headline finding: AI advice was "surprisingly good" when a user supplied a structured prompt — age, income, savings, specific goals. Advice built on that kind of input held up well against outcomes a real financial planner would recommend.

The gap was in who provided that structure. Advice generated from prompts written by men or by users with stronger financial literacy produced simulated outcomes worth roughly $50,000 more by age 60, compared with advice generated from prompts written by women or less experienced users. Same model, same tool, same access — the input determined the output, and the people least likely to already know how to write a detailed prompt got the worse one back.

The skill this exposes

This isn't a story about the model being unfair. It's a story about what happens when a tool rewards a skill — writing a complete, specific prompt — that nobody taught as a skill. A vague question gets a vague, hedged answer. A structured one gets something you can actually act on. The model isn't withholding quality from anyone; it's mirroring back whatever specificity it was given.

Monday morning, the fix is a checklist, not a personality trait. Before you ask AI for advice on anything with real stakes — money, a hiring decision, a contract term — write down the facts a competent human advisor would ask you for first, and put them in the prompt before you ask the question.

Two people, two prompts

Renata manages payroll and benefits for a 60-person logistics company and asked ChatGPT for advice on rebalancing her personal 401(k) after a job change. Her first prompt was "how should I invest my 401k rollover" — she got back a generic list of index-fund categories and a suggestion to speak to an advisor. On a colleague's advice, she rewrote it: her age, current balance, employer match, risk tolerance, and the specific date she needed to decide by. The second answer named allocation ranges tied to her actual timeline, not a generic one. Nothing about the model changed between the two prompts. What changed was that the second one gave the AI something to reason about.

Contrast that with Dev, a first-year associate at a regional accounting firm, who asked an AI tool to draft investment guidance for a client meeting the next morning. He included every number the client had sent him — but not the client's actual goal, which was funding a child's education in eight years, not general retirement savings. The advice came back competent and well-organized, and wrong for the client's real timeline, because the one fact that mattered most wasn't in the prompt. Complete data isn't the same as relevant context; a prompt needs both.

The takeaway

The $50,000 gap in MIT Sloan's study wasn't a gap in the AI. It was a gap in who already knew how to ask it a complete question — and that's a skill anyone can learn before their next high-stakes prompt, not a trait anyone is stuck with.