Vignesh August 4, 2026
The most useful finding in the AI-money research is not that AI is risky. It is that better prompts substantially improve the answer— that is from Choukhmane, de Silva and co-authors, the same team whose data shows AI advice helps on average. So part of the fix is simply in how you ask. Here are the three things to always tell it, because it usually will not ask.
1. Give it your take-home, not your gross
$120,000 gross in Massachusetts is about $7,100 a monthafter tax. Prompt an assistant with the gross number and every plan it builds sits on money you do not actually have. Lead with take-home, and name your state — tax is not a rounding error.
2. Tell it what happens to a payment when a loan clears
When your car loan ends, that $520 a month does not vanish — it has to be re-deployed, and where it goes changes your ten-year outcome. The model will pick a path silently unless you say it out loud. Add one line: “when the loan clears, keep contributing the freed payment.” Otherwise the two options it compares may not be spending the same money.
3. Ask it to check feasibility before it plans
Add: “given my take-home and fixed costs, is this monthly even possible — and if not, say so.” A plan that assumes a monthly you cannot sustain is a wish with a chart on it. Make the model prove the number is livable before it projects thirty years of it.
The uncomfortable part
Notice that all three are questions a good human planner would ask you. The reason to hand them to yourself is fairness: the research found the people who gained the least from AI were those with the least financial fluency — the ones least likely to know to add these lines. The better you can prompt, the better the answer, which is exactly backwards from who needs the help most.
AI is genuinely good at explaining a plan once the inputs are honest. Give it those three and you will get more out of any assistant. It is also the reason we bake all three into InvestEd— take-home tax, the freed-payment decision, and a feasibility check — so the good answer is not gated by how well you happened to ask.
This is about how to ask, not what to decide. It makes no recommendation about your specific situation.

