by AdvisorAnalyst.com Editorial Team
Rawley Z. Heimer, an economist at Arizona State University, just put together a massive review of household finance research. And here's the interesting part: he's not just cataloging the usual behavioural puzzles, he's asking how generative AI is about to scramble all of them.
Heimer calls his subject "financial decision making, at the dawn of AI,1” basically how regular households use markets to handle money over time, saving, investing, borrowing, the works. The textbook version says people should smooth their spending, diversify their portfolios, and manage debt efficiently. In practice? People miss the mark constantly. And Heimer is careful to point out this isn't always about irrational behaviour. Sometimes it's just friction, transaction costs, confusing paperwork, advisors with their own agendas, markets that don't offer the right products.
Three Problems That Won't Go Away
On the savings side, there's a trio of stubborn puzzles. Only 54 percent of American households actually hold retirement accounts, according to 2022 Survey of Consumer Finances data. Then there's the asset de-accumulation puzzle: retirees sit on their savings way longer than the models say they should. And almost nobody buys annuities, even though the math says they probably should. Heimer traces these patterns back to things like hyperbolic discounting (basically, people procrastinating on their future selves), sticky defaults, the urge to leave money to kids, and a weird tendency to underestimate how long we'll actually live.
Investing tells a different story. Stock market participation jumped from about 32 percent of households in 1989 to nearly 58 percent by 2022, thanks largely to commission-free trading apps. But here's the catch: once people are in the market, plenty of them trade like they're at a casino instead of building a diversified portfolio. Overconfidence, chasing whatever's trending, hunting for lottery-style payoffs, Heimer walks through all of it.
And then there's debt. The co-holding puzzle is a classic head-scratcher: people carry expensive credit card balances while sitting on cash that's earning next to nothing. Doesn't make sense on paper. Add to that some genuinely bad repayment habits, like paying down balances proportionally instead of attacking the highest interest rate first, and you've got a debt landscape full of avoidable mistakes.
Now Throw AI Into the Mix
This is where the paper gets really interesting. Heimer breaks the AI story into three threads.
First, AI seems to be cutting down on information overload. When Italy temporarily banned ChatGPT, researchers noticed investors got more cautious, fewer new positions, more concentrated bets. That suggests AI access was actually helping people process information better. Pair a robo-advisor with an LLM chatbot, and you get better diversification and people actually following the advice.
Second, and this is the part advisors should pay close attention to: AI advice carries its own kind of bias, one that's "programmable." LLMs mostly give sound financial guidance, but they also make weird assumptions, like overestimating how patient people are, and the recommendations shift depending on which model you're using, how you phrase the prompt, even how the AI was trained to behave. In other words, the bias isn't fixed. It's a design choice, whether intentional or not.
Third, not everyone's benefiting equally. Heimer dug into 2025 SHED survey data himself and found that GenAI users initially looked way better off financially. But once you control for income, job type, and industry, most of that advantage disappears. Only the equity-participation bump holds up, and it's small. His takeaway: this is selection, not causation. People already doing well are the ones using AI, not the other way around.
Five Practical Things Advisors Can Actually Do With This:
1. Use AI to prep, not to decide. Let it help you research, summarize reports, or draft client communications faster. But the actual recommendation, the thing you tell a client to do with their money, should still come from you. Heimer's research backs this up: people trust AI more when a human reviews it first, and that trust doesn't come at the cost of worse advice.
2. Don't assume every client benefits the same way. The clients already comfortable with tech and already doing well financially are the ones picking up AI tools fastest. That means AI risks becoming another advantage for people who already have advantages, not a great equalizer. If you've got clients who are older, less tech-savvy, or just skeptical, don't assume they're getting the same informational edge from AI that your younger or wealthier clients are. You may need to bridge that gap yourself.
3. Spot-check AI output before it reaches a client. Different AI tools, even different versions of the same tool, can give noticeably different financial advice depending on how they were built or trained. If you're using AI to draft anything client-facing, don't just copy-paste it. Read it like you'd read a junior analyst's work: useful starting point, but it needs your judgment on top.
4. Watch out for AI-generated scams and fake content. It's now cheap to produce convincing fake testimonials, fake news clips, or fake "expert" endorsements pushing bad investments. If a client mentions something they saw online that sounds too polished or too persuasive, that's worth a second look before they act on it.
5. Use AI conversations as an opening, not a replacement. Some people feel more comfortable asking an AI "dumb" financial questions than asking a human advisor, less judgment, less awkwardness. That's actually useful information. If a client's been using AI to explore questions they were embarrassed to ask you, treat that as an invitation to have the real conversation, not a sign they don't need you anymore.
Footnote:
1 Heimer, Rawley Z. Financial Decision Making, at the Dawn of AI. Arizona State University, 16 June 2026. Manuscript prepared for submission to Oxford Research Encyclopedia of Economics and Finance.
Copyright © AdvisorAnalyst
