The human advisors gave the portfolio a once-over and offered a simple qualitative assessment, while the AI offered a more detailed analysis, including a breakdown of the different sectors represented and the income Dotty receives from each of her dividend stocks. This more specific response reassured Talcove that he was on the right track.
“I have much more confidence in AI,” says Talcove, CEO of government business at LexisNexis Risk Solutions, who employs two major financial firms he declined to name to manage his own money. Not only was AI’s assessment more thorough, but it also was free of potential conflict. “The money managers are like, ‘Let me manage it for your mother,’” Talcove says.
As advances in artificial intelligence promise to democratize financial advice, anyone with a computer or a smartphone has access to the kind of models and recommendations that used to cost a lot of money—or, at least, a lot of time making spreadsheets.
This progress comes as AI investment has fueled massive gains in the stock market. Big players like Meta Platforms and Microsoft have committed to spending billions more in the coming months, even as OpenAI CEO Sam Altman mused in late August about an AI bubble.
While the upshot for markets remains under debate, large language models available through apps like OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude have been streamlining all manner of tasks, from writing wedding toasts to designing and managing a retirement portfolio. A recent survey from robo-advisor Betterment found that just over half of respondents reported using AI for some kind of financial-related purpose at least once a month, although only 30% said they trusted AI to give them actual financial advice.
The remaining 70% are on to something. For now, entrusting your financial life to AI would be a risky move. Large-language models are good for brainstorming and troubleshooting, but there’s still no substitute for an experienced human who can review your entire financial life, spending goals, and portfolio and then bundle it all together to create a financial plan.
Don’t expect AI models available to the public to be much use in picking stocks, either. “The dream scenario is to let AI create a model portfolio to beat the S&P 500 and make you rich,” says Sam Dogen, founder of the Financial Samurai website and author of Millionaire Milestones: Simple Steps to Seven Figures. “I don’t think that will happen yet.”
The promise of AI in managing your finances is undeniable—as are the perils. Here’s a look at how you can harness the technology to work for you:
How to Use AI Effectively
Today’s large language models are like that kid back in high school who loved to hear himself talk. No matter what the teacher asked, he’d raise his hand and give a confident answer. Often he knew what he was talking about, but sometimes he was just plain wrong. The fine print at the bottom of ChatGPT warns, “ChatGPT can make mistakes. Check important info.”
These large language models, or LLMs, are a type of AI that’s trained on vast data sets, hoovering up text in order to generate plausible answers to the questions users pose. “Generative AI” refers to when a chatbot creates original content from the mash-up of texts it has digested. While they may seem sentient—some people have “dated” chatbots or used them as therapists—they don’t have the feelings, morals, or a sense of personal responsibility you could reasonably expect to find in a financial advisor. Nor are they required to act as fiduciaries—legally bound to act in users’ best interests when offering financial advice—the way registered investment advisors are.
“What I think people forget is these LLMs are word calculators,” says Eric Ludwig, director of the Center for Retirement Income at The American College of Financial Services.
When it comes to investing basics, these models generally spit out sound advice. Large language models are only as good as the questions they’re fed, so mistakes at this level are largely errors of omission, like not including your age when asking about retirement planning. (That’s why AI “prompt engineer” has quickly become a hot job category.) A beginning investor might need to ask follow-up questions (“What if I already own a rental property?”) to arrive at a more accurate, personalized answer.
Large language models’ skill at summarizing and synthesizing can offer retirement investors a good way to get started, or a gut check on an existing plan. Tell ChatGPT that you’re 20 years from retirement, and it will suggest an appropriate mix of 50% to 60% U.S. stocks, 20% to 30% international stocks, 10% to 20% bonds, and zero to 10% alternative assets or real estate investment trusts, or REITs. Then it will recommend specific low-cost index funds to build the portfolio, like the Vanguard Total Stock Market exchange-traded fund and the iShares Core U.S. Aggregate Bond ETF.
Chatbots will also take a crack at projections. Until recently, you’d need to hire a financial advisor to run software to determine the probability of your money lasting throughout your retirement. Now, you can tell ChatGPT that you have a $1 million nest egg at age 65, and it will tell you there’s a 40.7% chance that your money will last until you’re 100, assuming $50,000 annual withdrawals adjusted for 2.5% inflation and 6% market returns.
But the advice will stop there—unless you ask for more. If you ask ChatGPT about the pitfalls in its analysis, it will point to the dangers of a stock downturn early in retirement, the wild card of long-term care needs, and other potential threats. It then offers to stress test its simulation against high inflation years, a period of bad market returns, and unexpected expenses—an exercise that any human advisor would do without prompting.
Released in late July, ChatGPT’s new “study mode” does a better job of asking questions to help users arrive at a personalized response. It promises to help users work through problems step by step rather than spitting out an answer. If you give it the same prompt about having a $1 million nest egg at age 65, it will start with asking questions about your asset mix and expected longevity.
The bottom line? If used thoughtfully, AI can help you set up a retirement plan, stress test it, shed light on any blind spots and even talk you off the ledge during market turbulence. “It can act like a great financial coach,” says Dogen, the Millionaire Milestones author, who lives in the Bay Area and invests in AI in addition to experimenting with it for his own portfolio.
Good for Research, Not Stock-Picking
Stanford researchers recently showed the promise of AI as a stockpicker, but only in the right hands. In a study released in June, their AI analyst outperformed 93% of active fund managers by an average of 600% from 1990 to 2020 using only publicly available information.
But they didn’t use an off-the-shelf large language model. Instead, researchers built a sophisticated Random Forest machine learning model to do the hard work of translating reams of Excel spreadsheets, earnings transcripts, regulatory disclosures, and investor presentations into actionable stock picks.
Institutions retain an edge in harnessing AI’s power to synthesize information, says Ed deHaan, a professor of accounting at the Stanford Graduate School of Business and a researcher on the project. Whatever retail investors can do, institutions generally have the resources to do it better.
“It’s tempting to think that generative AI will level the playing field, but that’s not realistic,” deHaan says.
While AI might not help you outperform the pros, it can help you evaluate investments. In fact, equity research is where the technology really shines. Felix Xu, a general partner at ZX Squared Capital, a hedge fund, used to build spreadsheets in his former job as an equity analyst. ChatGPT now makes it so much easier to track metrics like company profits and historic price-to-earnings ratios, and to summarize earnings transcripts.
Many people just want a quick answer of where to invest, Xu says. But rather than asking a large language model for stocks to buy and blindly following the recommendation, ask it to analyze companies’ challenges and key competitors, he says. You could even ask, “If you’re an analyst, what would you look for?” Xu says. “You should always keep challenging ChatGPT.”
To reduce the possibility of errors, Xu uploads actual company documents into large language models for them to analyze, rather than relying on the LLMs to compile them.
Do Your Own Math
Because they’re trained on words, large language models can fall short when attempting math. Ludwig asked ChatGPT and Claude to calculate the quarterly tax payments of a married couple with capital-gains taxes under Republicans’ new tax and spending law. The two calculations came out very different from one another. Andrew Lo, a finance professor at MIT, recently discovered that a large language model he declined to name didn’t know how to calculate compound interest when he asked it to compute the accrued interest on a loan for a nuclear power plant.
While ChatGPT is often used as a synonym for generative AI—like Kleenex or Xerox are for tissues and photocopying, respectively—it’s just one of several large language models. Each has different strengths and weaknesses, Lo says: “Almost like personalities, you need to get to know them.”
In a recent Barron’s test, ChatGPT and Gemini responded differently to a request to calculate required minimum distributions for a 73-year-old who has accounts across three brokerages. The RMD is the government’s way of finally getting its share of tax-deferred retirement savings, and it’s based on an investors’ balance across all of their tax-deferred retirement accounts. Financial firms will calculate the amount for their clients, but they only know about the money that’s with them.
ChatGPT offered to do the RMD calculation based on uploaded statements, but Gemini demurred. “I cannot directly calculate your Required Minimum Distribution (RMD), because I cannot securely access or process your financial statements. Please do not upload them or share any personal financial information,” Gemini said, offering instead to provide the formula for the user to do the math themselves.
It’s best not to put personally identifying information into LLMs, experts say. OpenAI’s Altman warned this summer that ChatGPT doesn’t offer users any legal confidentiality, and that its chats can be subpoenaed.
OpenAI, maker of ChatGPT, didn’t respond to a request for comment on the RMD illustration. A spokeswoman for Google, parent of Gemini, said users shouldn’t rely on Gemini’s responses as medical, legal, financial, or other professional advice.
Robots may one day operate on patients independently, or create and manage investment portfolios on their own. Lo predicts that within five years, so-called agentic AI will be able to act in a fiduciary capacity, managing users’ money in their best interests. But safeguards aren’t in place yet.
For now, most brokerage firms employ AI primarily for support tasks like taking notes at client meetings. But we aren’t too far off from a “hybrid” future where AI agents work alongside advisors, says Parker Ence, CEO and co-founder of Jump, an AI assistant for advisors. The goal, he adds, is delivering an AI experience that’s safe and compliant.
“In general, the technology is moving faster than regulations,” says Michelle Bonat, chief AI officer of AI Squared, a firm that helps companies integrate AI into their workflows.
Beyond the concerns with accuracy, privacy and legal liability, there are plenty of gray areas where experienced advisors can add value. A major job of a financial advisor is making sure a client follows through on the advice, such as executing a document or rolling over a 401(k). “A lot of your AIs will say, ‘this is what you should do.’” says Steve Parrish, a professor of practice at The American College of Financial Services. “A lot of retirement decisions aren’t yes or no.”
Write to Elizabeth O’Brien at elizabeth.obrien@barrons.com