AI Assistant Steers Retail Investors to Index Funds
Investors using a Mezzi AI assistant raised index funds’ share of monthly purchases by 7.6 percentage points and cut single-stock purchases by 5.9 points, an NYU Stern paper found.
Investors using an AI assistant shifted new purchases from individual stocks to index funds, according to a working paper from NYU Stern. Index funds’ share of monthly purchases rose by 7.6 percentage points after adoption, while the share allocated to individual stocks fell by 5.9 points.
The paper, “Generative AI and Retail Investment,” was written by finance PhD student Valerie Baldinger. It uses de-identified data from Mezzi, a U.S. wealth management app and SEC-registered investment adviser. Mezzi added an assistant using OpenAI’s ChatGPT and Google’s Gemini models in November 2024.
The dataset covers about 1,900 investors and 69,598 prompts matched with transaction and portfolio records. Users connected brokerage and retirement accounts to the app, allowing the assistant to consider their holdings. The assistant did not place trades; investors decided whether to follow its recommendations.
Index funds accounted for about 33% of purchases before investors adopted the assistant. Baldinger used a staggered difference-in-differences analysis to compare investors who began using it at different times. The paper estimates that adoption raised the index fund share by 7.6 percentage points and reduced the share of individual stocks by 5.9 points.
The shift was associated with a decline in estimated “return loss,” or the returns investors give up by holding diversifiable risk that the market does not reward. Annual return loss fell by 43 basis points from a pre-adoption average of about 407 basis points. Among the less-diversified half of investors, it fell by 101 basis points. The result for more-diversified investors was not statistically significant.
Investors mentioned buying an exchange-traded fund or mutual fund in 3.5% of prompts, while the assistant recommended a fund in 18% of its answers. When prompts named no ticker and did not mention stocks, 89.7% of buy recommendations were for ETFs or mutual funds.
The Vanguard Total Stock Market ETF, or VTI, was the most frequently recommended ticker for six of the seven ChatGPT and Gemini models examined. Within the Lipper Multi-Cap Core category, VTI accounted for 53.9% of investor ticker recommendations, compared with 15.8% of category assets. Across the models, the 10 most recommended tickers made up half of all ticker recommendations.
Investors bought at least one recommended security within 10 trading days in 36.2% of answers containing buy advice. The paper estimates that a one-standard-deviation increase in an answer’s personalization raised the probability of action by about 12 percentage points. The estimate was imprecise, and the analysis could not rule out other explanations, including model quality, user trust and access to advice on demand.
Mezzi’s instructions also affected recommendations. In mid-December 2025, the assistant was told to name specific tickers more often. Ticker recommendations rose from less than one-fifth of answers to about one-third. During the following four weeks, the fund share of new purchases by the platform’s heaviest existing users increased by 8.1 percentage points relative to non-users.
In January 2026, Mezzi instructed the assistant to avoid recommending tickers. Ticker recommendations fell to about 5% of answers, and the fund share of heavy users’ new purchases declined by 7.7 percentage points, returning close to its pre-December level.
Fund purchases after adoption had expense ratios about 3.6 basis points lower. Purchases also became more concentrated: the share allocated to the 20 most widely recommended funds rose by 7.2 percentage points.
The sample may not represent U.S. retail investors as a whole. Mezzi users had a median portfolio of about $556,000, compared with $155,000 for equity holders in the Federal Reserve’s Survey of Consumer Finances. Adoption was voluntary, and the paper has not been peer-reviewed.
Baldinger worked as an unpaid contractor for Mezzi and reported no financial interest in the company. The findings may differ for people using ChatGPT or Gemini directly because the Mezzi assistant operates under the platform’s instructions and can access account data. The study covered U.S. investors. European retail investors generally cannot buy U.S.-domiciled ETFs such as VTI because of European Union rules requiring a key information document for packaged retail investment products.








