One advisor for the whole world?

1Introduction

Retail investors are increasingly turning to artificial intelligence (AI) for financial advice, through chatbots built on large language models (LLMs) (OECD 2026). The share of U.S. adults who use AI tools specifically for financial advice roughly doubled between 2024 and 2025 (Financial Health Network 2025), and over half of U.K. adults now use such tools to help manage their money (Lloyds Banking Group 2025). This represents a fundamental shift in the nature of financial advice for households all around the world, from national markets in which advice reflects local norms and institutions to a global market where advice comes from a handful of foundation models trained overwhelmingly on U.S.-English text. This shift raises several questions about the new global financial advisor: What financial advice do AI models give in different countries and do they tailor the advice to local circumstances?

To answer these questions, I use a pre-registered experiment where thirteen frontier AI models from six vendors answer three financial problems posed in twenty-one countries. I ask the questions in the local language, in English, and with local context added, and also vary the characteristics of the investor asking for advice to identify the characteristics that the models respond to. The findings speak to the supply side of financial advice from AI by identifying how models respond to the same questions in a variety of different contexts.

My first finding on the new supply side of financial advice is that AI financial advice is largely the same everywhere for a simple portfolio choice question, but that questions requiring more local context generate more dispersion. Asked how to split the local-currency equivalent of $50,000 between a broad stock-market index fund and government bonds, the models give advice that differs only modestly across countries. Compared to the United States, every significant contrast is negative, ranging from 2.7 percentage points less equity in Mexico to 11 points less in Brazil. What instead moves the advice is the identity of the model, which accounts for 69.5 percent of the variance against 5.7 percent for country. Asked whether to invest savings or use them to prepay a 4 percent mortgage, the models recommend investing 10 to 29 percentage points less than in the United States in each of the fifteen countries where that rate is realistic. The largest gaps are in Japan, China, Switzerland, and Germany. When the rate is removed from the question, the models assume local mortgage rates themselves, the gaps in the low-rate countries largely close, and the high-rate countries move toward prepaying. Asked how to protect savings from 3 percent inflation, the models switch from equities to inflation-linked bonds in Brazil, Turkey, and Mexico, but overwhelmingly recommend equities in all other countries. On these two questions model identity explains only 15 and 26 percent of the variance. The pattern suggests that the models localize when the question depends on the local economy.

The lack of cross-country differences raises the question of what, if anything, the advice should track. Giving the same advice everywhere on what share to invest in stocks is not a mistake per se: under integrated markets the risky portfolio is the same world index everywhere and uniform advice is warranted. Of course, there are actually some differences across countries in the advice, so the models are not just giving exactly the same advice everywhere. As a simple (but not realistic) benchmark, a Merton benchmark with a currency-hedged world fund prescribes 0.52 in every country, within about six points of the advice everywhere. However, this model cannot explain why some countries receive less equity than the United States. A more realistic benchmark would take local pension systems and background wealth into account, but as I show later, the models do not measurably adapt their advice based on these variables. An alternative approach to understanding what models track across countries is to relate cross-country differences to country-level fundamentals. To test this, I relate the estimated country fixed effects to country-level risk-taking from the Global Preferences Survey (Falk et al. 2018), to financial literacy from the S&P Global FinLit Survey (Klapper et al. 2015), and to fourteen other country-level fundamentals. However, these country-level fundamentals have only limited explanatory power: country differences in financial advice on the risky share do not track measured risk preferences, financial literacy, local market returns, income levels, equity volatility, the interest rate, or the Merton benchmark. Advice on the mortgage and inflation questions tracks the ten-year yield and the local mortgage rate, again showing that the models can incorporate local context if it is relevant.

The next finding is that AI models give the kind of “popular financial advice” studied by Canner et al. (1997) and Choi (2022). The evidence for this comes from several different tests. First, I insert statements about the investor into otherwise identical English prompts for each country, covering risk tolerance, currency hedging, pension composition, pension generosity, the share of her liquid wealth that the stake represents, and investment horizon. Stated risk tolerance raises the recommended equity share by 46 percentage points, consistent with the dominance of risk tolerance in Ross and Lo (2026). A twenty-year rather than one-year horizon raises the equity share by 40 points, in contrast to the classic Merton–Samuelson irrelevance result where horizon should not affect the share at all under CRRA preferences with i.i.d. returns. Second, I ask the models to list the questions they would put to the investor before advising. In every country, the models ask the same four things: risk tolerance, what the money is for, whether the five-year horizon is firm, and whether an emergency fund exists. Every response except one asks about risk tolerance, and three quarters of the responses ask all four. These questions are reminiscent of popular financial advice, and it is striking that there are only negligible differences in the main questions the models ask across countries, but the local content shows up as currency and inflation risk in emerging markets. The models ask less often about the investor’s pension, or about local circumstances, both of which would be highly relevant in, for example, life-cycle models.

I then conduct several checks that directly ask how the balance sheet affects the advice. A stated generous pension adds safe background wealth, so it should raise the equity share if the model integrates the investor’s balance sheet. However, the advice barely moves when information about pensions is added to the prompt. This is not because the models ignore pensions, as nearly every explanation mentions the investor’s existing pension. In additional tests, added after the pre-registered results, I find that the advice moves strongly and in the predicted direction when a stated balance-sheet cue raises the investor’s exposure, as with a pension invested in stocks or a stake equal to her entire liquid wealth, but it barely moves on cues that signal safety, such as a guaranteed annuity, a currency hedge, or a cash cushion. The models appear to read the balance sheet as a source of hazard instead of seeing the balance sheet as a two-sided input, as portfolio theory would prescribe. Overall, AI advice fits popular financial advice more than life-cycle models that incorporate the household’s balance sheet.

The final finding is that the models appear to solve the portfolio problem in a shared, English-anchored representation regardless of the prompt’s language, and then explain the answer in the local language and with local context. This is consistent with the interpretability literature, which finds that multilingual models solve tasks in a shared, English-leaning concept space (Wendler et al. 2024; Tang et al. 2024; Wu et al. 2025; Lindsey et al. 2025). One response in Swedish, a language I speak, describes a “glidväg” (‘glide path’), a direct translation of the English industry term with no particular meaning in Swedish. More formally, I machine-code the explanation that each model gives for its recommendation into four binary indicators. Local financial context appears in 18 percent of U.S. explanations but 29 to 78 percentage points more often in every other country. Naming local instruments in the prompt raises local references by 23 percentage points, but moves the allocation by less than a point. The translation changes the words of the answer but not the advice, suggesting that the explanations localize far more than the advice itself.

These results hold up across a set of pre-registered and exploratory robustness checks. For example, I use a VPN to change the location where the prompt is sent from, which does not change the result. Sending the prompts through a U.S. rather than a German VPN exit moves the advice by 0.4 points. I also adjust the stake by purchasing power raises the share by a marginal one point and accounts for a sixth to a third of the Brazil, Poland, and Mexico contrasts, where the adjustment is largest. The results are robust across a number of similar checks, I describe more of these robustness checks when discussing the main results. The interested reader can go to this website to explore the results themselves and see, for example, how the results change by model generation or what the prompt was in each country. You can also download the code for running your own experiment. With papers based on AI-generated data, the question is always whether the results will hold for future model generations. I argue that the structure of the advice will, for two reasons. First, the portfolio question has no clear answer. Models can be trained to ignore gender or race, because these should not affect a financial recommendation. It is much less clear how country or language should affect the optimal equity share. The answer depends on theory, where there is no clear recommendation, and on preferences that the prompt does not state. What the models learn instead is the body of written financial advice, which changes slowly. The departures from theory that Canner et al. (1997) documented remain in the popular advice that Choi (2022) surveys twenty-five years later. Second, the results do not reflect a lack of knowledge that a more capable model would fill. The models assume the local mortgage rate, switch to inflation-linked bonds where inflation is high, and cite the local pension system in their explanations. They have the local information and do not use it on the portfolio question. The evidence so far supports this view. A panel of twenty-one models collected eleven weeks after the main collection, fourteen of them new, reproduces the variance split, with 70 percent for model identity against 5 percent for country, and 77 percent of its non-U.S. cells lie below the same model’s U.S. cell. The fourteen new models in that panel differentiate more across countries, but not in the direction of the benchmark. This argument concerns the advice rule and not the level of advice, which differs across models and can move between versions of the same model.

Summing up, this paper characterizes the supply side of an emerging global market for AI financial advice. Existing research primarily examines the quality and determinants of AI financial advice within particular countries, models, or investor populations. This paper instead examines its geographic structure across countries, models, and financial decisions. The findings reveal a tension between standardization and localization. AI models apply broadly similar investor-assessment rules across countries, asking the same questions and responding similarly to investor characteristics wherever the investor is located. Yet their recommendations incorporate local economic conditions to very different degrees across financial decisions. The explanations are more geographically differentiated still, frequently invoking local financial institutions even when the recommended allocation barely changes. These findings suggest that we are moving to one advisor for the whole world, with AI models globalizing not only the supply of financial advice but also the rules used to assess investors, while selectively adapting the recommendations to national financial environments.

Historically, financial advice has been a nationally segmented service industry, shaped by regulation, relationships, and language (Reuter and Schoar 2024). AI is changing this market by making advice from a handful of foundation models available across countries and languages. This transformation not only means that investors everywhere have broader access to financial advice, but also that the rules used to evaluate investors are becoming global. Standardization is not necessarily undesirable, but the concentration of advice production in a small number of models creates the potential for common recommendations and common errors across households in otherwise distinct national markets.