One advisor for the whole world?
Digital edition · November 2026 · working paper

One Advisor for the Whole World?

Cross-Country Evidence on Financial Advice from AI

Abstract

Households around the world now turn to AI models for financial advice. I pose three financial questions to frontier models in twenty-one countries to understand if that advice differ across countries and languages. AI advice on portfolio allocations varies little across countries, but varies more if local economic conditions warrant. Explanations invoke local financial institutions without changing allocations. Models apply remarkably similar investor-assessment rules across countries, but only selectively incorporate household balance sheets. The findings suggest that AI is globalizing the supply of financial advice and the rules used to assess investors, while selectively adapting recommendations to national financial environments.

JEL: · Keywords:
Proceed to Introduction →
The result in one chartMean recommended share by model and country

Each row is one of the thirteen models and each dot is one of the twenty-one countries: the mean recommendation that model gives when asked the question in that country's language, with the country named (Version B, age 40, fifty repetitions per dot). The black tick is the model's mean across countries. On the portfolio question the dots of a model sit close together while the rows sit far apart: the model matters far more than the country. Switch to the mortgage question and the dots spread out. Click a dot to follow one country across models.

How to read this site

  • The paper is split into the pages listed on the left, in the paper's own section numbering. Figures are drawn from the estimates and can be switched to a table or downloaded. Tables are the paper's tables.
  • Footnotes open inline when you click the small numbered buttons. Citations link to the reference list. Equation, figure and table numbers match the PDF.
  • The country effects explorer shows every country contrast for the three questions, the same contrast model by model, the leave-one-model-out estimates, and the cell means behind the language, local-context, age and stated-investor comparisons.
  • The prompt explorer shows the exact text every model received, in every country, version and question, with the DeepL back-translation for the non-English prompts.
  • The model tracking page sets the July 2026 panel beside the September 2026 wave of twenty-one models, fourteen of them new, on the same English slice of the primary question.
  • The raw model responses are not public. The data and code page says what the site does and does not contain, and gives a prompt for asking an AI assistant about the paper.