One Size Fits None: Do LLMs Provide Suitable Financial Advice?

Large language models (LLMs) are increasingly used for financial advice, yet their ability to provide suitable recommendations remains unclear. Financial suitability -- a core fiduciary requirement -- demands that financial advisors tailor recommendations to each client's unique circumstances. We evaluate whether LLMs can satisfy this standard by prompting GPT 4o with 1,000 synthetically generated client profiles spanning diverse demographics, financial situations, and goals. Using reverse-engineered machine learning models, we find that LLMs rely on simple heuristics and act as decision trees when recommending investments. Equipping LLMs with agentic tool use partially mitigates this issue but creates a diversification paradox: while portfolios have better cross-asset diversification, recommendations homogenize across clients. Our preliminary findings suggest that LLMs may not satisfy suitability requirements, raising the question of what guardrails would enable LLMs to meet fiduciary standards.