Relying on AI for Financial Suggestions

Would you trust your portfolio to a machine? Shifting from thought experiment to a live debate is the question of whether sophisticated AI models could clear one of finance’s hardest hurdles of knowledge: the Chartered Financial Analyst Level III exam.

Several frontier systems aced a mock CFA Level III, the notoriously difficult test of portfolio management judgment conducted by New York University’s Stern School of Business and wealth-tech shop GoodFin in a joint study. OpenAI’s o4-mini scored 79.1% and Google’s Gemini 2.5 Flash notched 77.3%, both above the pass threshold of about 63%. For comparison, the CFA Institute says just about half of human candidates clear Level III in an average window.

Why AI Is Becoming Harder to Ignore in Finance

The analysis looked at 23 popular models and found a stark divide. About three-quarters of them clustered between 71 and 75 percent on multiple-choice questions, but ranged more broadly in their performance on essays that test structured reasoning, analysis, synthesis and professional judgment. Top “reasoning” models pulled away on this segment.

That makes a difference, because real financial advice on the whole doesn’t sound like trivia. It mixes rules, probabilities and trade-offs: balancing the tax consequences, risk budgets, time horizons and investor behavior. When an AI reveals its mettle on essays — not just answer keys — it allows us to see increasing competence in the murky middle where advice really dwells.

Advice Versus Answers in Real Client Settings

Yet passing an exam is not equivalent to serving a client. Advice is situational and regulated. In the United States, registered investment advisers are fiduciaries and brokers must abide by Regulation Best Interest; both sets of standards require suitability as well as disclosure and documentation that head off some math.

Even GoodFin’s chief executive has stressed the human element in understanding context and intent — skills that models still find difficult to consistently nail. Body language, perceptions of risk and life transitions do not make for strong essays either. That’s why the near-term trajectory looks like copilot, not replacement.

Where AI Can Already Help Everyday Investors

AI is already very good at things that reward scale and discipline. Think categorization of spending by expense, forecasting of cash flows, alerts when it’s time to rebalance and tax-loss harvesting. Robo-advisors have racked up trillions of dollars in assets under management worldwide, according to industry trackers such as Statista and Morningstar, by algorithmically handling diversified portfolios.

There is also evidence from classical studies — most notably, the work of Brinson, Hood and Beebower — to suggest that asset allocation explains the bulk of variation in returns between institutional portfolios. This is a viable task if you decide that allocation and cost control are all that matter and an AI machine that never forgets to rebalance and whose single mission is endless tax lot optimization can be a useful ally.

The Risks of Outsourcing Decisions to Artificial Intelligence

They can still hallucinate or overfit patterns that simply echo the past, or they might miss breaking news. AI got caught making things up about legal or tax citations; a reminder that confident tone is not necessarily credible advice. Mistakes in tax and estate planning have genuine, sometimes irremediable, costs.

Conflicts of interest also matter. If a platform is paid to steer you into certain products, the “recommendation” isn’t neutral. Regulators such as the SEC and FINRA in the U.S., FCA in the UK, and authors of the EU’s AI Act are raising expectations over explainability, suitability, and fair marketing. Anticipate increased scrutiny as models transition from education to advice.

How to Vet Artificial Intelligence Advice Before You Act

• Start with provenance. Who built the model and who tuned it, and on what data? Responsible providers will make clear which training methods they used, on what schedule of updating (rather than merely when), and any shortcomings they are aware of. If the system can’t explain its origins or assumptions, consider outputs as brainstorming rather than guidance.
• Demand traceable reasoning. A strong assistant will tell you in simple language what trade-offs are involved and also offer the tax and fee assumptions underlying recommended paths, comparing scenarios for stress tests. If you can’t see the why, you can’t judge the risk.
• Protect privacy. Stop pasting account numbers, entire statements, or personally identifiable information into public models. Prefer tools that analyze data locally or under traceable enterprise confidentiality controls, and check how long your information is kept.
• Apply a human backstop for decisions that alter your financial trajectory. An AI tool’s output should be sanity-checked against your goals, constraints, and behavior — areas in which empathy and accountability are important. A fiduciary adviser or a CFP can provide that sanity check.

Bottom Line: Treat AI as Copilot, Keep Humans Accountable

Trust is something that is built and AI is building trust faster than anyone thought it could. Passing an imitation CFA Level III hints toward reasoning; yet the span from “knowledge” to “advice,” it still travels so far along context, regulation and ethics. For now, the prudent approach is to treat AI as a powerful copilot — one who calculates numbers constantly for the seat of your pants with no need for sleep — while you and if necessary a human adviser provide judgment, accountability and a firm hand.