When AI Gives Investment Advice, Who Checks the Maths?

By Ildar Farkhshatov, Co-Founder and Chief Technology Officer at IVM Markets

Earlier this year, at a conference in Paris, I came across a company comparing investment themes recommended by different AI models. The answers varied, sometimes considerably. Each model had a perfectly reasonable explanation for its choices.

I was more interested in what happened when you checked the numbers.

Researchers had been asking similar questions a year earlier. A 2025 study presented at ACM’s International Conference on AI in Finance found model-specific biases in investment analysis. Some persisted even when the models were shown contradictory evidence. [1]

Investors are adopting the technology anyway. In eToro’s 2025 survey of 11,000 retail investors across 13 countries, 19% said they used AI to select or adjust investments, up from 13% the previous year. [2]

I understand why. Financial research takes time, and AI makes a great deal of it accessible to people who would otherwise need a Bloomberg terminal and several years of experience.

A Good Story Can Sell a Bad Investment

Consider a situation from my everyday work.

A wealth adviser needs a structured product for a client seeking regular income, conditional downside protection and a three year maturity. A bank offers a note with an attractive coupon.

The adviser asks an AI model to assess the underlying stocks and explain why the investment might suit the client. Within seconds, the model produces a polished argument. The companies have credible growth prospects, the coupon looks appealing, and the downside protection sounds reassuring.

So far, so good.

Then another bank comes back with a competing note. Its coupon is lower, which makes the first offer look like the better deal.

Except the higher coupon may come with exposure to losses that matter more to this particular client. Under the scenarios most relevant to their circumstances, the supposedly less attractive note might produce better outcomes.

Now the adviser has a real decision to make. Ask AI which product to recommend, and it may find a persuasive argument for either one. Financial markets have never suffered from a shortage of persuasive arguments.

The question is what happens when you put the two products side by side and run the numbers. Regulators have reasons to care. ESMA has warned about bias and opaque decisions in AI-assisted investment services. MiFID II requires client-specific suitability assessments. In the US, SEC staff guidance addresses the consideration of reasonably available alternatives under applicable standards of conduct. [3–5]

If a client challenges the recommendation later, a polished AI explanation will be a poor substitute for evidence of why that product was chosen.

Let the Numbers Go First

Before AI offers an opinion, the adviser needs to establish which notes meet the client’s needs. A quantitative engine can compare eligible products across market scenarios, using their payoff terms and the prices quoted by issuing banks, with fees accounted for. The measures should reflect the client’s priorities within the firm’s risk framework.

Now AI has something useful to explain: where the extra income comes from and what the client gives up to get it.

Another analyst should be able to run the calculations and reach the same result. But I have seen AI confidently argue for an investment when the numbers pointed in the opposite direction.

I know what to look for. Someone new to financial modelling may not. Some days, working with AI feels like detective work: you know something is wrong, but finding exactly where it went wrong takes longer than getting the original answer.

Of course, reproducible rubbish is still rubbish. Bad market data or a convenient volatility assumption can make a poor trade look cheap. Keeping the original inputs and model settings allows someone else to rerun the analysis and challenge the conclusion.

Show the Alternatives

The client also needs a way to question the choice. Imagine an interactive chart plotting modelled returns against losses for the suitable notes. The recommended product appears alongside its competitors, with the scenarios available for inspection.

If another note shows a higher return and a lower modelled loss, the adviser has some explaining to do. Perhaps issuer credit risk matters, or the recommended note behaves better in a scenario the chart does not capture. Fine. Show the evidence.

A client may accept the maths and still prefer a different trade-off. They may buy nothing. That is their decision to make. The chart should help them interrogate the advice, rather than decorate an AI-generated sales pitch.

The Test Comes Later

I work with quantitative models and AI in structured investments. Language models can make specialist analysis accessible to advisers and clients who would otherwise struggle to interrogate it. That is worth pursuing.

But the firm or adviser remains responsible for the recommendation. [3] When a trade loses money, the question should be whether the decision made sense with the information available at the time. Were the assumptions reasonable? What suitable alternatives were considered? Can someone reconstruct the analysis after the adviser has moved on and the AI model has been replaced?

Markets can punish a sound decision. They can also reward a terrible one. Neither outcome tells you much about the quality of the original advice.

AI will keep getting better at explaining investments. Financial firms should get equally serious about checking the maths behind those explanations.

References

1. Lee, H., et al. Your AI, Not Your View: The Bias of LLMs in Investment Analysis. 6th ACM International Conference on AI in Finance (ICAIF ’25), 2025. DOI: 10.1145/3768292.3770375. Open-access manuscript: arXiv:2507.20957.

2. eToro. Retail investors flock to AI tools, with usage up 46% in one year. 1 October 2025.

3. European Securities and Markets Authority. ESMA provides guidance to firms using artificial intelligence in investment services. 30 May 2024.

4. European Securities and Markets Authority. MiFID II, Article 25: Assessment of suitability and appropriateness and reporting to clients. Interactive Single Rulebook.

5. U.S. Securities and Exchange Commission. Staff Bulletin: Standards of Conduct for Broker-Dealers and Investment Advisers—Care Obligations. 2023. See “Considering Reasonably Available Alternatives”.