As AI becomes more embedded in wealth management, explainability is no longer optional. Daniele Grassi, CEO at Axyon AI explores why transparency is needed in how AI models reach their conclusions.
In high-stakes, professional environments like medical, legal and financial services, characteristics such as logic, fairness and accountability are not just desired but often mandatory to maintain regulatory compliance.
If we want to demonstrate conviction in our opinions, we cite reliable sources to back up our arguments. If we need a watertight rationale for making a recommendation, we set out the supporting points stacked up against any alternatives to make the strongest, most objective case possible. And heaven forbid should anything ever go wrong: at best we should understand the process in order to prevent it happening again; at worst we want someone to blame.
This is true of human-based decision-making and extends as comprehensively across tools using artificial intelligence (AI).
The different elements that contribute to this clarity and how much weight each holds will vary depending on your role and area of responsibility but they all reduce friction across any wealth or investment business.
Investment committees need to understand the logic behind any AI-driven insights; compliance teams need evidence and auditability to strengthen their paper trail; and sales, marketing and relationship managers need clear explanations of what they are communicating to clients and advisors. Explainability acts as a common language between these different components of your business and partners with whom you’re working – from the quantitative and investment teams, to risk and client-facing functions.
Long before today’s generative AI existed, algorithmic trading and quantitative hedge funds ballooned in popularity in the run up to the financial crisis. As many suffered huge liquidations, it exposed how many competing strategies held identical positions. The Quant Crisis of August 2007 (and other market interrogations that have happened since) highlighted the need for greater transparency and is one of the reasons investors became cautious over relying too heavily on so-called ‘black box’ strategies.
Of course, when things are going well, people care less about how well their risk mitigation is articulated. Most investors will typically welcome solid outperformance with a less-intrusive grilling on where it has come from. But when underperformance occurs, explainability becomes essential to build trust, document the decision-making process and demonstrate the drivers on which a model’s outputs are grounded.
It’s important to understand that signals typically informing decisions are not static or reliably predictable. Outcomes (and therefore the reasons that sit behind them) are at the mercy of multiple factors, so it’s important to not just think about what a model is doing but how it’s evolving, any signals the market is sending and how an AI model is learning from all of those inputs.
In a world where information transparency and disclosure now form an essential part of any financial services value exchange, it’s critical to not just understand what AI is telling us to do but why it has given the option it has. Not just giving a prediction but casting light on the drivers behind the prediction to help us understand how it got there is reassuring, if nothing else.
A human portfolio manager might favour an asset because of valuation, momentum or macroeconomic conditions. This may depend on internal factors like their management style or external factors, such as market conditions. What informed these preferences? Was it fundamentals, sentiment, technical indicators or a combination? Similarly, AI will use the same information to make its decisions but often lack the transparency that spells it out for us. Their models might be very accurate but, again, the ‘black box’ stigma still exists in some corners.
Explainability also provides a framework for justifying over- and underperformance that is rooted in fact, not forecast. In much the same way as members of an investment team might review their rationale for a trade after they have assessed how it delivered, explainable AI can disclose the degree of influence different variables had on a model at certain points in time, how they changed and the effects, if any, on the model.
This creates a structured and accurate audit trail over time, rather than relying only on trader notes or retrospective summaries, which can be affected by hindsight bias. All these distinctions are important because the more an investment professional understands a model’s logic and the reasoning behind its allocations or signals, the more familiar it becomes.
That familiarity will deepen users’ understanding of the tools they are using. In turn this will enable them to challenge the model with greater confidence, assess how the signals compare with their own market views and decide how much faith to place in the output. In wealth management, explainability therefore turns AI from a signal-generation engine into a more transparent decision-support tool.
