Katherine Maslova

Family office advisor, founding partner of Bourgeois Bohème Fintech

AI Risks: The Questions to Ask

Most conversations about AI risk in wealth management focus on the obvious: algorithmic bias, regulatory exposure, and data leaks. These matter. But having built and tested AI agentic infrastructure from the ground up, I want to flag a quieter set of risks that family offices are significantly underestimating.

The first is data residency. A 2025 survey found that fewer than 30% of wealth management firms could identify the precise jurisdiction in which their AI vendor’s model was hosted. When a family office connects client data to an AI tool, most principals have no idea where that data sits or who has contractual access to it. The question every family should be asking every vendor is simple: where is the model hosted, and who can see what? In Europe and the Middle East, that is not a theoretical concern. It is a regulatory and reputational exposure hiding in plain sight.

The second is what I would call governance by assumption. Many offices are using tools quietly integrated into daily workflows with no oversight structure, no accountability framework, and no human override protocol. When something goes wrong, nobody knows who is responsible. And things do go wrong. I have a documented case: one autonomous agent breached an internal corporate chatbot in under two hours, accessing tens of millions of confidential messages through a single unpatched API. Not a sophisticated nation-state attack. Just one AI agent.

The third is the absence of practical boundaries. NVIDIA’s internal security team applies what they call a Rule of Two: AI agents can access files, access the internet, and execute code, but never all three at once. A simple boundary that most family offices have not thought about at all.

Before any firm-wide rollout, start small. Use ready-made tools to solve your own pain points first. Understand what access they require, what they touch, and how they behave before bringing in consultants who are often more motivated to build than to protect. The organisations that navigate this well are not those who block everything or chase everything. They are those who learn early by making small, deliberate steps, drawing smart boundaries, and staying curious without losing judgment.

One rule to start with today: never paste private client information into a public AI tool. It sounds obvious. It is not yet standard practice.

Dion Bailey

Co-founder & Chief Product and Technology Officer

Do You Actually Understand What You Have Already Adopted?

I sit across multiple industries simultaneously, and the thing I keep bumping into is the same conversation, dressed differently depending on the room. Everyone is talking about AI adoption. Boards, advisers, executives. I understand it. But adoption was the conversation from a few years ago. The question that actually matters now is whether you understand what you have already adopted.

Here is what I mean. Open your Gmail. AI is in there. Your CRM, your compliance screening, your portfolio analytics. AI is already in there too. These were not decisions most people made consciously. They came bundled with the product, turned on by default, and now they are informing the work. That is not a criticism. That is just where we are. The real question is whether anyone has actually looked at what these systems are trained on, or what assumptions they are carrying.

Because here is the part that does not get said enough. These are prediction engines. They sound like they understand things. They do not. They have no grasp of context, culture, or consequence. They reflect the data they were trained on, and that data has blind spots baked in.

In 2024, Goldman Sachs faced regulatory penalties over the Apple Card credit algorithm. The system was found to have produced discriminatory outcomes across applicant profiles. Goldman could not clearly explain why, even when the decisions were technically within the rules. The institution did not know what it was running. You cannot challenge something you cannot see.

Scale that into multigenerational wealth decisions across jurisdictions and cultures, and the stakes feel different. These models are overwhelmingly built on Western, English language data. There is documented evidence that the same tool gives materially different answers depending on what language you ask it in. That is not a footnote for a family operating across geographies. That is a structural problem sitting in the middle of your process.

Literacy is not about knowing what AI is. It is about knowing when it is wrong, when it has drifted, when it is confidently telling you something that does not hold.

The system you are trusting most right now was built by someone, trained on something, and designed to see certain things clearly and others not at all. The question worth asking, whether you are a principal or the adviser beside one, is whether you actually know which is which.

We work with families and advisers who value clarity over noise. If you would like to explore whether we might be useful to you, the conversation starts here.