By Peter Alder, Marcuard Heritage

Independent wealth managers handle highly sensitive personal data every day, from portfolio holdings to the family circumstances of their clients. At their core, they bear the same data-protection and supervisory obligations as the big banks, yet they command only a fraction of the resources for data protection, cybersecurity, and data quality.

Generative artificial intelligence promises considerable efficiency gains in exactly these areas, precisely where small teams still do a great deal by hand. When approved AI tools are missing or fail to convince, «shadow AI» almost inevitably emerges: the use of freely available AI applications without the company's consent. It is not a ban on AI but its managed, compliant use that protects client trust and, at the same time, secures the industry's competitiveness.

What Lawmakers and Regulators Demand

Switzerland's Data Protection Act is deliberately technology-neutral, which means it applies directly to AI-driven data processing as well. Anyone deploying AI tools must disclose the purpose, the way they work, and the data sources, and must uphold the principles of transparency, purpose limitation, proportionality, and data security.

On top of this come obligations for commissioned data processing, for disclosing personal data abroad, for data-protection impact assessments, and for reporting breaches, all backed by the personal criminal liability of those responsible. Decisive in practice: entering personal data into a web-accessible AI tool counts, legally, as disclosure to a third party, and usually abroad at that.

Professional secrecy, to which independent wealth managers are also subject, bites even harder. Anyone who uploads client data to an AI tool the company has not approved may breach it, and good intentions or efficiency arguments offer no protection.

Supervisory organizational duties also require the management of all risks, including legal and reputational ones, with effective internal controls that must explicitly cover the use of AI tools.

Finma, Switzerland's financial market regulator, has likewise spelled out its expectations. It calls for AI governance with a central inventory and risk classification, clear responsibilities that can be delegated neither to the AI itself nor to third parties, along with data quality, testing, ongoing monitoring, documentation, and explainability.

According to Finma's survey of April 2025, around half of the roughly 400 institutions polled already use AI in daily operations or have first applications in development; a further quarter plans to deploy it within the next three years. Engaging with AI risks is therefore no longer optional but mandatory. The good news: regulation does not prohibit AI, it sets the framework for how AI may be used.

«Shadow AI»: the Rule, Not the Exception

«Shadow AI» arises wherever freely available AI tools meet the pressure to be efficient and a lack of – or unappealing – approved alternatives. Private, familiar applications get folded into the working day without anyone asking.

The typical pattern: the company-sanctioned AI tool, say an application with data storage and a data-processing agreement in Switzerland or Europe, convinces on data protection, while other tools often deliver better results in everyday work. It is precisely this performance gap that encourages the retreat into «shadow AI».

A study by HP Switzerland among some 1,300 office workers in German-speaking Switzerland shows the scale: more than half use AI tools at least weekly, only 48 percent report clear company guidelines, and 27 percent enter client-specific or personal data. «Shadow AI» is thus not the misconduct of a few individuals but the norm every organization must reckon with, and one that can turn into a security risk.

Ways Out of the Dilemma

For the industry, protection begins with visibility. An inventory of approved AI tools and use cases lays the foundation.

On top of that come clear AI guidelines that spell out prohibited data entry, such as personal data in unapproved tools, along with technical controls like blocking and data loss prevention, plus a deliberately simple approval process for new use cases governed by a «human in the loop». Equally important are attractive, approved alternatives built on corporate contracts that rule out the use of data to train language models and that guarantee data residency and confidentiality in Switzerland or the EU.

Providers' terms differ considerably and change constantly, which demands continuous attention from companies. The supposed dilemma between efficient AI and the protection of client data can be resolved on a risk basis: by anonymizing or pseudonymizing client data before entry, by cleanly separating use cases with and without personal data, and by putting contractual and technical safeguards in place with the AI-tool providers.

How Marcuard Heritage Contains the Problem

At Marcuard Heritage, everything starts with data quality. The heterogeneous data streams from various suppliers are consolidated in the core application, with clear data ownership, classification, and matching controls along the entire data value chain.

Poor data quality leads to poor output, and that holds for in-house data analysis just as much as for any AI application. Building on this, the firm invests in its people and fosters an understanding of AI through regular training, covering the approved use of AI tools, the anonymization of client data, and how to handle AI-generated output.

The principle stays the same: responsibility for the result always rests with the employee, and thus with Marcuard Heritage, never with the machine. This interplay of a consolidated data base, clear rules, and trained staff lets the firm deploy AI where it brings efficiency, without ceding control over client data.

Breaches of data protection or secrecy can trigger reporting duties and criminal liability. Above all, though, they can damage client trust and reputation.

Is Trust a Real Differentiator?

Switzerland's and Europe's stricter data-protection rules do, admittedly, look like an efficiency and cost disadvantage next to those in the US and China. Whether this actually translates into a competitive disadvantage for the two is hard to gauge today.

Trust in data security, financial markets, and legal certainty has long been a mark of quality, and one that should endure. Those who demonstrably handle client data in exemplary fashion position themselves as trustworthy service providers.

A look into the crystal ball suggests that in ten years know-how will have become a commodity, with investment analysis, research, and performance reporting freely available. The successful wealth manager of the future will likely combine human expertise, the controlled use of AI, and the trust of clients.

The balancing act between regulation and «shadow AI» succeeds when regulation is read as a compass: the Data Protection Act, the Financial Institutions Act, and Finma's expectations set out how AI may be used, they do not ban it. Those who replace «shadow AI» with compliant, attractive alternatives, clear policies, and trained staff gain efficiency while preserving the industry's most valuable asset: the trust of clients.

Peter Alder is chief operating officer and head of compliance at Marcuard Heritage. He has worked for the firm since 2012 and is responsible for, among other things, compliance, risk management, and the regulatory matters of the independent, internationally active wealth manager headquartered in Zurich.