Over the past eighteen months, generative AI tools have moved into daily business workflows faster than almost any technology in recent memory. Employees draft contracts with ChatGPT, summarise client calls with Claude, and paste financial data into whatever assistant is fastest to hand. Most of this happens without IT’s knowledge, without a policy governing it, and without anyone in the organisation able to say with confidence what data has left the building. This is Shadow AI, and for asset managers and SMEs handling sensitive client information, it has quietly become one of the more consequential governance gaps of the year.
How This Actually Plays Out
Shadow AI rarely arrives as a dramatic breach. It arrives as convenience. A relationship manager pastes a client’s portfolio summary into a public chatbot to draft a quarterly letter faster. A junior analyst uploads a term sheet to get a second opinion on structure. A developer feeds proprietary code into an assistant to debug a script. None of these actions are malicious, and most employees genuinely believe they are simply working efficiently. But each one is an uncontrolled data egress event, often to a third party processor whose data retention, training use, and jurisdictional handling are unknown to the organisation, and in many cases unknown even to the employee.
The scale compounds the risk. Unlike a single unauthorised software installation, AI tool use tends to spread horizontally and quickly across teams, because the barrier to adoption is a browser tab, not an IT ticket. By the time a compliance function notices, usage is often already embedded in daily workflows across several departments.
Why Traditional Controls Don’t Catch It
Most SMEs and boutique asset managers built their security stacks around a different threat model: malware, phishing, unauthorised devices, network intrusion. Shadow AI does not trip any of those wires. It uses sanctioned devices, sanctioned browsers, and often personal or unmanaged accounts on otherwise reputable AI platforms. Data Loss Prevention tools calibrated for file transfers or email attachments frequently miss text pasted directly into a web interface. The result is a blind spot that sits squarely inside the perimeter organisations believe they have already secured.
Regulatory exposure follows directly. For asset managers handling client financial data, pasting portfolio details or personal information into a third party AI tool without a data processing agreement or legal basis can constitute a breach under Swiss data protection law (nDSG) or, for EU facing business, the GDPR, regardless of whether any incident ever occurs. The exposure exists the moment the data leaves the organisation’s control.
What Effective Governance Looks Like
The answer is not a blanket ban, which tends to fail quietly as employees simply move usage to personal devices, further outside the organisation’s visibility. Effective governance instead starts with visibility: understanding which AI tools are actually in use today, through network level discovery rather than a survey employees may answer inaccurately. From there, organisations need a tiered policy that distinguishes between data classifications: what can never leave the organisation’s own environment, what may be used with an enterprise agreement in place, and what carries no meaningful sensitivity at all.
Enterprise tier agreements with AI vendors, where available, typically include contractual guarantees against training on submitted data and defined data residency terms, a materially different risk profile from consumer tier access to the same underlying model. Where sensitive client or proprietary data is involved, self hosted or vetted enterprise deployments should be the default, not the exception.
Finally, this is a training and culture problem as much as a technical one. Employees adopt Shadow AI because sanctioned alternatives are slower or don’t exist. Providing a fast, approved tool with clear boundaries closes the gap far more effectively than a policy document alone.
Key message: Shadow AI is not a future risk, it is a present, ongoing one, operating inside organisations that believe their perimeter is secure. Closing the gap requires visibility into actual usage, a tiered data policy, enterprise grade tooling where sensitivity demands it, and treating this as a governance priority now, before a regulator or a client does it for you.
