The Shadow AI Problem: You Are Already Running Systems You Cannot Name
If you have 500 employees, you probably have 500 unsanctioned AI pilots running right now — and the largest exposure is not data leakage, it is unaccountable decision-making.
Every organisation has a version of the same conversation happening internally. Leadership believes AI adoption is at the pilot stage — carefully scoped, with a steering committee. Meanwhile, in finance, someone is pasting draft board commentary into a consumer chatbot. In HR, a recruiter is running CVs through a browser extension nobody procured.
None of these people are behaving badly. They are behaving rationally. The tools work, and no one has told them where the line is. That is not an employee problem. It is a governance vacuum — and vacuums get filled.
Why "data leakage" undersells the risk
The standard framing of shadow AI is confidentiality: sensitive data leaves the perimeter. That is real, but it is the smallest part of the exposure. The larger exposure is unaccountable decision-making. When an unsanctioned tool influences a hiring shortlist or a credit decision, your organisation has made a decision it cannot reconstruct — no record of what model, what version, what prompt, or what a human did with the output.
The inventory is the first governance artefact
ISO/IEC 42001 Clause 4 requires you to define the boundaries of your AI Management System and identify the AI systems in scope. Every downstream requirement depends on it. A credible AI inventory captures: system and purpose, risk tier, data in and out, human oversight point, a named owner, and the third-party dependency behind it.
The amnesty approach
The instinct is to block. Blocking drives usage further underground and onto personal devices where you have no visibility at all. A declared amnesty — a fixed window in which anyone can register any AI tool with an explicit guarantee of no disciplinary consequence — consistently works better. The typical mid-sized enterprise finds three to five times more AI tooling in active use than leadership estimated, and the majority of it turns out to be low-risk and worth sanctioning.